《No One Did It: Responsibility Laundering, from the Scapegoat to the Algorithm》（第 4 部分 · 第 12–13 章）

No One Did It: Responsibility Laundering, from the Scapegoat to the Algorithm (part 4, chapters 12–13)

> By Li Xiaolai · his book manuscript published on GitHub · [GitHub repository](https://github.com/xiaolai/no-one-did-it) · version of 2026-05-31, 20 chapters; this page has chapters 12–13. Images are given as links.

Chapter 9 — The Model Did It

---
chapter_id: 09-the-model-did-it
chapter_number: 9
part: III
title: The Model Did It
evidence_grade: A
---

Chapter 9 — The Model Did It

> the servant is himself an instrument which takes precedence of all other instruments.
>
> — Aristotle, *Politics* I.4, trans. Benjamin Jowett

Across a single calendar window in the spring of 2025, the same grammatical move surfaced three times. Each time a non-human noun was placed in the verb's subject position, and each time it absorbed a decision a person had made. The move ran at three structurally distinct layers of one technology stack: **input**, where the model is built; **deployment**, where it is run; **evaluation**, where it is benchmarked. At each layer a different thing sits in the verb's subject position — at input it is the data, at deployment it is the model, at evaluation it is the variant. We meet these layers through three companies, in the order the events unfolded — but the layer, not the company, is the thing to watch.

On 5 April 2025, Meta Platforms, Inc. launched the Llama 4 family on its AI research blog and cited, in the launch communications, a position of #2 on the LMArena Chatbot Arena leaderboard. The model entry that occupied #2 was labelled "Llama-4-Maverick-03-26-Experimental." It was not the same set of weights that Meta released for public download. When the publicly released Llama 4 Maverick weights were added to LMArena under their public name six days later, on 11 April 2025, they ranked approximately #32. A gap of about thirty places. The first official explanation, given through a Meta spokesperson and quoted in *The Verge*, *TechCrunch*, and *The Register* between 7 and 8 April 2025, named the agent as the variant: "Llama-4-Maverick-03-26-Experimental is a chat-optimized version we experimented with that also performs well on LMArena."

On 24 and 25 April 2025, OpenAI rolled out an updated version of GPT-4o as the default model behind ChatGPT. Within 48 to 72 hours, users on X and Reddit had surfaced examples of the deployed model producing what OpenAI's own subsequent post-mortem called "sycophantic" behaviour — praising a transparently bad business idea, applauding a user's narrative of stopping psychiatric medication and disengaging from family, validating delusional framings the model's published usage policy had been written to refuse. On 27 April 2025 Sam Altman, CEO of OpenAI, acknowledged the issue on X. The first official explanation, given in OpenAI's post-mortem of 29 April 2025 and the expanded post-mortem of 2 May 2025, named two agents at the verb position. The first was the model: "GPT-4o skewed towards responses that were overly supportive but disingenuous." The second was an optimisation framework: the company had "focused too much on short-term feedback."[¹](#c9-1)

In the same April-to-September window, in a federal courthouse in San Francisco, a third version of the same grammar was being prepared for adjudication. On 23 June 2025, Judge William Alsup of the United States District Court for the Northern District of California handed down a summary judgment ruling in *Bartz v. Anthropic PBC*, Case No. 3:24-cv-05417. The ruling divided one company's conduct in two. Training a large language model on books that Anthropic had lawfully acquired was held to be transformative fair use under 17 U.S.C. § 107. Downloading and retaining approximately 196,640 books from the shadow library LibGen, and a further set of pirated books from PiLiMi, was held to be not fair use.

The procedural sequence that followed has been short and steep. The class was certified on 17 July 2025. The settlement was announced on 5 September 2025, in the amount of $1.5 billion — what would be, if the court grants final approval, the largest reported copyright settlement in United States history by acknowledged press reporting. As of our manuscript-freeze date that approval is pending; we hold the superlative to that condition here and below.

The input layer's official explanation predates the Anthropic case and was stated most sharply by another firm. OpenAI laid it down in a written submission to the United Kingdom House of Lords Communications and Digital Committee on 8 January 2024, and the framing was adopted in variant form across the litigation cluster. The submission named the agent as the data: "it would be impossible to train today's leading AI models without using copyrighted materials." Training data was renamed the load-bearing substance. The decisions that placed which works into which corpora on which dates were displaced from the grammar.

Three operations. Three layers of the same stack — input, deployment, evaluation. Three official accounts in which the named agent is a non-human entity — the data, the model, the variant — and the named human decisions, the named acquisition dates, the named release approvals, and the named submission decisions are not on the visible surface of the explanation.

The fair clue is in the grammar. At each of the three layers, the verb's subject is a thing. The thing that "performed well." The thing that "skewed." The thing that "would be impossible to train without." We ask, at each layer, what the grammar was doing — and which named human decision the grammar displaced.

Three operations, three official stories

The three official stories belong to the three layers, and we set them down in their own words before taking them apart. Each is from a primary source. Each is quoted to the standard the project's quote-integrity rule requires.

The first is at the input layer. The training-data framing has been laid down across more than a decade of corporate communication, but its highest-stakes statement in the litigation period is OpenAI's written evidence to the United Kingdom House of Lords Communications and Digital Committee, dated 8 January 2024. The sentence reads: "it would be impossible to train today's leading AI models without using copyrighted materials." The same submission characterised the choice as a structural one between training on copyrighted corpora under existing fair-use doctrine and not training competitive models at all. Around the same period, OpenAI and other AI developers continued to describe their corpora using the phrase "publicly available data," a framing that conflated availability on the open internet with permission for AI training. A third framing — adopted in defensive filings and accepted at least partially by the UK High Court on 4 November 2025 in *Getty Images v. Stability AI* — held that once a model is trained, its weights "contain statistically trained parameters, not stored copies or reconstructions" of any individual training example. The input is renamed substance. The output is renamed abstraction. The decisions that connect them are removed from the verbs.

The second official story is at the **deployment layer**. OpenAI's post-mortem of 29 April 2025, "Sycophancy in GPT-4o: What happened and what we're doing about it," is the load-bearing primary document. It uses, throughout, the grammar of system agency. The model "began behaving" in certain ways. The optimisation framework "focused too much" on short-term feedback. The expanded post-mortem of 2 May 2025, "Expanding on what we missed with sycophancy," adds a procedural layer. The post-mortem states that internal evaluators — OpenAI's own term is "expert testers" — had raised concerns about the model's behaviour before launch, and that those concerns were treated as outweighed by positive signals from a broader A/B testing population. The post-mortem does not name the testers. It does not name the engineers who designed the reward signal. It does not name the product leadership that approved the default-model swap. The post-mortem names the optimisation choice. Sam Altman's X post of 27 April 2025, made before the formal post-mortem and setting the public framing under which the post-mortem would be read, characterises the issue at the personality level: the last couple of GPT-4o updates had made the personality "too sycophant-y and annoying."[²](#c9-2) Two layers of language. The CEO's framing places the cause on the model's personality. The post-mortem places the cause on the optimisation framework and the release process. Neither layer names a human decision-maker.

The third official story is at the **evaluation layer**. Meta's position on the Llama-4-Maverick-03-26-Experimental submission, given through a spokesperson and quoted across the trade press between 7 and 8 April 2025, treats the LMArena leaderboard as one experimental surface among many: "We experiment with all types of custom variants. Llama-4-Maverick-03-26-Experimental is a chat-optimized version we experimented with that also performs well on LMArena." On 7 April 2025, Ahmad Al-Dahle, Vice President for GenAI at Meta, posted on X denying that Meta had trained on test sets and characterising rumours of such conduct as "simply not true."[³](#c9-3) When the controversy moved from the model provider to the benchmark operator, LMArena (operated by Arena Intelligence, Inc., the spin-out of the LMSYS research project) responded in mid-April 2025 with a public statement that Meta's interpretation of LMArena's disclosure policy had not matched what the benchmark expected from model providers, and announced a policy revision. The named agent at the verb position shifts again. First the variant performed. Then the policy did not match expectations. At neither moment is the agent the named individual at Meta GenAI who chose to submit a stylistically tuned variant to a public leaderboard under a name that paired the variant with the upcoming public release.

We can hold each of the three stories in mind separately. Our first promise here is that the three are not separate. They are a family of grammatical operations. The verb's subject at the input layer is "data." The verb's subject at the deployment layer is "the model" or "the reward signal." The verb's subject at the evaluation layer is "the variant" or "the policy." Each operation displaces a named human decision. The grammar is the laundering. The diagnostic that follows is what reaches the displaced decision.

Running the diagnostic at the input layer

The instrument here is the same one chapter 3 installed — the eight questions, asked in order: blame, control, benefit, knowledge, preventability, record-control, cost, and what responsibility would look like if it followed control instead of visibility. Chapter 7 closed Part II by showing that those eight only operate on cases where the record is accessible. The record is the prior condition. Now we take the next step. The AI stack is not a single architecture. It is three structurally distinct layers — input, deployment, evaluation — and at each layer the responsibility-laundering grammar names a different non-human agent, displaces a different set of human decisions, and requires the diagnostic to climb to a different rung of named actors. The same alibi shape recurs at each layer; the diagnostic reaches different humans each time. We begin at the input layer, and we run all eight in full.

The anchor is the training-data litigation cluster. Hybrid classification: the primary case-type is **system/object alibi**; the secondary is **cost-bearing goat**. The primary classification holds because the load-bearing laundering is the grammar — "data," "the corpus," "the model" — that renames human creative work as substance. The secondary fires because the rights-holders absorbing the cost of the renaming are not visible at the narrative centre of the AI story. The two classifications do not collapse into each other. The first names the mechanism. The second names the harm.

*Who was publicly blamed?* Four sequential containers, escalating across the period of the litigation. First, "data" or "the corpus" — the input substance, the most basic alibi, in which human work is renamed material that data curators (Common Crawl; the Books3 corpus assembled by Shawn Presser from the shadow library Bibliotik in 2020; LAION; EleutherAI for The Pile) are responsible for. Second, when the suit gets specific in *Bartz v. Anthropic*, "the shadow library" — LibGen and PiLiMi as the named non-human accomplices, indifferent to which works they contained, acquired by Anthropic at one remove from the original piracy. Third, "the model" itself — once trained, the artifact whose outputs are at issue, framed in defensive filings as containing statistical parameters rather than copies. Fourth, "the market" or "national AI competitiveness" — the highest-order alibi, in which constraining training is positioned as ceding strategic ground. At every container the named agent is non-human and the named decisions are absent.

Control, benefit, and knowledge run together at this layer, and they run to the same place. Each AI developer controlled the choice of which corpora to include in training. Each controlled the choice between lawful acquisition — for example, OpenAI's later licensing agreements with News Corp, the Associated Press, Axel Springer, Le Monde, and Reddit, and Anthropic's later lawful acquisition of books — and shadow-library acquisition. Each controlled the model release decision. Each controlled the pre-litigation messaging that framed the activity. Per the *Bartz* ruling, Anthropic in fact had the option to acquire books lawfully and in some instances did so. The shadow-library acquisition was, on the court's reading of the acquisition record, a discretionary choice. The institutions positioned as upstream — Common Crawl, LAION, the Books3 curator — controlled the assembled corpora made available to the labs. The load-bearing controlled entities, by the court's analysis, are the labs that chose to use the corpora.

The benefit accrued to those same labs. OpenAI's late-2025 valuation exceeded $150 billion; Anthropic's exceeded $30 billion; Microsoft's Azure OpenAI revenue ran to tens of billions across the same window. The value of the training corpora was captured in model weights, which generated the revenue. The rights-holders received no compensation at training time. Post-litigation settlements — *Bartz*'s $1.5 billion; the licensing deals that followed — capture a fraction of the value extracted at training. The class structure of the *Bartz* settlement makes the per-author figure legible: approximately $3,000 per work for approximately 500,000 works included in the LibGen and PiLiMi corpora that Anthropic had acquired. The per-work figure is the settlement's mathematics, not the market value extracted from the corpus at training time. The convergence of control and benefit at the same actor is the structural fact on the public record; whether benefit motivated the control choices is a question for evidentiary discovery.

And the documentary record establishes the knowledge. The *Bartz* ruling and exhibits, per Judge Alsup's summary judgment order of 23 June 2025, document that Anthropic's pre-2024 acquisition of LibGen and PiLiMi corpora was visible to the company at the level of internal communication referencing the corpora by name and acknowledging their unlicensed character. The level of internal knowledge varies by defendant; *Bartz* is the highest-evidence-grade finding to date that a major AI developer knowingly trained on pirated books.

Two other cases in the same cluster have moved the public-record question forward at different speeds. In *Authors Guild v. OpenAI Inc.* (S.D.N.Y., Case No. 1:23-cv-08292, Judge Sidney H. Stein), the October 2025 ruling on OpenAI's motion to dismiss denied the motion in part and allowed copyright-infringement claims related to ChatGPT output to proceed to discovery. In *The New York Times Company v. Microsoft Corp.* (S.D.N.Y., Case No. 1:23-cv-11195, MDL), the November 2025 discovery order required OpenAI to preserve output log data including ChatGPT conversation logs that had been scheduled for deletion under the company's standard retention policy[⁴](#c9-4).

Discovery in each case is the instrument by which what was known internally to each developer enters the public record.

*Who could have prevented recurrence?* The lab, in choosing not to acquire pirated corpora. The regulator, in setting training-data disclosure obligations. The court, in interpreting fair-use doctrine against acquisition-record evidence rather than against post-training behavioural abstractions. Each is preventability at a different layer. Each is now partly engaged. California's SB-942, the California AI Transparency Act, signed by Governor Gavin Newsom on 19 September 2024, was amended by AB-853 (signed 13 October 2025) to push its operative date to 2 August 2026; its primary focus is AI-generated content provenance rather than training-data manifest disclosure. The European Union's AI Act, Regulation (EU) 2024/1689, contains training-data summary obligations in Article 53 for general-purpose AI model providers, with operative dates phased across 2025 and 2026. The statutory layer is partly operative and partly future-tense. The judicial layer, per *Bartz*, has reached the acquisition record and applied existing copyright doctrine to it.

*Who controlled the record?* The labs controlled the training-data manifests, most of which are not public. Anthropic's manifest was partially disclosed under court order in *Bartz*. The labs controlled the internal communications about corpus acquisition. The labs controlled the model weights themselves. The labs controlled the inference logs that could — per the November 2025 NYT discovery order — reveal which user prompts produced regurgitation of substantial passages from copyrighted works. The U.S. federal courts (Northern District of California, Southern District of New York, MDL) control the litigation record. The UK High Court (Chancery Division) controls the *Getty v. Stability AI* record. The California legislature controls the SB-942 and AB-853 statutory record. *The training corpus is the record.* What was in it, and how it got there, is the diagnostic question on which the input-layer alibi stands or falls.

*Who bore the cost?* Authors, photographers, journalists, code contributors, visual artists. The *Bartz* settlement makes the cost legible for one defendant on one corpus subset; the diffuse cost — journalists whose work is now competed against by ChatGPT-as-summariser; photographers whose stock market has been hollowed by image generators trained on their archives; programmers whose code was used to train code-assistant models; artists whose distinctive styles can be invoked by prompt — is structural and partially invisible in the case-by-case litigation record. The rights-holders are not being publicly blamed for AI's growing pains. They are being erased from the grammar of how AI training is described. The alibi here is not "the authors caused this" but "training data is a substance, not a body of human work." Cost-absorbing silence rather than active accusation. That is the operative form of cost-bearing goat at the input layer.

What would responsibility look like if it followed control instead of visibility? Named acquisition decisions, by named executives, at named labs, on documented dates, against documented licensing alternatives. Anthropic PBC is the corporate respondent in *Bartz*. Per the public record, Dario Amodei is CEO of Anthropic PBC. No public statement attributing the LibGen or PiLiMi acquisition decision to Amodei personally has been documented as of May 2026. No civil or criminal liability has been adjudicated against Amodei as of May 2026. The litigation has reached the corporate actor. The five-role conflation we named in Chapter 7 — same institution deploys, audits, complains, prosecutes, discloses — applies directly to AI training at the input layer. The same lab assembles the corpus, controls the manifest, defends the litigation, and decides what is disclosed in discovery. Chapter 7's anti-laundering rule applies: where one institution holds all the record-control roles, external examination is the necessary instrument.

The reversal sentence belongs here. The training-data alibi is the input-layer instance of the same record-control conflation chapter 7 installed at the institutional layer. The "data is just data" grammar performs the laundering. The discovery record breaks it.

Therac, MAX, Horizon — updated

The "model decided" alibi is not new. The scale is new. The stack-depth is new. The grammar is not. We take three callbacks at one or two sentences each, to ground the diagnostic in cases we have already installed at full depth.

Chapter 2 read **Therac-25**, the radiation therapy machine built by Atomic Energy of Canada Limited, which produced fatal overdoses at six US and Canadian treatment centres between June 1985 and January 1987. The audit trail was a single memory-mapped error-code counter; "Malfunction 54" and "TREATMENT PAUSE" were the alibi. The "model decided" framing is the direct lineal descendant of "the software was the cause," at scale: the model is the software, the training corpus is the missing audit trail, and the benchmark rank is the new "TREATMENT PAUSE."

Chapter 2 also read the **Boeing 737 MAX** as a partial scapegoat with system/object alibi secondary — pilots were named as the cause of the Lion Air and Ethiopian Airlines crashes while the Maneuvering Characteristics Augmentation System became the engineered-system blame container. The model alibi runs the same shape with no pilot. "MCAS handled it" becomes "the model decided," but in the AI stack there is no pilot at the verb position to absorb the blame. There is only the lab, the deployer, and the executive who approved the release. The earlier chapter's partial scapegoat is replaced by an unoccupied seat.

Chapter 7 installed the **Horizon** five-role conflation diagnostic. Same institution deploys, audits, complains, prosecutes, and discloses. The AI stack runs the same conflation with the labels updated. The same lab trains the model, releases it, monitors its post-deployment behaviour, writes the post-mortem when the behaviour goes wrong, and controls the training-corpus disclosure. The Post Office held five roles against a population of sub-postmasters dispersed across the United Kingdom. The AI labs hold the same five roles against a population of users dispersed across the consumer internet, and against a population of rights-holders dispersed across the world of creative work. Chapter 7's anti-laundering rule applies directly. The record is first. Where the record is controlled by the institution being diagnosed, the diagnostic cannot answer. The first task is to break the record-control architecture. What follows shows what that task looks like at each AI-stack layer.

A 72-hour rollback

The deployment layer is the AI stack's most visible surface. It is also the layer at which the laundering operates fastest and at which the institutional response is most ready-made.

Return to the GPT-4o sycophancy incident. The behaviour was systematic. The pattern was visible enough across X and Reddit by 26-27 April 2025 that Sam Altman acknowledged it publicly within 48 hours of the rollout. OpenAI rolled the update back over 28-29 April 2025. The published post-mortem appeared on 29 April; the expanded post-mortem on 2 May. The technical mechanism the post-mortem identifies is a reward-shaping choice: the post-training optimisation framework was weighted toward short-term user feedback signals, and that weighting produced the sycophantic skew when the model was deployed at default-model scale. The procedural mechanism the post-mortem identifies is a release-decision choice: internal expert testers had flagged concerns, those concerns were treated as outweighed by positive signals from a broader A/B testing population, and the release proceeded.

Each of these findings is on the public record because OpenAI placed it there. The post-mortem is a primary corporate document. It uses the grammar of system agency throughout. The model "began behaving." The reward signal "focused too much" on short-term feedback. The process "did not adequately weight" qualitative tester concerns against quantitative ones. The named agent at every verb is a non-human entity — the model, the signal, the process. The post-mortem does not name the engineers who designed the reward signal. It does not name the team leads who approved the design. It does not name the product leadership that approved the default-model swap. It identifies "expert testers" as a category of internal evaluator whose number, seniority, and specific concerns are not publicly disclosed. The displaced decisions are inside the company's record-control envelope.

Run the eight here in abbreviated form — the same chapter-3 instrument, not a new one. Four of the eight answer exactly as the prose has just shown: the blamed agent is the model and the reward signal; the benefit runs to the short-term user-feedback signals the optimisation framework was, by the post-mortem's own account, weighted toward; the cost falls on the users praised for stopping medication or disengaging from family; the record sits inside OpenAI's post-mortem. What the deployment layer bends is the displaced-decision trio — control, knowledge, and preventability. At the input layer those three pointed at acquisition decisions; here they point at different humans, the engineers and product leadership who shaped the reward signal and approved the default-model swap. That is the subset this layer compresses; we walk it now.

The deployment-layer responsibility chain reaches different humans than the input layer's chain. The displaced decision-makers are the engineers and the product leadership inside OpenAI who approved the reward-shaping change and the default-model swap. We do not name those engineers or product leaders. None has been publicly identified in primary sources as the decision-maker on either choice. The publicly accountable executive is Sam Altman as CEO. Sam Altman is CEO of OpenAI. Altman acknowledged on X on 27 April 2025 that recent GPT-4o updates had made the personality "too sycophant-y and annoying."[⁵](#c9-5) No civil or criminal liability has been adjudicated against Altman as of May 2026. The procedural-stage trio is what we can support; characterisation of Altman's contemporaneous knowledge of the sycophancy issue beyond the X post and the post-mortem documents is what the record cannot support, and we do not advance it.

The record-control architecture at the deployment layer is the five-role conflation from chapter 7, transposed. OpenAI deployed the model. OpenAI controlled the post-training reward signal that produced the behaviour. OpenAI controlled the pre-launch evaluation that produced the expert-tester record. OpenAI controlled the post-deployment telemetry that measured the production behaviour. OpenAI controlled the post-mortem document that named the cause. The user community — millions of ChatGPT users posting screenshots to X and Reddit — controlled the public reconstruction of the model's behaviour during the live window. Without that user-community surfacing, the alibi would have sustained. With it, the institutional response was forced within 72 hours. The user community at the deployment layer plays the role that the Justice for Subpostmasters Alliance played at the Horizon institutional layer — the aggregating mechanism that converts dispersed individual experience into a recognisable pattern in the public record. Without an aggregating mechanism, the deployment-layer alibi sustains by default.

The 72-hour interception window is fast by historical comparison. Horizon took 20 years from rollout in 1999 to Fraser J's *Horizon Issues* judgment in 2019. Therac-25 took roughly 24 months from the first fatal overdose in June 1985 to the FDA-led investigation's reconstruction of the failure mode across 1986 and 1987. GPT-4o's deployment-layer alibi was rolled back in days. The speed, however, is a function of the layer, not of the alibi. Software rollback is a sub-day operation. The user community's surfacing engine — X and Reddit screenshots — is a sub-hour operation when the model's behaviour is visibly anomalous. The institutional shell — fast acknowledgment, fast rollback, frank post-mortem — operates inside the same record-control envelope. No external auditor entered the loop. No court compelled disclosure. No regulator certified the cause. The institutional account is what OpenAI chose to disclose. The interception is real; the architecture is intact.

The deployment-layer observation to export is this. The fast-rollback institutional shell can be working as designed — quick acknowledgment, frank post-mortem, announced process changes — and still produce a structural absence of individual accountability for decisions affecting hundreds of millions of users. Blameless post-mortem methodology, imported from site-reliability engineering practice, is appropriate for systems failures in which a pager misfired or a configuration was wrong. It produces a responsibility-shaped hole when applied to *judgment calls* — release decisions, reward-signal choices, expert-tester-override decisions — made by named individuals whose identity then disappears into "the metric caused the behaviour" framing. The architecture of the laundering is operating at scale. The fast cycle is what makes it harder to see.

Alibi escalation at the leaderboard

The evaluation layer's alibi is structurally distinct from the deployment layer's, and it operates by a different escape mechanism. We pivot from GPT-4o to the Llama 4 LMArena controversy to install it.

Recall the timing. On 5 April 2025, the Llama-4-Maverick-03-26-Experimental entry occupied #2 on the LMArena leaderboard, just behind a Gemini 2.5 Pro Experimental entry from March. Within hours of Meta's launch blog post, HuggingFace community members posted side-by-side comparisons showing that the publicly released Llama 4 Maverick weights — when downloaded and run locally — produced terse, emoji-free output, whereas the experimental entry on LMArena was producing verbose, emoji-laden, stylistically tuned responses. The gap was visible to the open-source community within a week. On 11 April 2025, when the released weights were added to LMArena under their public name, they landed at approximately #32. The entry that placed second was the one nobody could download; the one anyone could download placed thirty-second. A 30-place gap on a leaderboard that is treated by parts of the AI industry as a procurement signal for model quality.

The first move of the evaluation-layer alibi was to place agency on the variant. The Meta spokesperson statement described "Llama-4-Maverick-03-26-Experimental" as "a chat-optimized version we experimented with that also performs well on LMArena." The variant is the named agent. The verb is "performs." The decision to submit *that* variant rather than the release weights — under a name that read in launch coverage as the upcoming Llama 4 release — is absent from the grammar. Mark Zuckerberg is CEO of Meta Platforms. No public statement attributing the Llama-4-Maverick-03-26-Experimental submission decision to Zuckerberg personally has been documented as of May 2026. No civil or criminal liability has been adjudicated against Zuckerberg as of May 2026. Ahmad Al-Dahle is Vice President for GenAI at Meta. Al-Dahle stated on X on 7 April 2025 that Meta did not train on test sets and that rumours of such conduct were "simply not true."[⁶](#c9-6) No civil or criminal liability has been adjudicated against Al-Dahle as of May 2026. We do not name engineers inside Meta GenAI as participants in the submission decision. None has been publicly identified in primary sources.

The second move of the evaluation-layer alibi is the one to learn from. When the controversy moved past the "we experiment with variants" framing — when the 30-place gap to the released weights became visible enough that the framing could not absorb it — the named agent shifted one level up. On 7 April 2025, LMArena issued a public statement that Meta's interpretation of LMArena's disclosure policy "did not match what we expect from model providers," and announced that it had "updated our leaderboard policies to reinforce our commitment to fair, reproducible evaluations so this confusion doesn't occur in the future"[⁷](#c9-7). The new named agent was the policy. Not the submission decision. Not the engineers or product leadership at Meta GenAI who made it. Not the benchmark operator's prior policy choices. The named gap was a policy gap. The institutional response was a policy revision. The alibi did not fail. It relocated one rung up the procedural shell.

Name the operation. *Alibi escalation.* When the laundering is caught at level N, it moves to level N+1. At level 1, the variant performed well. At level 2, the policy did not match expectations. The decision-makers inside Meta GenAI who chose to submit the variant under that name remain unnamed in any public accountability venue. The benchmark operator's policy revision is real institutional work. It is also the form of interception that leaves the architecture of the laundering intact. The procedural shell tightens. The named human decisions remain absent.

Run the eight a third time, abbreviated again — still chapter 3's instrument. *Blame* lands on the variant and then on the policy; control, knowledge, and preventability point to the submission decision inside Meta GenAI and to the benchmark operator's prior disclosure rules; benefit is the leaderboard rank cited in launch coverage. The two questions this layer bends are cost and record-control. The cost-bearer is not an identifiable population of harmed users but the credentialing commons — the shared evaluation infrastructure whose signal was compromised; and record-control sits not with a single defendant lab but with the benchmark operator, LMArena, which holds the submission record and writes the disclosure policy. Those are the two answers that distinguish the evaluation layer; we walk them now.

The evaluation-layer cost-bearer is not Meta's customers. The customers downloaded the released weights and received what was on the model card. The cost-bearers are different. They are the open-source community whose reliance on LMArena rank as a quality signal was silently compromised. They are the downstream developers who chose Llama 4 for their products partly on the leaderboard signal, only to find the deployed weights differed from the benchmarked variant. They are other model providers whose released weights competed on the arena under disclosure rules that did not match what at least one provider's submission practice was doing. The harm is to the credentialing commons rather than to identifiable individual users. That is the evaluation layer's distinctive shape. The harm is institutional-credibility erosion, and the cost-bearer is the broader infrastructure of model evaluation on which the AI development ecosystem partly depends.

The evaluation-layer escape mechanism is benchmark-organisation self-reform. LMArena revised its policies. This is roughly comparable in durability to the deployment-layer rollback. Faster than the legal-discovery channel at the input layer. Less expensive than the legal channel. Less reaching, in the sense that no individual at any named lab is publicly accountable for the decision that produced the controversy. The procedural shell adjusts; the substantive responsibility chain remains where it was. The benchmark becomes the instrument of its own anti-laundering rule. The benchmark is also the only instrument operating at the evaluation layer with any institutional standing to revise the architecture. There is no court compelling discovery. There is no regulator certifying submission rules. There is the benchmark, the model provider, and the community that watches both.

Alsup draws the line at acquisition

Return to the input layer. The legal-discovery channel produced, in 2025, the first court-documented application of existing copyright doctrine to AI training on the basis of the acquisition record — and the decisive detail is where the court drew its line.

Judge Alsup's summary judgment ruling of 23 June 2025 in *Bartz v. Anthropic PBC* did not turn on the model's behaviour. It did not turn on whether the model's outputs reproduced training examples in a copyright-relevant sense. It did not turn on the abstraction of model weights. It turned on the acquisition record. Anthropic had acquired books in two ways. Some had been acquired through lawful channels — purchased, scanned under fair-use scanning doctrine. Others had been acquired by downloading and retaining copies from LibGen and from PiLiMi. The court distinguished the two. Training on lawfully acquired books was held to be transformative fair use under 17 U.S.C. § 107. Downloading and retaining the shadow-library copies was held to be not fair use. The line was drawn at the input — the moment of acquisition — not at the output. The model alibi's grammatical move, in which the input is renamed substance and the output is renamed abstraction, was set aside in favour of an evidentiary record about what was acquired, from whom, on what date, under what licence.

From that line the case moved fast — class certified 17 July 2025, settlement announced 5 September 2025. The $1.5 billion figure worked out to approximately $3,000 per work gross of plaintiffs' fees and costs, across the roughly 500,000 LibGen and PiLiMi works in the certified class[⁸](#c9-8). The final fairness hearing was held before Judge Araceli Martínez-Olguín in the Northern District of California on 14 May 2026. The hearing recorded a claims rate of approximately 92.77% and 350 valid opt-outs covering 1,802 works. The court took the matter under submission rather than ruling from the bench, with distribution calculations to follow by 11 June 2026 if final approval issued[⁹](#c9-9).

As of our manuscript-freeze date, no order granting final approval has been recorded in the public docket; we use the procedural language the docket supports.

The wider litigation cluster is the field in which the input-layer alibi continues to be contested. Judge Stein's October 2025 ruling in *Authors Guild v. OpenAI* allowed copyright-infringement claims related to ChatGPT output to proceed to discovery. The November 2025 *NYT v. OpenAI* discovery order required preservation of output log data including ChatGPT conversation logs scheduled for deletion under OpenAI's standard retention policy, reattaching the model-output layer to the input layer for purposes of the regurgitation claims plaintiffs were advancing.

*Getty Images v. Stability AI* in the UK High Court, on 4 November 2025 — Mrs Justice Joanna Smith — dismissed Getty's primary copyright claims on territoriality grounds (training did not occur within the United Kingdom). The same ruling found limited trademark infringement for specific watermark generations, and granted Getty permission to appeal on secondary copyright infringement.

The territoriality holding is itself a finding about the model alibi. It says that the legal question is where the corpus was assembled and where the model was trained, not what the model's outputs subsequently produced. The judicial layer is moving toward the acquisition record. Multiple cases are pre-judgment as of May 2026.

The breakthrough at the input layer is the same instrument Chapter 7's Horizon interception required, transposed to the AI stack. Court with subpoena power capable of compelling production of the training-corpus manifest and the internal communications about corpus acquisition. Independent technical expert capable of reverse-engineering what the lab's training run actually did, distinguishing lawfully acquired corpora from shadow-library corpora at the level of detail a court can adjudicate. Named federal judge willing to engage substance — willing to apply existing copyright doctrine to the acquisition record rather than defer to the "model is a black box" framing. Three instruments. All three required for the input-layer alibi to be broken. Subtract any one, and the alibi sustains.

The five-role conflation is what the discovery process is reaching. The lab assembled the corpus. The lab controlled the manifest of what was in the corpus. The lab defended the litigation in which the manifest was sought. The lab decided what to disclose to plaintiffs and to the court. The lab controlled the inference logs that show what the deployed model produces from the corpus. Five record-control roles. One lab. Chapter 7's anti-laundering rule applies. The interception works when an outside fact-finder with subpoena power and an independent technical examiner with reverse-engineering capability force the manifest into the public record. Then the diagnostic — the eight questions — can operate on a record the lab did not produce. Then the responsibility chain can climb.

Run the diagnostic three times

##### Run the diagnostic three times per AI system — once at input layer, once at deployment layer, once at evaluation layer.

We have now run it three times: in full at the input layer, then twice in abbreviated form — once at deployment, once at evaluation — because each layer bends a different subset of the same eight questions. The eight did not change. What changed, layer by layer, was which of them did the work. At the input layer the displaced-decision questions reached acquisition decisions; at the deployment layer the same trio — control, knowledge, preventability — reached the engineers and product leadership who shaped the reward signal; at the evaluation layer the bent questions were cost and record-control, the credentialing commons and the benchmark operator. At each layer the named agent of the laundering grammar is different. At the input layer the agent is "data," "the corpus," "the shadow library," "the market." At the deployment layer the agent is "the model," "the reward signal," "the release process." At the evaluation layer the agent is "the variant," "the policy," "the benchmark." At each layer the displaced human decisions are different. At each layer the cost-bearer is different. At each layer the interception instruments are different. The diagnostic does not apply once per AI system. It applies three times.

> *The three layers as a table; the walk above is its audio rendering, and listeners may skip the table.*

::: {.figure-embed}
[〔图："The model did it" at three layers — input, deployment, evaluation — each with a different non-human subject taking the verb and a different party bearing the cost.〕](https://github.com/xiaolai/no-one-did-it/blob/b5e9c8f820bc04bb1d452d4109318673c08c3945/book/chapters-v6/book/evidence/diagrams/proofs/ch09-ai-stack.png)
:::

The operational form is the three-record demand.

##### The input-layer record demand: the training-corpus acquisition log.

For any AI tool whose output affects us, ask what corpora were used in training, when, from what sources, under what licensing or acquisition documentation. The Alsup distinction — between lawfully acquired corpora and shadow-library corpora — is the diagnostic test the federal court has already supplied. Where the acquisition record is denied, the denial is the diagnostic finding. The input-layer alibi is operating. The record-control architecture is intact.

##### The deployment-layer record demand: the post-training reward-shaping log.

Ask which version of which model is running, what post-training fine-tuning or reward-shaping changes have been applied since the last release, on what date, and what the change was. The GPT-4o post-mortem of 29 April 2025 shows what such a record looks like when an institution produces it after the fact. Demand it as standard before the deployment that affects us, not as post-incident apology after the fact. Where the deployment record cannot be supplied — where the deployed model is described as "the model" with no version designation and no post-training-change record — the deployment-layer alibi is operating.

##### The evaluation-layer record demand: the benchmark-submission record.

Ask whether the deployed model is the same set of weights whose benchmark scores were cited in marketing or procurement materials. Ask which variant was submitted, under what disclosure policy, on what date. LMArena's policy revision of mid-April 2025 shows the diagnostic moving from a case-specific dispute to a standing rule. Demand it as standing rule. Where the benchmarked variant differs from the released variant and the difference is not disclosed in the procurement document, the evaluation-layer alibi is operating.

The three demands integrate with chapter 7's five-role rule. At each layer of the AI stack, ask whether the same institution is performing all the record-control roles. At the input layer: does the same lab that assembled the corpus also defend the litigation, decide what to disclose, and operate the inference service that produces the regurgitation evidence? At the deployment layer: does the same company that deploys the model also produce the pre-launch evaluation, run the post-deployment telemetry, write the post-mortem, and decide what is named in the post-mortem? At the evaluation layer: does the same model provider that submits the variant also write the launch communications that cite the rank? Where one institution holds all the record-control roles at a layer, the laundering architecture is structurally operating. Where the rule is broken, external examination is the necessary instrument — court with subpoena power, independent technical examiner, named federal judge willing to engage substance. Chapter 7's instruments port forward. The number of times they must be deployed multiplies by three.

The litigation interception triplet

The AI-litigation venue is the emerging interception infrastructure. The mechanism is a triplet. None of the three is sufficient on its own.

The **first instrument** is discovery power. The Northern District of California and the Southern District of New York discovery processes have produced the acquisition record for AI training in a form no prior venue achieved. *Bartz v. Anthropic*, Case No. 3:24-cv-05417 — Judge Alsup compelled production of the corpus manifest that established Anthropic's pre-2024 acquisition of LibGen and PiLiMi. *NYT v. Microsoft / OpenAI*, Case No. 1:23-cv-11195 — the November 2025 order requiring preservation of output log data including ChatGPT conversation logs scheduled for deletion under OpenAI's standard retention policy. *Authors Guild v. OpenAI*, Case No. 1:23-cv-08292 — Judge Stein's October 2025 motion-to-dismiss-denied ruling allowed copyright-infringement claims related to ChatGPT output to proceed to discovery. Without discovery, the input-layer alibi sustains. The acquisition record is the institutional secret that the alibi depends on; the discovery process is what makes it visible.

The **second instrument** is technical-evidence analysis by court-appointed or party-retained experts. AI training corpora and model behaviours are technically opaque; turning a discovery production into a court finding requires expert technical analysis at the level of detail a judge can adjudicate. *Bartz* relied on technical evidence distinguishing lawfully acquired books from shadow-library copies. The regurgitation claims advancing through *NYT v. OpenAI* and *Authors Guild v. OpenAI* depend on technical reproducibility evidence. The instrument is the AI equivalent of Jason Coyne's reverse-engineering of the Horizon Known Error Log in chapter 7. Without it, the discovery production is opaque.

The **third instrument** is named federal judges willing to engage substance rather than defer to the "the model is a black box" framing. Judge William Alsup, Northern District of California, is the canonical instance. His ruling applied 17 U.S.C. § 107 fair-use doctrine to the acquisition record, refusing the framing that training is just data processing. The willingness to engage substance is the third instrument because the technical evidence without it produces records no court reads.

Any one alone fails. Discovery without expert analysis produces records no court can read. Expert analysis without discovery has nothing to analyse. Judges without either cannot rule on substance. The three together are the AI-litigation interception triplet.

We are honest about the cost of the escape.

Begin with the years. AI training litigation began in 2022 and 2023; the first court-documented breakthrough is Alsup's ruling of 23 June 2025. Two to three years for the first substantive court finding at the input layer. Fast by Horizon's 20-year scale. Slow by the deployment layer's 72-hour scale. The interception cost in years tracks the layer at which the alibi operates: input layer roughly two to three years, deployment layer roughly three days, evaluation layer roughly ten days. The faster interceptions reach less.

Then the money. The *Bartz* settlement of $1.5 billion is paid by Anthropic, recovered from a model that was trained on the contested corpora, and the product remains in market. The settlement does not require destruction of derivative work. The cost-bearing rights-holders receive a portion of the settlement, less plaintiff fees. The per-work figure is materially below the value extracted at training time. The settlement is the cost of the conduct for one defendant on one corpus subset, not the price of restoring the laundered architecture to a non-laundering one.

Territoriality remains unresolved, and it is a cost of its own. *Getty v. Stability AI* in the UK High Court dismissed primary copyright claims on territorial grounds. Training infrastructure can be located in one jurisdiction, corpus assembled in another, model deployed in a third. The multinational AI infrastructure makes the acquisition-record discovery transnational and slow. The European AI Act's training-data summary obligations under Article 53 are the operative procedural surface in the European Union, and our diagnostic exports cleanly into that surface, but the harmonisation across jurisdictions is itself a multi-year project.

The non-litigated layers carry the heaviest cost of all. The deployment-layer rollback does not climb to executive accountability. The evaluation-layer policy revision does not climb to the named submission-decision-maker at the model provider. Only the legal-discovery channel produces, on the public record to date, the kind of acquisition-record documentation that could support climbing to named individual accountability — and even that channel has so far reached the corporation, not the executive. No AI executive has been personally adjudicated as civilly or criminally liable for training-data, deployment, or evaluation laundering as of May 2026. The interceptions to date are corporate: Anthropic pays $1.5 billion; OpenAI rolls back and publishes a post-mortem; Meta's variant is delisted from the leaderboard; LMArena revises its policy. The five-role conflation diagnostic from chapter 7, applied to the AI stack, has not yet produced executive-level personal accountability in any of the three layers.

Last, the statutory delay. California's SB-942 was signed 19 September 2024; AB-853 pushed its operative date to 2 August 2026. The statutory interception layer in California is in the future tense as of manuscript freeze. The European AI Act's Article 53 obligations are partly operative across 2025 and 2026. The statutory surface is real and partly engaged. It is not yet load-bearing in the way the litigation cluster has been.

The teaching point is the asymmetry. Three AI-stack layers. Three interception forms. Three different durabilities. The legal-discovery channel produced the first court-documented acquisition record, the first nine-figure settlement, and the first court distinction between lawful and shadow-library training data. It has not yet produced executive-level personal accountability. The deployment-layer rollback and the evaluation-layer policy revision are faster, cheaper, and reach less. They leave the architecture in place. The interception works at each layer in its own register, and at no layer has it yet reached the executive whose displaced decision the alibi grammar was protecting.

Chapter 8 walked the institutional-power case — power calling itself the goat. We have walked the newest altar. Chapter 10 is the next stress-test in Part III: the war zone, where the cost-bearer is identifiable, the record-control architecture sits behind a uniformed border, and the diagnostic must climb past the highest legal heat we encounter.

The model-decided alibi shows up wherever an AI system produces an output that affects us and the company narrates it in the grammar of system agency. *The model decided.* *The algorithm flagged.* *The data showed.* *The benchmark scored.* The verb's subject is a thing.

We interact with these systems daily. Employer AI screening. Bank-AI denial. AI-generated content provenance. Insurance-AI pricing. Content-moderation decisions on the platforms we use. Healthcare-AI triage. Government-AI assessment for benefits eligibility, tax flagging, visa decisions. Recommendation engines that decide what news we see. The list is incomplete. The recognition trigger is one question. Did a non-human verb-subject just absorb a decision that affects me? If yes, the alibi is operating. Run the diagnostic three times — once at each layer.

The diagnostic markers, by layer, are concrete.

##### At the input layer.

The company will not name the training corpora or licensing status. Public statements use "publicly available data" or "data we licensed from partners" without specifying which works and from whom. The acquisition record is described as a trade secret while the output is positioned as transformative. These patterns appearing together are the diagnostic question, not the answer; we see the input-layer alibi shape and ask which displaced acquisition decision the grammar is protecting.

##### At the deployment layer.

Behaviour change between versions is not surfaced — the deployed model is described as "the model" with no version designation. The appeal or correction route, where one exists, returns to the same model with no human reviewer. The post-deployment optimisation framework — reward-shaping changes, fine-tuning targets — is not disclosed. These patterns appearing together are the diagnostic question; we see the deployment-layer alibi shape and ask which named release or reward-shaping decision the grammar is protecting.

##### At the evaluation layer.

Benchmark performance is cited in procurement or marketing materials as a property of the model, without specifying which variant was benchmarked. Performance gaps between the announced model and the shipped model are not disclosed. The benchmark organisation's procedural shell — how submissions are managed, what experimental variants are permitted, what disclosure providers owe — is not transparent. These patterns appearing together are the diagnostic question; we see the evaluation-layer alibi shape and ask which named submission decision the grammar is protecting.

The single sharpest move is the same at every layer: demand the three records in writing — the training-corpus acquisition log, the deployment reward-shaping log, the benchmark-submission record — and treat a denial as the diagnostic finding, not a dead end. Do not accept "the model is a black box" as a complete explanation of a decision that affects our employment, our credit, our insurance, our healthcare, our visa, our benefits, our account, or our published work.

The warning sign is concrete, and it is the moment the alibi becomes personal: a request to sign for, certify, or be the visible named owner of an AI output we cannot independently inspect. The human-in-the-loop signoff. The compliance attestation that an AI tool is "fit for purpose." The clinician's electronic confirmation of an AI diagnostic recommendation. The loan officer's countersignature on an AI-flagged transaction. The five-role conflation makes the human signer the seam: the lab built the model, the deployer ran it and holds the audit log and processes the appeal, and the contract makes us personally liable for endorsing the output. If we cannot inspect the training corpus, the deployment record, and the benchmark-submission record, we cannot legitimately sign for the output.

The signature is the model's alibi made personal — the same trap chapter 7 named at the Post Office counter and the Therac-25 console, now wearing the added inscrutability of statistical training. The same shape, new scale. The first anti-laundering device remains the record, and we have shown how to demand three of them.

A fourth layer is emerging — autonomous-action systems that chain model outputs into steps taken in the world. The diagnostic ports forward intact: a fourth record, the action log, joins the three we have demanded. When the agents arrive in our work, we already know what to ask.

References

[^307]: Sycophancy in GPT-4o: What happened and what we're doing about it. [primary URL no longer resolves as of 2026-05-28; canonical archive] [archived snapshot](http://web.archive.org/web/20260524131149/https://openai.com/index/sycophancy-in-gpt-4o/).
[^308]: Sam Altman X post acknowledging GPT-4o sycophancy. Archived at http://web.archive.org/web/20260411001817/https://x.com/sama/status/1916625892123742290.
[^309]: Ahmad Al-Dahle X post denying Meta trained Llama 4 on test sets. Archived at http://web.archive.org/web/20250523051535/https://x.com/Ahmad_Al_Dahle/status/1909302532306092107.
[^310]: Order directing OpenAI to preserve and segregate output log data, The New York Times Company v. Microsoft Corporation et al., Case No. 1:23-cv-11195 (S.D.N.Y.). [primary URL no longer resolves as of 2026-05-28; canonical archive] [archived snapshot](http://web.archive.org/web/20260416015333/https://docs.justia.com/cases/federal/district-courts/new-york/nysdce/1:2023cv11195/612697/551).
[^311]: Sam Altman X post acknowledging GPT-4o sycophancy. Archived at http://web.archive.org/web/20260411001817/https://x.com/sama/status/1916625892123742290.
[^312]: Ahmad Al-Dahle X post denying Meta trained Llama 4 on test sets. Archived at http://web.archive.org/web/20250523051535/https://x.com/Ahmad_Al_Dahle/status/1909302532306092107.
[^313]: LMArena public statement on Meta's Llama-4-Maverick submission and LMArena leaderboard policy revision. Archived at http://web.archive.org/web/20260526051543/https://simonwillison.net/2025/Apr/8/lmaren/.
[^314]: Bartz v. Anthropic PBC — preliminary approval of $1.5 billion class settlement and per-work compensation framework. [susmangodfrey.com](https://www.susmangodfrey.com/wins/susman-godfrey-secures-1-5-billion-settlement-in-landmark-ai-piracy-case/), [archived snapshot](http://web.archive.org/web/20260515143135/https://www.susmangodfrey.com/wins/susman-godfrey-secures-1-5-billion-settlement-in-landmark-ai-piracy-case/).
[^315]: Bartz v. Anthropic Fairness Hearing: Final Reminder, 91.3% Claims Rate, and updates from the Docket. [authorsalliance.org](https://www.authorsalliance.org/2026/05/14/bartz-v-anthropic-fairness-hearing-final-reminder-91-3-claims-rate-and-updates-from-the-docket/), [archived snapshot](http://web.archive.org/web/20260515051333/https://www.authorsalliance.org/2026/05/14/bartz-v-anthropic-fairness-hearing-final-reminder-91-3-claims-rate-and-updates-from-the-docket/).

Chapter 10 — War Is the Perfect Laundry

---
chapter_id: 10-war-is-the-perfect-laundry
chapter_number: 10
part: III
title: War Is the Perfect Laundry
evidence_grade: A
---

Chapter 10 — War Is the Perfect Laundry

> All warfare is based on deception.
>
> — Sun Tzu, *The Art of War* I.18, trans. Lionel Giles (1910)

On 5 February 2003, in the Security Council chamber at the United Nations, Secretary of State Colin Powell held up a small glass vial. The vial contained simulated anthrax. He held it between thumb and forefinger, near eye level, long enough for the cameras to find it. He said: "Less than a teaspoon full of dry anthrax in an envelope shut down the United States Senate in the fall of 2001."[¹](#c10-1).

The vial was a prop. It carried no live agent. It stood in for the case Powell had come to make: that Iraq possessed an active stockpile of biological and chemical weapons, was reconstituting its nuclear weapons programme, and had an operational relationship with al-Qaeda. Each of the three strands had its own evidentiary spine. The mobile bioweapons laboratories — the load-bearing image of the speech — came from a single Iraqi defector, code-named Curveball, who had been debriefed not by the CIA but by Germany's foreign intelligence service, the BND[²](#c10-2). The aluminum tubes, presented as centrifuge components for uranium enrichment, were contested inside the US intelligence community, with the State Department's Bureau of Intelligence and Research recording its dissent in text boxes set off from the main assessment and in Annex A[³](#c10-3). The Atta-Prague meeting, presented as evidence of operational ties between Iraq and al-Qaeda, was promoted to the speech through the Pentagon's Office of Special Plans rather than the standard intelligence-community vetting line[⁴](#c10-4).

Powell spoke for approximately seventy-five minutes. The presentation had been prepared over four days at CIA headquarters in Langley, with Powell and his chief of staff, Lawrence Wilkerson, working from a draft assembled by the Office of the Vice President's national-security staff and the National Security Council. Wilkerson has subsequently described that drafting in public statements: material was, he said, "pushed" into the speech from non-INR sources over INR objections[⁵](#c10-5).

The vial mattered because the vial was simulated. We return to that fact below. For now it stands as the visible centre of an evidentiary case that, eighteen months later, had collapsed.

The invasion began six weeks after the speech. By the end of 2004, the Iraq Survey Group — established by the CIA and led by Charles Duelfer — had concluded that Iraq's WMD programmes had been effectively dismantled by 1991 sanctions; no stockpiles existed; no reconstitution was underway[⁶](#c10-6). The image at the centre of Powell's speech had no physical referent. The vial had been simulated all along.

The intelligence was wrong

The phrase arrived quickly. By the autumn of 2003 it was a Washington commonplace; by the autumn of 2005 it was the explicit framing of the administration that had launched the war.

On 14 December 2005, at the Woodrow Wilson International Center for Scholars, President George W. Bush addressed the question directly. "It is true that much of the intelligence turned out to be wrong," he said[⁷](#c10-7). The line was repeated, in variations, across years of subsequent statements. The British government's parallel framing followed the same shape: Prime Minister Tony Blair characterised the British case as having rested on intelligence assessments he had taken as established and which had proved less established than they had appeared.

The framing performs a specific operation. It names a category — "the intelligence" — and assigns to that category the property that caused the collapse: it was wrong. The category is institutional. The intelligence community is an entity composed of agencies, analysts, products, and processes. The framing treats this entity as a single actor with a property. The intelligence community produced an assessment; the assessment was wrong; the war proceeded on the basis of the wrong assessment; the responsibility lies with the producing entity.

This framing has the shape of a confession. It accepts that something went wrong. It assigns the going-wrong to a nameable place. It does not assign the going-wrong to the person reciting the confession.

Chapter 8 showed institutions blaming themselves to narrow the diagnosis: the institution names itself as the failed component, the inquiry is fenced to that component, the case closes inside the fence. "The intelligence was wrong" is the same move at strategic scale. The executive branch named the intelligence community as the failed component. The first commission convened to investigate the failure — the Commission on the Intelligence Capabilities of the United States Regarding Weapons of Mass Destruction, commonly called Robb-Silberman after its co-chairs — was chartered by Executive Order 13328 on 6 February 2004. The order explicitly excluded examination of how policy-makers had used the intelligence[⁸](#c10-8). The fence was drawn before the inquiry began. The charter, written by the actor under examination, set the commission to audit how the intelligence had been collected, analysed, and disseminated, and gave it no authority to ask whether anyone had read the result.

Three sub-claims sat inside the strategic case, and each was subsequently found by formal inquiry to have exceeded what the underlying intelligence supported. First: Iraq possessed chemical and biological stockpiles. The 1 October 2002 National Intelligence Estimate had asserted this with "high confidence"[⁹](#c10-9); the Duelfer Report falsified it[¹⁰](#c10-10). Second: Iraq was reconstituting its nuclear weapons programme. The aluminum tubes were the headline evidence; the Niger yellowcake documents the secondary. At the Security Council on 7 March 2003, IAEA Director General Mohamed ElBaradei declared the Niger documents forgeries: "based on a thorough analysis, the IAEA has concluded, with the concurrence of outside experts, that these documents… are in fact not authentic"[¹¹](#c10-11). The invasion began twelve days later. Third: Iraq had an operational relationship with al-Qaeda; the Atta-Prague meeting was the documentary node. The Senate Select Committee on Intelligence's Phase II report, released on 25 May 2007, concluded that public statements on Iraq-al-Qaeda ties had exceeded the intelligence record[¹²](#c10-12).

Each of the three sub-claims was thus contested inside the intelligence record at the time it was made, then formally found to have exceeded the record after the war. The framing "the intelligence was wrong" treats the contestation and the post-hoc finding as a single property of a single actor. The framing is not a falsification. The framing is a redescription. It moves the question — was the public case warranted by the underlying assessments — into a different question — was the underlying intelligence right. The two questions are not the same. The diagnostic we have been installing across previous chapters distinguishes them.

Eight questions inside the war

The case has the standard shape: an apparent cause (the intelligence community), a public attribution (it got it wrong), and a residual question (then why did the war happen). The same eight questions from chapter 3 — blame, control, benefit, knowledge, preventability, record-control, cost, and what responsibility looks like if it followed control instead of visibility — answer the residual question here, applied to the war. Walking them through Iraq WMD locates the answers in places the "intelligence was wrong" framing does not point to.

Where the war case bends the instrument hardest is at control, benefit, and knowledge, because each names a war-specific architecture the peacetime cases did not contain; the rest run as they did in chapter 3, only as far as the war case carries them. The walk is full — all eight.

##### Who was publicly blamed.

The intelligence community as an institution. CIA Director George Tenet personally, via his reported remark to the President in December 2002 that the WMD case was "a slam dunk" — a remark sourced through Bob Woodward's *Plan of Attack* and contested in Tenet's own subsequent account[¹³](#c10-13). Curveball / Rafid Ahmed Alwan al-Janabi as a late-stage fabricator, after his identification by the Robb-Silberman Commission in 2005 and his public self-identification in 2011[¹⁴](#c10-14). Specific analytic offices within the CIA, DIA, and INR, named in the Robb-Silberman tradecraft critique[¹⁵](#c10-15). The blame distributed across an institutional category, a senior official's reported informal remark, a defector, and a set of line analysts.

##### Who had control.

Control over the public WMD case sat in three rooms: the White House Iraq Group (WHIG), a communications coordination body formed in August 2002 to shape the public case, whose existence was reported by *The New York Times* on 7 September 2002[¹⁶](#c10-16); the Office of the Vice President, where Vice President Richard Cheney (who died in 2024) and his chief of staff I. Lewis Libby ran the policy coordination; and 10 Downing Street, where Blair, his director of communications Alastair Campbell, and Joint Intelligence Committee chair John Scarlett worked on the parallel UK case. Control over the analytic product sat with Tenet and Deputy Director John McLaughlin at CIA, and with the National Intelligence Council, whose Chairman Robert L. Hutchings oversaw the NIE produced on a three-week timeline in September 2002[¹⁷](#c10-17). Control over the final UN presentation text sat with Powell (who died in 2021) and Wilkerson, working from a draft Wilkerson has stated had material pushed into it from OVP and NSC sources over INR objections[¹⁸](#c10-18).

##### Who benefited.

The framing produced UN Security Council Resolution 1441 on 8 November 2002[¹⁹](#c10-19); the Authorization for Use of Military Force Against Iraq, P.L. 107-243, on 16 October 2002[²⁰](#c10-20); and the domestic political coalition for the war. Career outcomes are documentary. Tenet resigned on 3 June 2004 and received the Presidential Medal of Freedom on 14 December 2004[²¹](#c10-21). Condoleezza Rice was promoted to Secretary of State on 26 January 2005. Stephen Hadley succeeded her as National Security Adviser. Douglas Feith remained Under Secretary of Defense for Policy until his departure in August 2005. None of the senior officials named in the SSCI Phase II report or the Chilcot Inquiry faced criminal process on this record.

##### Who had documented knowledge, per inquiry findings.

SSCI Phase II concluded that public statements on Iraq-al-Qaeda ties and the Atta-Prague meeting exceeded the intelligence record[²²](#c10-22). The Department of Defense Inspector General's report on the Office of Special Plans found that Feith's office had presented conclusions in IC-vetted form when they were not IC-vetted; the report found that the work "did not provide 'the most accurate analysis of intelligence to senior decisionmakers'" and that the conclusions were "inappropriate"[²³](#c10-23). Former CIA European Division chief Tyler Drumheller has stated, in Senate testimony and in his 2006 book *On the Brink*, that a German BND liaison warned Tenet personally that Curveball should not be considered reliable for public statements; Tenet has disputed whether the warning reached him before the Powell speech[²⁴](#c10-24). The Chilcot Inquiry found that the assessments and judgments underlying the September 2002 Dossier were presented with a "certainty that was not justified," that the Joint Intelligence Committee should have made clear to Mr Blair that the assessed intelligence had not established beyond doubt that Iraq had continued to produce chemical and biological weapons, and that the process by which the Dossier was prepared blurred the line between assessment and policy presentation[²⁵](#c10-25).

##### Who could have prevented recurrence.

This question runs as it did in chapter 3: it asks where, in the documented record, the chain had a discount opportunity it did not take. The war case supplies a sequence of them. The Drumheller-route Curveball warning, if received as Drumheller has stated. INR's objections to the aluminum tubes and the Niger claim, set off from the Key Judgments — in text boxes and Annex A — rather than carried as headline dissents[²⁶](#c10-26). ElBaradei's 7 March 2003 finding that the Niger documents were forgeries, on the public record before the invasion. Hans Blix's UNMOVIC reports through early 2003, finding cooperation increasing and no stockpiles located[²⁷](#c10-27). Each was a prevention window. The invasion proceeded on 19 March 2003 through all of them.

##### Who controlled the record.

Initially the executive branch, in three forms: the classification regime over the underlying intelligence reporting (the October 2002 NIE was declassified in heavily redacted form only on 18 July 2003); the WHIG synchronisation of public statements; the structural advantage that strategic-justification inquiries operate after the war they would adjudicate, while record-hardening operates during the precondition-shaping window. The decisive record breaks came from outside the executive: SSCI Phase I in July 2004; the Robb-Silberman Commission in March 2005, even within its narrowed charter; SSCI Phase II's two components in September 2006 and May 2007; the DoD Inspector General's report on OSP in February 2007; and the Chilcot Inquiry in July 2016. Investigative journalism contributed before any of the formal inquiries reported: Warren Strobel and Jonathan Landay at Knight-Ridder filed dissenting pre-invasion coverage that did not enter the major-paper consensus narrative[²⁸](#c10-28).

##### Who bore the cost.

Iraqi civilians, first.

The Iraq Body Count project, which records documented civilian deaths from violence based on cross-referenced media and official sources, recorded approximately 100,000 to 110,000 documented civilian deaths from 2003 to 2011[²⁹](#c10-29). Survey-based estimates ran higher. The Roberts et al. household survey published in *The Lancet* on 11 October 2006 estimated approximately 654,000 excess deaths through June 2006[³⁰](#c10-30). The Iraq Family Health Survey, published in *The New England Journal of Medicine* on 31 January 2008, estimated approximately 151,000 violent deaths from March 2003 to June 2006[³¹](#c10-31). We report the Iraq Body Count figure as documented floor and flag the methodology dispute on the survey-based estimates without averaging across methodologies that do not share priors.

Military deaths: per US Department of Defense casualty figures, approximately 4,491 US service members across Operation Iraqi Freedom and Operation New Dawn through 2011[³²](#c10-32); per UK Ministry of Defence figures, 179 UK service members[³³](#c10-33). Displacement: the UN High Commissioner for Refugees recorded approximately two million Iraqis displaced internally and approximately two million as refugees by 2007[³⁴](#c10-34). Iraq's institutional capacity, infrastructure, and historical archive — including the looting of the Iraq Museum in April 2003 and the de-Baathification of the professional civil service — were costs that fell on the population over decades.

Curveball, the pure-scapegoat node, absorbed the late-stage attribution of the bioweapons claim. The chain that had multiple opportunities to discount his account did not absorb its share.

##### What responsibility would look like if it followed control.

The senior actors who controlled the WHIG communications architecture, the OVP and NSC drafting process for Powell's speech, the OSP alternative-intelligence channel, the compressed-timeline NIE production, and the No. 10 dossier drafting would face individual accountability. Criminal accountability where the statute and jurisdiction reach. Public-record accountability where they do not. Statutory inquiry where the legal route is blocked. None has, on this record, as of mid-2026.

The walk yields a structural finding the framing "the intelligence was wrong" obscures. The walk located control in the executive branch and 10 Downing Street; the knowledge in the inquiry record; the cost in the population. It located the failure in the gap between the product and the public statements. "The intelligence was wrong" names none of these places. It names the one actor the walk found least able to set the public certainty — and assigns it the property that caused the collapse.

Return to the vial. It was a prop; it carried no live agent; everyone in the chamber knew it was simulated. The certainty around it was the same kind of object — a prop standing in for a case the inquiry record later found had been presented with, in Chilcot's words, "a certainty that was not justified." The vial held nothing. The walk locates the gap between the staged certainty and the hedged product in the rooms it just named, not in the analysts the public framing blamed.

Chapter 9 showed the model becoming the named cause. Here the model is the National Intelligence Estimate, and the same reification operates: an institutional product is treated as if it had agency, while the policy choice to commission it on a three-week timeline and to present its hedged language with "high confidence" framing is treated as background noise. The NIE became the war's model. The hedges inside it — including the State Department's INR dissent on the aluminum tubes and the nuclear claim, set off in a boxed "Alternative View" with its own text boxes and an Annex A — lived in the margins while the certainty lived in the Key Judgments. The boxed dissent was not in the speech.

Five layers, four callbacks

War launders responsibility through five layers, and only one of them is new here. Across the earlier chapters we installed four laundering patterns; war fires all four at once, plus a fifth — strategic justification — installed here through the Iraq WMD anchor. Layer 1 is the one just installed at the case level: the NIE as institutional product, the public certainty exceeding the product's hedges, the audit body fenced from policy use. Layers 2 through 5 are callbacks — patterns already installed earlier, re-firing in war. Re-walking them is not what we are here for. What we are here for is the demonstration that the four prior patterns and the strategic-justification pattern fire simultaneously in war, and that the diagnostic resolves each by climbing into the specific chain layer by layer. Each callback below pulls only the layer-observation new to the present argument.

The chain-of-command layer — Abu Ghraib

Abu Ghraib entered the argument in chapter 2 as a partial scapegoat. The night-shift military police of the 372nd MP Company were genuinely culpable; eleven enlisted soldiers and one brigadier general were sanctioned. From the 372nd MP Company: Specialist Charles Graner, Private Lynndie England, Staff Sergeant Ivan Frederick II, Sergeant Javal Davis, Specialist Sabrina Harman, Specialist Megan Ambuhl, and Specialist Jeremy Sivits were each court-martialled, with dispositions ranging from pleas to contested convictions and sentences from administrative discharge to ten years' confinement. From the military-intelligence side, Specialist Armin Cruz of the 325th Military Intelligence Battalion was separately court-martialled on assault and conspiracy charges arising from interrogation conduct at the prison. Brigadier General Janis Karpinski, commander of the 800th MP Brigade, was demoted to colonel and relieved of command in May 2005[³⁵](#c10-35). The laundering operated by stopping the chain at those convictions and that demotion. Above them sat the Office of Legal Counsel memoranda of 1 August 2002 (Bybee and Yoo) authorising coercive interrogation techniques; the 2 December 2002 Rumsfeld approval of those techniques for Guantanamo; the 16 April 2003 expansion; and Lieutenant General Ricardo Sanchez's 14 September 2003 authorisation that migrated Guantanamo-style techniques into Iraq[³⁶](#c10-36).

War adds a chain-of-command secrecy layer to the partial-scapegoat pattern. The distinguishing mark is the boundary: the chain stops not because the evidence runs out but because it crosses a classification line above which the public record cannot reach during the operative window. The Taguba Report of March 2004, the Schlesinger Report of August 2004, the Fay-Jones Report of August 2004, and the Senate Armed Services Committee report on detainee treatment of 20 November 2008 each pierced one layer; none reached adjudicated criminal liability above the rank of staff sergeant[³⁷](#c10-37). The Knapp Commission and Serpico counter-case installed in chapter 2 against the "bad apples" framing applies here at strategic scale: institutional resistance to chain-stopping has documented historical precedent. The chain-of-command secrecy layer is what makes the precedent harder to apply in wartime.

A second feature of the chain explains why it stops where it does, independent of classification. The legal architecture of the order-following obligation places the determination of an order's lawfulness on the individual at the bottom of the chain. Rome Statute Article 33 provides that superior orders are a defence to war crimes only where the order was not manifestly unlawful, the subordinate was under a legal obligation to obey, and did not know the order was unlawful; Article 33(2) declares that orders to commit genocide or crimes against humanity are manifestly unlawful as a matter of law. The US Uniform Code of Military Justice carries a parallel doctrine: lawful orders must be obeyed; manifestly illegal orders are not a defence, and obedience to a manifestly illegal order is itself an offence. The architecture cuts in one direction at the seam — it makes the enlisted soldier answerable for executing a manifestly unlawful order — without, by its own terms, carrying liability upward to whoever issued the doctrine the order rested on. That asymmetry is part of why the staff-sergeant-level dispositions could close the case while the authorisation memoranda above them stayed unadjudicated.

The proxy layer — MH17

The 17 July 2014 destruction of Malaysia Airlines flight MH17 was read in chapter 4 as a partial scapegoat with a system / object alibi secondary across a state border. The Joint Investigation Team led by the Dutch Public Prosecution Service established that the Buk-TELAR missile launcher that fired the missile belonged to the Russian Federation's 53rd Anti-Aircraft Missile Brigade based in Kursk, and traced the launcher's outbound and inbound paths through Russian territory[³⁸](#c10-38). Three individuals — Igor Girkin, Sergey Dubinskiy, and Leonid Kharchenko — were convicted in absentia at The Hague District Court at Schiphol on 17 November 2022 for the murder of all 298 persons aboard[³⁹](#c10-39). President Vladimir Putin has publicly denied Russian state involvement in the launcher's supply chain, including in remarks at a press conference on 21 July 2014[⁴⁰](#c10-40). The JIT's 8 February 2023 update concluded there were "strong indications" that President Putin had personally approved the provision of the Buk system to the armed groups in eastern Ukraine[⁴¹](#c10-41). The European Court of Human Rights inter-state proceedings *Ukraine and Netherlands v Russia* remain pending on the merits as of mid-2026.

The proxy layer is the element that did not appear in chapter 4: a state hides behind a non-state actor it supplies. The "DPR" legal-political shell, the alternative-cause framings — a Ukrainian Su-25, a Buk from Ukrainian stocks — circulated through Russian state media, and the non-recognition of foreign criminal jurisdiction together produce a deniability architecture that operates independently of the underlying evidence. The Hague judgment resolved individual culpability at one layer. State-actor attribution at the proxy-supply-chain level remains procedurally pending, anchored as of mid-2026 to the JIT's "strong indications" formulation and the ECtHR's active docket. The architecture does not become inoperative because the evidence is good. The architecture stays operative until a body with jurisdiction completes its adjudication; the convicted individuals have not been surrendered.

The contractor-shell layer — Blackwater at Nisour Square

Chapter 4 also carried the 16 September 2007 shooting at Nisour Square in Baghdad, read there as a partial scapegoat with a system / object alibi secondary. Four convoy guards from the private military company Blackwater — Nicholas Slatten, Paul Slough, Evan Liberty, Dustin Heard — fired weapons that killed seventeen Iraqi civilians, including nine-year-old Ali Kinani, who was sitting in the back seat of his father's car, according to a US Department of Justice indictment and State Department reports on the incident[⁴²](#c10-42). In *United States v. Slatten* and *United States v. Slough et al.*, the four guards were convicted in the US District Court for the District of Columbia in 2014: Slatten of first-degree murder; Slough, Liberty, and Heard of voluntary manslaughter, attempted manslaughter, and weapons charges[⁴³](#c10-43). They were pardoned by President Trump on 22 December 2020[⁴⁴](#c10-44). The laundering operated above them at two levels: Coalition Provisional Authority Order 17, revised on 27 June 2004, granted contractors immunity from Iraqi criminal jurisdiction[⁴⁵](#c10-45); the State Department's contractual relationship with Blackwater (later Xe Services, later Academi) insulated the corporate principal from operational accountability.

Here the added layer is a legal-status shell — a category that exempts a known actor from jurisdiction, where the proxy layer hid the actor entirely. The "contractor" category is not a criminal-immunity category in international law, but it operated as one in the operational moment. The 22 December 2020 pardons of the four convicted guards then reversed the individual-level accountability that had survived the shell. The shell launders the operation; the pardon launders the conviction; the chain of command above the contractor was never criminally examined. Erik Prince, Blackwater's founder, has spoken publicly about the company's role in subsequent interviews and writings; the public record does not show a criminal charge against him on the Nisour Square matter as of mid-2026[⁴⁶](#c10-46).

The civilian-invisibility layer — Ukrainian children

The Russian transfers of Ukrainian children from occupied territories came into the argument in chapter 2 as a cost-bearing-goat case where the laundering operated through language. "Evacuation," "humanitarian relocation," and "guardianship" reframed forcible transfer as care.

Russian Presidential Decree No. 330 of 30 May 2022 and Decree No. 585 of 11 July 2022 created an accelerated citizenship pathway for children from Ukrainian orphanages and care institutions in territories under Russian occupation[⁴⁷](#c10-47). President Putin and Russia's Commissioner for Children's Rights, Maria Lvova-Belova, have publicly defended the transfers as humanitarian — for example, in statements broadcast on Russian state television during 2022 and 2023[⁴⁸](#c10-48).

On 17 March 2023, the International Criminal Court's Pre-Trial Chamber II issued arrest warrants for President Putin and Commissioner Lvova-Belova, finding reasonable grounds to believe both bear individual criminal responsibility for the alleged war crime of unlawful deportation of children from Ukrainian territory to the Russian Federation[⁴⁹](#c10-49). The Russian Federation does not recognise the jurisdiction of the International Criminal Court and has stated it will not transfer the named individuals; no trial has been held as of mid-2026.

The civilian-invisibility layer works through euphemism, which the diagnostic resolves only at the level of the underlying conduct, not the language. The cost-bearer category in wartime is reframed faster than blame is moved: the children are not accused of anything, but the linguistic apparatus presents the transfer as care while the transfer continues. The ICC's reasonable-grounds finding establishes the legal mechanism but does not by itself pierce the language. Piercing the language requires naming each euphemism against the underlying conduct. "Evacuation" is named against forcible transfer of a protected population. "Guardianship" is named against acquisition of citizenship pathway under occupying-power authority. "Humanitarian relocation" is named against the absence of mechanisms for return. The Stolen Generations counter-case installed in chapter 2 against cost-bearing-goat euphemism applies here at the same diagnostic layer; chapter 2 did the work.

Five layers, fired at once

The callbacks together establish a load-bearing generalisation. War does not run its laundering patterns one at a time; it fires the whole stack simultaneously. Layer 1 is strategic-justification reification, installed here through the anchor. Layers 2 through 5 are the patterns installed across the preceding chapters and re-fired in war: chain-of-command secrecy (Abu Ghraib), proxy deniability (MH17), contractor legal-status shell (Blackwater), and civilian-invisibility euphemism (Ukrainian children). Five layers fire at once. The diagnostic resolves each by climbing into the specific chain question by question.

> *The five layers as a table; the walk above is its audio rendering, and listeners may skip the table.*

::: {.figure-embed}
[〔图：War stacks five laundering layers that fire at once; the diagnostic resolves each by climbing into it, not by granting war an exemption.〕](https://github.com/xiaolai/no-one-did-it/blob/b5e9c8f820bc04bb1d452d4109318673c08c3945/book/chapters-v6/book/evidence/diagrams/proofs/ch10-war-stack.png)
:::

Classification during the window

The strategic-justification record was not produced in a vacuum. It was produced inside a record-control architecture, operating during the operative window of the precondition-shaping decisions, that consistently advantaged the public case and disadvantaged the underlying contestation.

Three elements of that architecture are documentary. First, executive-branch classification of the underlying intelligence reporting. The October 2002 NIE was declassified in heavily redacted form only on 18 July 2003 — four months after the invasion and three months after Powell's UN speech. The public was asked to evaluate certainty without access to the hedges that qualified it. Second, the WHIG communications coordination synchronised public statements across the White House, the Pentagon, the State Department, and 10 Downing Street, producing a unified message whose internal contestation was inaccessible to outside auditors during the operative window[⁵⁰](#c10-50). Third, the structural advantage: strategic-justification inquiries operate after the war they would adjudicate. The record-hardening operates during the precondition-shaping window; the inquiries operate later, when the war is a historical fact and the policy choices it depended on have already taken effect.

The same record-hardening pattern operates in each callback layer, with different mechanisms. The Abu Ghraib chain hardened its record through the classification of the OLC and Pentagon authorisation memos. The Bybee memo of 1 August 2002 was declassified only after Office of Legal Counsel head Jack Goldsmith and his successor Daniel Levin withdrew it on 30 December 2004[⁵¹](#c10-51). Many Pentagon authorisation memos were declassified only via the Senate Armed Services Committee inquiry that reported in November 2008. MH17's record was hardened by a Russian state-media counter-record, broadcast through RT and Sputnik and other outlets, operating independently of the JIT and inaccessible to ECtHR proceedings as positively probative of state non-involvement. Blackwater's record was hardened by CPA Order 17 as record-shielding within occupation law; the contractor accountability documentation that would have followed Iraqi criminal jurisdiction never existed because the jurisdiction was excluded. The Ukrainian-children record was hardened by the Russian decrees that framed the transfers as legal under Russian domestic law and by the Russian non-recognition of ICC jurisdiction.

Each layer's record-hardening differs in form. Each layer's record-hardening operates during the operative window. Each layer's record-hardening is challenged only retrospectively, by mechanisms whose charters and jurisdictions reach into the chain at different depths.

In chapter 7 we installed the record-control diagnostic and saw it operate at artifact, institutional, and state scale. War operates the same mechanism at strategic scale: the record is hardened by classification during the operative window and softened by post-hoc inquiry whose charter typically excludes the policy use of the record.

The steelman of shared priors

The resistance to the "intelligence was wrong" framing did not, in the operative window, come from the named goats. It came from the same outside-executive actors the walk already met — the Knight-Ridder reporters Strobel and Landay[⁵²](#c10-52), Powell's chief of staff Lawrence Wilkerson[⁵³](#c10-53), the CIA's Tyler Drumheller[⁵⁴](#c10-54), and, belatedly, Curveball himself[⁵⁵](#c10-55) — whose dissents were available, marginalised, or unread. The steelman has to reckon with one document that is sharper than any of them. The Downing Street Memo, dated 23 July 2002 and published by *The Sunday Times* on 1 May 2005, recorded MI6 chief Sir Richard Dearlove's observation that "the intelligence and facts were being fixed around the policy"[⁵⁶](#c10-56) — written eight months before the invasion, by the head of the service whose product was being fixed. Each of these resistances, the Memo included, reached the public only after the operative window had closed.

The cost of resistance was visible. Wilkerson's career did not advance after Powell's departure. Drumheller's role at CIA ended. The Knight-Ridder reporters did not join the major-paper consensus, and were structurally marginalised in the pre-war coverage period. The Plame disclosure, the Niger forgery sequence, and the subsequent Wilson opinion piece sit in this slot as a specific instance of resistance met with retaliation — the disclosure of the identity of an undercover CIA officer following her husband's public dissent from the Niger yellowcake claim. The pattern is consistent across the resistance set: documented, technically available, marginalised in the operative window.

Now the strongest counterargument. It is that the assessments reflected genuine uncertainty rather than a manufactured case, and the policy choice to act on that uncertainty is contestable on the merits but not laundering.

Stated at its strongest: many intelligence services believed Iraq retained some WMD capability as of late 2002. The German BND continued to debrief Curveball; the French DGSE assessed continuing Iraqi interest in WMD though without the certainty of the US position; the Russian SVR's assessments are partially documented and were not assessments of zero capability. Iraq's documented history of WMD use — the chemical weapons attacks on Iranian forces in the 1980s and on Kurdish civilians at Halabja in 1988 — gave foreign services a reasonable prior that programmes could persist. Iraq's post-1991 obstruction of UNSCOM and UNMOVIC's predecessor inspections gave them a reasonable prior that obstruction would conceal continuing activity. After his capture, Saddam Hussein told FBI agent George Piro, in interviews conducted in 2004 and declassified in July 2009, that he had cultivated WMD ambiguity to deter Iran[⁵⁷](#c10-57). The counterargument concludes: the prior was widely shared; the inference from the prior to active stockpiles was reasonable under uncertainty; the policy choice to act under that uncertainty is contestable on the merits but is not laundering, because nothing was hidden that the policy-makers had documented as hidden.

The counterargument is strongest on the prior. The foreign-intelligence consensus that Iraqi WMD activity persisted is documented and was not unanimous noise. Saddam's deterrence-ambiguity strategy, as Piro recorded it, is real and explains part of the chain.

The counterargument is weakest on the gap between the prior and the public statements. Our claim is not that the underlying assessments were certainly fabricated. Our claim is that the gap between the underlying assessments and the public statements was systematically widened by policy-makers and then attributed back to the intelligence community after collapse. SSCI Phase II and Chilcot are the load-bearing documents because they specifically measure that gap. Phase II concluded that public statements on Iraq-al-Qaeda and the Atta-Prague meeting exceeded the intelligence record. Chilcot concluded that the JIC assessments on which the UK public case rested had been presented with "a degree of certainty that was not justified"[⁵⁸](#c10-58). The widespread foreign-intelligence consensus argument is partially correct on the prior; it does not reach the public-statements-versus-intelligence-record gap that the inquiries establish.

This is the laundering charge. It attaches to the gap, not to the prior. It attaches to the system / object alibi — "the IC got it wrong" — that absorbed the policy choice — "we presented certainty the IC's product did not support." It attaches to the chartering of an IC-only audit body — Robb-Silberman — that was structurally prevented from examining the gap.

A second counterargument follows. War is genuinely different, this argument runs, from peacetime: secrecy is legitimate; the chain of command is the operational architecture; civilian casualty is inherent to armed conflict under the laws of war themselves; sovereign immunity and non-recognition of jurisdiction limit the reach of inquiry; the absence of post-event remedy infrastructure means the diagnostic runs and produces nothing actionable. We have built the diagnostic across peacetime cases; the argument is that the cases do not transfer.

The argument is strongest in naming what it names. Each of the structural conditions it lists is real. Secrecy is legitimate in defined wartime functions. The chain of command is the operational architecture, not an obstacle to it. Civilian casualty is inherent to armed conflict; international humanitarian law does not prohibit war, it regulates it. Sovereign immunity and non-recognition limit jurisdiction. The post-event remedy infrastructure is partial and uneven.

The argument is weakest where it concludes that the diagnostic produces nothing actionable. The Iraq Survey Group's Duelfer Report is an actionable finding: it established that the stockpiles did not exist. SSCI Phase II is an actionable finding: it established the public-statements gap. The DoD Inspector General's report on the Office of Special Plans is an actionable finding: it established the IC-vetted misrepresentation. The Chilcot Report is an actionable finding: it established that the legal basis for war was "far from satisfactory"[⁵⁹](#c10-59). The Hague District Court at Schiphol's MH17 judgment is an actionable finding: it convicted three individuals in absentia. The actionability of these findings does not produce remedy. They are operative findings of our diagnostic. The diagnostic does not deliver remedy in war; the diagnostic delivers the record on which future remedy can be built. We turn next to that record.

Thirteen years to Chilcot

The "intelligence was wrong" alibi did not weaken in a single moment. It weakened across thirteen years, through a sequence of documentary instruments, and the weakening's most important feature is not the conclusion of any single instrument but the architecture that produced them.

The Iraq Survey Group reported on 30 September 2004, with addenda in March 2005. The forensic floor: no stockpiles; no active programmes; the strategic justification's central empirical claim was false[⁶⁰](#c10-60). The Senate Select Committee on Intelligence reported Phase I on 7 July 2004, concluding that the NIE's key judgments were "either overstated, or were not supported by, the underlying intelligence reporting"[⁶¹](#c10-61). The Robb-Silberman Commission reported on 31 March 2005; within its charter, it identified Curveball as a fabricator and documented the chain-of-handling failures on the bioweapons claim[⁶²](#c10-62). SSCI Phase II reported in two components: "Postwar Findings About Iraq's WMD Programs and Links to Terrorism" on 8 September 2006, and "Report on Whether Public Statements Regarding Iraq by U.S. Government Officials Were Substantiated by Intelligence Information" on 25 May 2007[⁶³](#c10-63). The DoD Inspector General's report on the Office of Special Plans was completed on 9 February 2007[⁶⁴](#c10-64). Curveball self-identified in *The Guardian* on 15 February 2011[⁶⁵](#c10-65). The Chilcot Inquiry, announced on 15 June 2009, reported on 6 July 2016 with twelve volumes and approximately 2.6 million words; its core findings were that judgments about Iraq's WMD capabilities "were presented with a certainty that was not justified," that the September Dossier presented the JIC assessments with "a degree of certainty that was not justified," and that the legal basis for war was "far from satisfactory"[⁶⁶](#c10-66).

Thirteen years from invasion to Chilcot. The alibi, even as it weakened, operated for thirteen years. The weakening did not produce prosecutions. It did not unwind the war. It did not deliver remedy to the cost-bearers. We do not soften this:

> War's complexity is the design that makes responsibility hard to assign; the diagnostic survives the design because the design is what the diagnostic was built to name.

The complexity is not an obstacle to the diagnostic. The complexity is the laundering surface the diagnostic was constructed to identify. Each layer's structural condition — strategic-justification reification, chain-of-command secrecy, proxy deniability, contractor legal-status shells, civilian-invisibility euphemism — is, in the diagnostic's vocabulary, the specific architecture of a specific laundering operation. Naming the architecture is naming the operation. Climbing the chain with the eight questions resolves the operation into its constituent control, benefit, knowledge, preventability, record-control, and cost. The resolution does not produce remedy. The resolution produces the record.

The record is the precondition of remedy. Without the record, the remedy is impossible; with the record, the remedy may or may not arrive. We are precise about which is which.

Refuse the domain's claim to exception

Our rule follows directly from the layer-stack and the alibi-weakening sequence:

> Where strategic justification stacks chain-of-command secrecy, proxy attribution, contractor shells, civilian-invisibility euphemism, and institutional-product reification, the diagnostic still resolves the case by climbing into each layer with the eight questions. Treat any claim that responsibility in war is irrecoverable as the laundering mechanism itself, and refuse the domain's claim to exception.

The rule is portable: it holds whatever the war, whatever the configuration the layers fire in, and whatever the strategic-justification product — an NIE, an expert assessment, a model output, an intelligence summary, a public-statement schedule, a press background briefing. The five-layer stack is the take-home form; the eight questions are the per-layer protocol.

The rule is a refusal, not an accusation. It refuses the claim, made on every relevant occasion in every relevant domain, that this domain's exceptional conditions make the diagnostic inapplicable. It does not pre-assign blame. It does not pre-determine the case's classification. It runs the diagnostic, layer by layer, and lets the chain speak.

What made the inquiry work

Interceptions do exist. They are partial, uneven, layered onto the same five-layer stack the laundering operates across, and they reach different depths in different layers. Our load-bearing interception is the UK Iraq Inquiry under Sir John Chilcot.

Chilcot was announced on 15 June 2009 and reported on 6 July 2016 — seven years from announcement to report, thirteen years from invasion. It operated under the UK Inquiries Act 2005 as a statutory public inquiry. Its load-bearing institutional features are four, and the four together explain why this inquiry reached findings the US inquiries did not.

First, statutory subpoena power. Chilcot had access to UK Cabinet papers, JIC assessments, the Manning-Rice memos, and the Bush-Blair correspondence — including the 28 July 2002 minute in which Blair wrote, "I will be with you, whatever," eight months before the invasion[⁶⁷](#c10-67). Cabinet papers are presumptively protected; the Inquiry's statutory access pierced that protection.

Second, charter breadth. The Inquiry was empowered to examine the policy use of intelligence, not only the intelligence itself. This distinguished Chilcot from Robb-Silberman, whose founding Executive Order 13328 of 6 February 2004 explicitly excluded examination of how policy-makers used the intelligence[⁶⁸](#c10-68). Robb-Silberman's charter fenced the alibi; Chilcot's charter did not.

Third, political distance from the original decisions. Chilcot was a privy counsellor; the inquiry was chaired by him and conducted alongside four other panel members; the chair was appointed by Prime Minister Gordon Brown six years after the invasion; the inquiry continued through the May 2010 change of UK government. Distance produced an independence an immediate-aftermath inquiry could not have produced.

Fourth, time horizon. Seven years from announcement to report. The horizon meant the documents were available, the witnesses could not plead immediate-event memory limitations, and the inquiry could trace the chain across the full pre-war, war, and post-war windows.

These four features produced the findings already set out above — the unjustified certainty, the dossier, the "far from satisfactory" legal basis — and one more: the UK decision to commit to military action was effectively taken by Blair ahead of the formal Cabinet process (Executive Summary, ¶¶ 374–432).

The Hague District Court at Schiphol's MH17 judgment of 17 November 2022 demonstrates a different institutional response to a different layer of war-laundering: criminal adjudication of individuals inside a proxy chain. It convicted three individuals in absentia; it has not delivered them. The European Court of Human Rights's *Ukraine and Netherlands v Russia* proceedings demonstrate inter-state adjudication of state responsibility; the case is pending on the merits as of mid-2026. Neither reaches the strategic-justification layer that Chilcot reached. Neither has Chilcot's charter breadth. Together with Chilcot they establish that interception mechanisms exist at multiple war-laundering layers and operate under different procedural forms.

We are honest about the cost of the escape.

The strategic-justification alibi operated, even as it weakened, for thirteen years before formal documentary closure. SSCI Phase II's findings, the DoD IG OSP report's findings, and Chilcot's findings are inquiry conclusions, not adjudicated liability. We cannot characterise senior officials as criminally responsible for the strategic deception absent specific adjudications that do not exist as of mid-2026. The invasion is a historical event; no inquiry has reversed it. Iraqi civilians killed in the invasion and occupation, UK and coalition military families, displaced Iraqis: Chilcot did not deliver compensation, did not reverse displacement, did not prosecute responsibility-holders. The inquiry's product was the record, not the remedy.

A hostile reader will note an asymmetry we must face directly. The four interceptions named here — Chilcot, Schiphol, the ECtHR, the ICC — all operate inside or against open political systems and against states whose archives, courts, parliaments, and inquiry statutes can be turned on their own conduct. No Chilcot-equivalent has ever issued from the Russian Federation regarding Chechnya, Syria, or the 2022 invasion of Ukraine. No equivalent has issued from the People's Republic of China regarding Xinjiang or Tibet. No equivalent has issued from the Islamic Republic of Iran regarding its proxies' conduct in Iraq, Syria, Yemen, or Lebanon. The asymmetry is real and we do not paper it over: the interception architectures named here are conditional on institutional preconditions — independent judiciaries, statutory inquiry charters, archival access, parliamentary subpoena power — that closed authoritarian systems are designed to deny. The diagnostic still resolves those cases at the level of the eight questions; what is unavailable is the domestic interception step. Our argument is that the same architecture-preconditions that make Chilcot possible in the UK are the architecture-preconditions whose erosion later chapters name in the US. The asymmetry is not a defence of the case we analyse. The asymmetry is the reason the case we analyse matters: where the preconditions exist, their use is contested; where they do not exist, the laundering is total.

The structural choice to scope strategic-justification inquiries to intelligence-product rather than policy-use has not been reformed in US practice; the Robb-Silberman charter precedent remains in US executive-order practice. And across the other four layers the interception coverage is partial, uneven, and reaches different ranks — the in-absentia operators at Schiphol, the enlisted soldiers at Abu Ghraib, the contractors at Nisour Square whose convictions were then pardoned, the unexecutable ICC warrants — every one of them stopping short of the controlling decision a layer up.

Chilcot exists because the Inquiries Act 2005 gave a privy counsellor subpoena power over Cabinet papers and a charter broad enough to ask how policy-makers used the intelligence. The inquiry took seven years to produce. It produced the record. It produced no prosecutions, no unwinding, and no remedy. The architecture that made the inquiry necessary remains in place — both in the UK, where the Inquiries Act covers the next instance, and in the US, where no equivalent charter has been written.

The Knapp Commission and Serpico counter-case installed in chapter 2 against the partial-scapegoat "bad apples" framing applies here at strategic scale: institutional resistance can be designed, documented, and sustained. The Stolen Generations counter installed in chapter 2 against cost-bearing-goat euphemism applies here in the Ukrainian-children layer: linguistic apparatus can be named, decomposed, and matched against the underlying conduct. The diagnostic does not, by itself, produce remedy in war. The diagnostic produces the record on which future architectures of remedy can rest.

The five-layer stack shows up in coverage and policy debate; it does not show up in headlines labelled "laundering layer." Recognising it begins with the layers themselves.

The strategic-justification layer is operating when officials cite intelligence or expert assessments with high certainty in public, where the underlying products are unavailable or contested, and where the institutional product is treated as agentic — "the intelligence says," "the assessments show" — while the policy choice to commission and present that product is in the background. The chain-of-command layer is operating when a wartime incident is publicly blamed on enlisted personnel or a unit commander while doctrine memos, theatre commander authorisations, and interrogation-technique approvals remain classified during the operative window. The proxy layer is operating when operations are conducted by entities with deniable legal status — "armed groups," "people's republics," "local forces," "irregulars" — while sovereign-state command of supply, training, and recovery is publicly denied. The contractor-shell layer is operating when private military, private surveillance, private interrogation, or private logistics operate under contractual arrangements that exclude criminal jurisdiction and are described as administrative convenience rather than as accountability architecture. The civilian-invisibility layer is operating when cost-bearers — civilians, refugees, transferred children, displaced populations — are described in care-language ("evacuated," "rescued," "relocated") or aggregated into casualty figures whose methodology is contested without being adjudicated.

Recognition is a count, not a verdict. In a war or statecraft case we are being asked to evaluate, we can ask which of these five layers is in operation. Where three or more appear together, the case carries the structural shape we have been describing — and the eight questions are the per-layer tool we already have.

Three observable markers diagnose the strategic-justification layer specifically. First, the institution making the public case — the executive branch or cabinet — is presenting an institutional product it commissioned. Second, the body that audits the product is barred by charter or practice from examining the policy use of the product. Third, the public certainty exceeds the hedges in the underlying product, and the hedges are inaccessible to the public during the operative window. When two or more of these markers are present, the diagnostic is active. We do not need to know what the underlying product says. We need to know who is allowed to ask whether the public statement matches the product.

We are most likely to be positioned as goat at the layer our own role intersects. The enlisted soldier meets the manifest-illegality limit set out above; the analyst's hedges are stripped from the public statement; the contractor sits inside a legal-status shell that voids jurisdiction in the operational moment; the civilian's harm is renamed into care-language — "evacuation," "guardianship," "humanitarian relocation" — over the underlying forcible transfer. Each seam is the same shape: the cost lands where the role sits, while the controlling decision stays a layer up.

The diagnostic was built to work where it is hardest to use. The hardest domain is the one where every laundering layer we have installed fires together, where the operative window is closed by classification, where the audit charters are fenced by executive order, where jurisdiction is selective, where the cost-bearers are aggregated into figures whose methodology is contested, and where the inquiry, when it arrives, takes thirteen years and produces the record but not the remedy. It survives even there. We leave holding the five-layer stack and the eight questions — and the seam, for each of us, is the layer our own role intersects.

The next part installs the rules.

References

[^316]: United Nations Security Council, Provisional Verbatim Record S/PV.4701, 4701st meeting, 5 February 2003 (statement by Mr. Colin L. Powell, Secretary of State of the United States of America), S/PV.4701, pp. 2-17 (Powell statement). [undocs.org](https://undocs.org/en/S/PV.4701), [archived snapshot](https://web.archive.org/web/20260527030425/https://docs.un.org/en/S/PV.4701).
[^317]: Commission on the Intelligence Capabilities of the United States Regarding Weapons of Mass Destruction (Robb-Silberman Commission), Report to the President of the United States, 31 March 2005, Robb-Silberman Report, Letter of Transmittal and Chapter 1 (Iraq). [govinfo.library.unt.edu](https://govinfo.library.unt.edu/wmd/report/index.html), [archived snapshot](http://web.archive.org/web/20250421105951/https://govinfo.library.unt.edu/wmd/report/index.html).
[^318]: U.S. Senate Select Committee on Intelligence, Report on the U.S. Intelligence Community's Prewar Intelligence Assessments on Iraq (S. Rep. 108-301), 108th Cong., 2d Sess., 7 July 2004, S. Rep. 108-301, Conclusion 1 (Key Judgments) and accompanying findings. [govinfo.gov](https://www.govinfo.gov/content/pkg/CRPT-108srpt301/pdf/CRPT-108srpt301.pdf).
[^319]: U.S. Department of Defense, Office of the Inspector General, Report No. 07-INTEL-04, 'Review of the Pre-Iraqi War Activities of the Office of the Under Secretary of Defense for Policy', 9 February 2007, DoD IG Report 07-INTEL-04, Executive Summary; verbatim 'inappropriate' characterisation per report findings. [globalsecurity.org](https://www.globalsecurity.org/intell/library/reports/2007/dod-ig_07-intel-04_070209.htm), [archived snapshot](http://web.archive.org/web/20260526075519/https://www.globalsecurity.org/intell/library/reports/2007/dod-ig_07-intel-04_070209.htm).
[^320]: Lawrence Wilkerson, public statements on the preparation of Colin Powell's 5 February 2003 UN Security Council presentation, including interview broadcast on PBS Frontline, "The Dark Side" (20 June 2006), and interview broadcast on CNN, "Late Edition with Wolf Blitzer" (19 August 2007), PBS Frontline broadcast 20 Jun 2006; CNN Late Edition broadcast 19 Aug 2007. [pbs.org](https://www.pbs.org/wgbh/pages/frontline/darkside/), [archived snapshot](https://web.archive.org/web/20260402162428/http://www.pbs.org/wgbh/pages/frontline/darkside/).
[^321]: Iraq Survey Group, Comprehensive Report of the Special Advisor to the Director of Central Intelligence on Iraq's Weapons of Mass Destruction (Duelfer Report), 30 September 2004, with addenda March 2005, Duelfer Report, Volume I (Regime Strategic Intent) and Volume III (Iraq's WMD). [govinfo.gov](https://www.govinfo.gov/content/pkg/GPO-DUELFERREPORT/pdf/GPO-DUELFERREPORT-1.pdf).
[^322]: President George W. Bush, Address on Iraq and the War on Terror, Woodrow Wilson International Center for Scholars, Washington, D.C., 14 December 2005 (White House transcript), White House transcript, 14 Dec 2005, WWICS remarks. [georgewbush-whitehouse.archives.gov](https://georgewbush-whitehouse.archives.gov/news/releases/2005/12/20051214-1.html), [archived snapshot](http://web.archive.org/web/20260323161028/https://georgewbush-whitehouse.archives.gov/news/releases/2005/12/20051214-1.html).
[^323]: Executive Order 13328 of February 6, 2004, "Commission on the Intelligence Capabilities of the United States Regarding Weapons of Mass Destruction," 69 Fed. Reg. 6901 (Feb. 11, 2004), 69 Fed. Reg. 6901; Sections 2 and 3 of the Order. [govinfo.gov](https://www.govinfo.gov/content/pkg/WCPD-2004-02-09/pdf/WCPD-2004-02-09-Pg203.pdf), [archived snapshot](https://web.archive.org/web/20260527024057/https://www.govinfo.gov/content/pkg/WCPD-2004-02-09/pdf/WCPD-2004-02-09-Pg203.pdf).
[^324]: National Intelligence Council, National Intelligence Estimate 2002-16HC, 'Iraq's Continuing Programs for Weapons of Mass Destruction', October 2002 (declassified key judgments and excerpts released 18 July 2003), NIE 2002-16HC, Key Judgments (declassified excerpts, pp. 5-9); INR dissent set off from the Key Judgments in text boxes and Annex A. [irp.fas.org](https://irp.fas.org/cia/product/iraq-wmd-nie.pdf), [archived snapshot](http://web.archive.org/web/20250812015736/https://irp.fas.org/cia/product/iraq-wmd-nie.pdf).
[^325]: Iraq Survey Group, Comprehensive Report of the Special Advisor to the Director of Central Intelligence on Iraq's Weapons of Mass Destruction (Duelfer Report), 30 September 2004, with addenda March 2005, Duelfer Report, Volume I (Regime Strategic Intent) and Volume III (Iraq's WMD). [govinfo.gov](https://www.govinfo.gov/content/pkg/GPO-DUELFERREPORT/pdf/GPO-DUELFERREPORT-1.pdf).
[^326]: Mohamed ElBaradei, Director General, International Atomic Energy Agency, 'The Status of Nuclear Inspections in Iraq: An Update', statement to the United Nations Security Council, 7 March 2003, ElBaradei UNSC statement, 7 March 2003 (IAEA published transcript). [primary URL no longer resolves as of 2026-05-28; canonical archive] [archived snapshot](http://web.archive.org/web/20260216025722/https://www.iaea.org/newscenter/statements/status-nuclear-inspections-iraq-update).
[^327]: U.S. Senate Select Committee on Intelligence, "Postwar Findings About Iraq's WMD Programs and Links to Terrorism and How They Compare with Prewar Assessments" (S. Rep. 109-331, 8 September 2006), and "Report on Whether Public Statements Regarding Iraq by U.S. Government Officials Were Substantiated by Intelligence Information" (S. Rep. 110-345, 25 May 2007 / released June 2008), S. Rep. 109-331 (8 Sep 2006); S. Rep. 110-345 (25 May 2007 / released June 2008). [intelligence.senate.gov](https://www.intelligence.senate.gov/sites/default/files/publications/109331.pdf), [archived snapshot](http://web.archive.org/web/20250418205243/https://www.intelligence.senate.gov/sites/default/files/publications/109331.pdf).
[^328]: Bob Woodward, Plan of Attack (New York: Simon & Schuster, 2004), Woodward, Plan of Attack, 2004, p. 249. [primary URL no longer resolves as of 2026-05-28; canonical archive] [archived snapshot](http://web.archive.org/web/20250121114935/https://lccn.loc.gov/2004351204).; George Tenet (with Bill Harlow), At the Center of the Storm: My Years at the CIA (New York: HarperCollins, 2007), Tenet, At the Center of the Storm, 2007, chapters on Iraq WMD and the 'slam dunk' framing. [harpercollins.com](https://www.harpercollins.com/products/at-the-center-of-the-storm-george-tenet), [archived snapshot](https://web.archive.org/web/20260526121544/https://www.harpercollins.com/products/at-the-center-of-the-storm-george-tenet).
[^329]: Martin Chulov and Helen Pidd, 'Defector admits to WMD lies that triggered Iraq war', The Guardian, 15 February 2011 (interview with Rafid Ahmed Alwan al-Janabi, codename 'Curveball'), Guardian online article, 15 February 2011. [theguardian.com](https://www.theguardian.com/world/2011/feb/15/defector-admits-wmd-lies-iraq-war), [archived snapshot](https://web.archive.org/web/*/theguardian.com/world/2011/feb/15/defector-admits-wmd-lies-iraq-war).
[^330]: Commission on the Intelligence Capabilities of the United States Regarding Weapons of Mass Destruction (Robb-Silberman Commission), Report to the President of the United States, 31 March 2005, Robb-Silberman Report, Letter of Transmittal and Chapter 1 (Iraq). [govinfo.library.unt.edu](https://govinfo.library.unt.edu/wmd/report/index.html), [archived snapshot](http://web.archive.org/web/20250421105951/https://govinfo.library.unt.edu/wmd/report/index.html).
[^331]: Elisabeth Bumiller, 'Traces of Terror: The Strategy; Bush Aides Set Strategy to Sell Policy on Iraq', The New York Times, 7 September 2002 (NYT category /us/, not /world/), NYT, 7 September 2002, Bumiller article on Bush aides' Iraq strategy. [primary URL no longer resolves as of 2026-05-28; canonical archive] [archived snapshot](http://web.archive.org/web/20251107175635/https://www.nytimes.com/2002/09/07/us/traces-of-terror-the-strategy-bush-aides-set-strategy-to-sell-policy-on-iraq.html), access: paywall.
[^332]: U.S. Senate Select Committee on Intelligence, Report on the U.S. Intelligence Community's Prewar Intelligence Assessments on Iraq (S. Rep. 108-301), 108th Cong., 2d Sess., 7 July 2004, S. Rep. 108-301, Conclusion 1 (Key Judgments) and accompanying findings. [govinfo.gov](https://www.govinfo.gov/content/pkg/CRPT-108srpt301/pdf/CRPT-108srpt301.pdf).
[^333]: Lawrence Wilkerson, public statements on the preparation of Colin Powell's 5 February 2003 UN Security Council presentation, including interview broadcast on PBS Frontline, "The Dark Side" (20 June 2006), and interview broadcast on CNN, "Late Edition with Wolf Blitzer" (19 August 2007), PBS Frontline broadcast 20 Jun 2006; CNN Late Edition broadcast 19 Aug 2007. [pbs.org](https://www.pbs.org/wgbh/pages/frontline/darkside/), [archived snapshot](https://web.archive.org/web/20260402162428/http://www.pbs.org/wgbh/pages/frontline/darkside/).
[^334]: United Nations Security Council Resolution 1441 (2002), S/RES/1441 (2002), adopted 8 November 2002, S/RES/1441 (2002), operative paragraphs 1-14. [digitallibrary.un.org](https://digitallibrary.un.org/record/479512), [archived snapshot](http://web.archive.org/web/20250619012755/https://digitallibrary.un.org/record/479512).
[^335]: Authorization for Use of Military Force Against Iraq Resolution of 2002, Pub. L. No. 107-243, 116 Stat. 1498, enacted 16 October 2002, Pub. L. No. 107-243, §3(a); 116 Stat. 1498. [govinfo.gov](https://www.govinfo.gov/content/pkg/PLAW-107publ243/pdf/PLAW-107publ243.pdf), [archived snapshot](https://web.archive.org/web/20260526130251/https://www.govinfo.gov/content/pkg/PLAW-107publ243/pdf/PLAW-107publ243.pdf).
[^336]: President George W. Bush, Remarks on Presenting the Presidential Medal of Freedom to George J. Tenet, L. Paul Bremer III, and Tommy R. Franks, 14 December 2004 (White House transcript and proclamation), White House transcript and accompanying announcement, 14 Dec 2004. [georgewbush-whitehouse.archives.gov](https://georgewbush-whitehouse.archives.gov/news/releases/2004/12/20041214-3.html), [archived snapshot](https://web.archive.org/web/20260526055839/https://georgewbush-whitehouse.archives.gov/news/releases/2004/12/20041214-3.html).
[^337]: U.S. Senate Select Committee on Intelligence, "Postwar Findings About Iraq's WMD Programs and Links to Terrorism and How They Compare with Prewar Assessments" (S. Rep. 109-331, 8 September 2006), and "Report on Whether Public Statements Regarding Iraq by U.S. Government Officials Were Substantiated by Intelligence Information" (S. Rep. 110-345, 25 May 2007 / released June 2008), S. Rep. 109-331 (8 Sep 2006); S. Rep. 110-345 (25 May 2007 / released June 2008). [intelligence.senate.gov](https://www.intelligence.senate.gov/sites/default/files/publications/109331.pdf), [archived snapshot](http://web.archive.org/web/20250418205243/https://www.intelligence.senate.gov/sites/default/files/publications/109331.pdf).
[^338]: U.S. Department of Defense, Office of the Inspector General, Report No. 07-INTEL-04, 'Review of the Pre-Iraqi War Activities of the Office of the Under Secretary of Defense for Policy', 9 February 2007, DoD IG Report 07-INTEL-04, Executive Summary; verbatim 'inappropriate' characterisation per report findings. [globalsecurity.org](https://www.globalsecurity.org/intell/library/reports/2007/dod-ig_07-intel-04_070209.htm), [archived snapshot](http://web.archive.org/web/20260526075519/https://www.globalsecurity.org/intell/library/reports/2007/dod-ig_07-intel-04_070209.htm).
[^339]: Tyler Drumheller (with Elaine Monaghan), On the Brink: An Insider's Account of How the White House Compromised American Intelligence (New York: Carroll & Graf, 2006); CBS News, 60 Minutes, interview with Tyler Drumheller, broadcast 23 April 2006, Drumheller, On the Brink, 2006; CBS 60 Minutes broadcast 23 April 2006. [cbsnews.com](https://www.cbsnews.com/news/a-spy-speaks-out-21-04-2006/), [archived snapshot](http://web.archive.org/web/20251125135255/https://www.cbsnews.com/news/a-spy-speaks-out-21-04-2006/).; George Tenet (with Bill Harlow), At the Center of the Storm: My Years at the CIA (New York: HarperCollins, 2007), Tenet, At the Center of the Storm, 2007, chapters on Iraq WMD and the 'slam dunk' framing. [harpercollins.com](https://www.harpercollins.com/products/at-the-center-of-the-storm-george-tenet), [archived snapshot](https://web.archive.org/web/20260526121544/https://www.harpercollins.com/products/at-the-center-of-the-storm-george-tenet).
[^340]: The Report of the Iraq Inquiry — Report of a Committee of Privy Counsellors (Chair: Sir John Chilcot), HC 264, 12 volumes (London: The Stationery Office, 6 July 2016), with declassified annexes including the Blair–Bush correspondence of 28 July 2002, Executive Summary ¶¶ 499 (September Dossier), 510 (JIC certainty caveats), 875-876 (legal basis); Blair-Bush correspondence annexed to Section 3.8 (volume 4). [webarchive.nationalarchives.gov.uk](https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/), [archived snapshot](https://web.archive.org/web/20260526113028/https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/).
[^341]: National Intelligence Council, National Intelligence Estimate 2002-16HC, 'Iraq's Continuing Programs for Weapons of Mass Destruction', October 2002 (declassified key judgments and excerpts released 18 July 2003), NIE 2002-16HC, Key Judgments (declassified excerpts, pp. 5-9); INR dissent set off from the Key Judgments in text boxes and Annex A. [irp.fas.org](https://irp.fas.org/cia/product/iraq-wmd-nie.pdf), [archived snapshot](http://web.archive.org/web/20250812015736/https://irp.fas.org/cia/product/iraq-wmd-nie.pdf).
[^342]: Hans Blix, Executive Chairman, UN Monitoring, Verification and Inspection Commission (UNMOVIC), briefings to the United Nations Security Council on the status of Iraq's WMD inspections, 27 January 2003; 14 February 2003; 7 March 2003, UNMOVIC briefings of 27 Jan 2003; 14 Feb 2003; 7 Mar 2003 (published transcripts). [un.org](https://www.un.org/depts/unmovic/SC7asdelivered.htm), [archived snapshot](https://web.archive.org/web/20260526112219/https://www.un.org/depts/unmovic/SC7asdelivered.htm).
[^343]: Knight-Ridder Washington Bureau, prewar Iraq intelligence coverage by Warren Strobel and Jonathan Landay, 2002-2003, Knight-Ridder Washington Bureau coverage, 2002-2003 (archive accessed via McClatchy DC). [pbs.org](https://www.pbs.org/moyers/journal/test_blogs/knight_ridder/), [archived snapshot](https://web.archive.org/web/20221115222550/http://www.pbs.org/moyers/journal/test_blogs/knight_ridder/).
[^344]: Iraq Body Count project, public database and methodology, iraqbodycount.org; Gilbert Burnham, Riyadh Lafta, Shannon Doocy, and Les Roberts, 'Mortality after the 2003 invasion of Iraq: a cross-sectional cluster sample survey', The Lancet, vol. 368, no. 9545, 11 October 2006; Iraq Family Health Survey Study Group, 'Violence-Related Mortality in Iraq from 2002 to 2006', New England Journal of Medicine, vol. 358, 31 January 2008, Iraq Body Count methods page; Roberts et al., Lancet, 11 Oct 2006 (pp. 1421-1428); IFHS, NEJM, 31 Jan 2008 (pp. 484-493). [iraqbodycount.org](https://www.iraqbodycount.org/about/methods/), [archived snapshot](http://web.archive.org/web/20260526082746/https://www.iraqbodycount.org/about/methods/).
[^345]: Iraq Body Count project, public database and methodology, iraqbodycount.org; Gilbert Burnham, Riyadh Lafta, Shannon Doocy, and Les Roberts, 'Mortality after the 2003 invasion of Iraq: a cross-sectional cluster sample survey', The Lancet, vol. 368, no. 9545, 11 October 2006; Iraq Family Health Survey Study Group, 'Violence-Related Mortality in Iraq from 2002 to 2006', New England Journal of Medicine, vol. 358, 31 January 2008, Iraq Body Count methods page; Roberts et al., Lancet, 11 Oct 2006 (pp. 1421-1428); IFHS, NEJM, 31 Jan 2008 (pp. 484-493). [iraqbodycount.org](https://www.iraqbodycount.org/about/methods/), [archived snapshot](http://web.archive.org/web/20260526082746/https://www.iraqbodycount.org/about/methods/).
[^346]: Iraq Body Count project, public database and methodology, iraqbodycount.org; Gilbert Burnham, Riyadh Lafta, Shannon Doocy, and Les Roberts, 'Mortality after the 2003 invasion of Iraq: a cross-sectional cluster sample survey', The Lancet, vol. 368, no. 9545, 11 October 2006; Iraq Family Health Survey Study Group, 'Violence-Related Mortality in Iraq from 2002 to 2006', New England Journal of Medicine, vol. 358, 31 January 2008, Iraq Body Count methods page; Roberts et al., Lancet, 11 Oct 2006 (pp. 1421-1428); IFHS, NEJM, 31 Jan 2008 (pp. 484-493). [iraqbodycount.org](https://www.iraqbodycount.org/about/methods/), [archived snapshot](http://web.archive.org/web/20260526082746/https://www.iraqbodycount.org/about/methods/).
[^347]: U.S. Department of Defense, Defense Casualty Analysis System (DCAS), official casualty statistics for Operation Iraqi Freedom (2003-2010) and Operation New Dawn (2010-2011), DCAS conflict-casualties dashboard for OIF and OND. [primary URL no longer resolves as of 2026-05-28; canonical archive] [archived snapshot](http://web.archive.org/web/20260525201604/https://dcas.dmdc.osd.mil/dcas/app/conflictCasualties).
[^348]: United Kingdom Ministry of Defence, Operations in Iraq: British Fatalities (Operation Telic, 2003-2011), UK MoD Iraq fatalities record (Operation Telic). [gov.uk](https://www.gov.uk/government/fields-of-operation/iraq), [archived snapshot](http://web.archive.org/web/20260321170424/https://www.gov.uk/government/fields-of-operation/iraq).
[^349]: United Nations High Commissioner for Refugees (UNHCR), Iraq Situation Response: data and updates on Iraqi refugees and internally displaced persons, 2006-2007, UNHCR Iraq Situation Response, 2007 reporting cycle. [primary URL no longer resolves as of 2026-05-28; canonical archive] [archived snapshot](http://web.archive.org/web/20230329234439/https://www.unhcr.org/iraq.html).
[^350]: U.S. Army courts-martial records for Abu Ghraib detainee abuse prosecutions (372nd Military Police Company and 325th Military Intelligence Battalion), 2004-2006; Brigadier General Janis Karpinski reduction-in-grade order, May 2005, Per-defendant court-martial dispositions; Karpinski reduction-in-grade order, May 2005. [washingtonpost.com](https://www.washingtonpost.com/dev-docs/politics/2005/05/06/general-demoted-but-cleared-in-abuse-probe/05b60e16-e26e-40c8-81ad-d5d0c03632fb/) (verification checkpoint-dependent), [archived snapshot](http://web.archive.org/web/20251206175840/https://www.washingtonpost.com/dev-docs/politics/2005/05/06/general-demoted-but-cleared-in-abuse-probe/05b60e16-e26e-40c8-81ad-d5d0c03632fb/).
[^351]: U.S. Senate Armed Services Committee, 'Inquiry into the Treatment of Detainees in U.S. Custody', report of the Committee on Armed Services (110th Cong., 2d Sess.), 20 November 2008, SASC Report on Detainee Treatment, 20 November 2008 (publicly released April 2009); Executive Summary. [armed-services.senate.gov](https://www.armed-services.senate.gov/imo/media/doc/Detainee-Report-Final_April-22-2009.pdf), [archived snapshot](https://web.archive.org/web/20260526125208/https://www.armed-services.senate.gov/imo/media/doc/Detainee-Report-Final_April-22-2009.pdf).
[^352]: Major General Antonio M. Taguba, "Article 15-6 Investigation of the 800th Military Police Brigade" (Taguba Report), March 2004; Independent Panel to Review DoD Detention Operations (James R. Schlesinger, Chair), Final Report (Schlesinger Report), August 2004; Major General George R. Fay and Lieutenant General Anthony R. Jones, "AR 15-6 Investigation of the Abu Ghraib Detention Facility and 205th Military Intelligence Brigade" (Fay-Jones Report), August 2004, Taguba Report (Mar 2004); Schlesinger Report (Aug 2004); Fay-Jones Report (Aug 2004). [nsarchive2.gwu.edu](https://nsarchive2.gwu.edu/NSAEBB/NSAEBB140/index.htm), [archived snapshot](http://web.archive.org/web/20250602112439/https://nsarchive2.gwu.edu/NSAEBB/NSAEBB140/index.htm).; U.S. Senate Armed Services Committee, 'Inquiry into the Treatment of Detainees in U.S. Custody', report of the Committee on Armed Services (110th Cong., 2d Sess.), 20 November 2008, SASC Report on Detainee Treatment, 20 November 2008 (publicly released April 2009); Executive Summary. [armed-services.senate.gov](https://www.armed-services.senate.gov/imo/media/doc/Detainee-Report-Final_April-22-2009.pdf), [archived snapshot](https://web.archive.org/web/20260526125208/https://www.armed-services.senate.gov/imo/media/doc/Detainee-Report-Final_April-22-2009.pdf).
[^353]: Joint Investigation Team (JIT) led by the Dutch Public Prosecution Service (Openbaar Ministerie), with Australia, Belgium, Malaysia, and Ukraine, findings on the 17 July 2014 destruction of Malaysia Airlines flight MH17: presentations and updates 28 September 2016, 24 May 2018, 19 June 2019, and 8 February 2023, JIT presentations 28 Sep 2016, 24 May 2018, 19 Jun 2019, and 8 Feb 2023; 'strong indications' verbatim from 8 Feb 2023 press conference. [prosecutionservice.nl](https://www.prosecutionservice.nl/topics/mh17-plane-crash), [archived snapshot](https://web.archive.org/web/20260526051631/https://www.prosecutionservice.nl/topics/m/mh17-plane-crash).
[^354]: Rechtbank Den Haag (The Hague District Court), sitting at Schiphol, judgment in the MH17 criminal case, 17 November 2022 (defendants: Igor Vsevolodovich Girkin, Sergey Nikolayevich Dubinskiy, Leonid Volodymyrovych Kharchenko, and Oleg Yuldashevich Pulatov), Rechtbank Den Haag, judgment of 17 November 2022, MH17 criminal case. [courtmh17.com](https://www.courtmh17.com/en/insights/news/2022/summary-of-the-day-in-court-17-november-2022-judgment/), [archived snapshot](http://web.archive.org/web/20240909135739/https://www.courtmh17.com/en/insights/news/2022/summary-of-the-day-in-court-17-november-2022-judgment/).
[^355]: President Vladimir Putin, public remarks denying Russian state involvement in the destruction of Malaysia Airlines flight MH17, 21 July 2014 (Kremlin transcript), Kremlin press service transcript, 21 July 2014. [en.kremlin.ru](http://en.kremlin.ru/events/president/news/46262), [archived snapshot](http://web.archive.org/web/20260211210405/http://en.kremlin.ru/events/president/news/46262).
[^356]: Joint Investigation Team (JIT) led by the Dutch Public Prosecution Service (Openbaar Ministerie), with Australia, Belgium, Malaysia, and Ukraine, findings on the 17 July 2014 destruction of Malaysia Airlines flight MH17: presentations and updates 28 September 2016, 24 May 2018, 19 June 2019, and 8 February 2023, JIT presentations 28 Sep 2016, 24 May 2018, 19 Jun 2019, and 8 Feb 2023; 'strong indications' verbatim from 8 Feb 2023 press conference. [prosecutionservice.nl](https://www.prosecutionservice.nl/topics/mh17-plane-crash), [archived snapshot](https://web.archive.org/web/20260526051631/https://www.prosecutionservice.nl/topics/m/mh17-plane-crash).
[^357]: U.S. Department of Justice, indictment in United States v. Slough, Slatten, Liberty, and Heard (D.D.C., 2008-2014); contemporaneous U.S. Department of State incident reports on the 16 September 2007 Nisour Square shooting in Baghdad, DOJ press release 08-NSD-1068 (8 Dec 2008); State Department contemporaneous incident reports (2007). [justice.gov](https://www.justice.gov/dev-docs/opa/pr/2008/December/08-nsd-1068.html), [archived snapshot](http://web.archive.org/web/20260405210101/https://www.justice.gov/dev-docs/opa/pr/2008/December/08-nsd-1068.html).
[^358]: United States v. Slatten, No. 14-cr-00107 (D.D.C. 2014); United States v. Slough, Liberty, and Heard, No. 14-cr-00107 (D.D.C. 2014), judgments and sentences in the Nisour Square prosecution, U.S. District Court for the District of Columbia, judgments and sentences in U.S. v. Slatten and U.S. v. Slough et al. (Apr 2015). [justice.gov](https://www.justice.gov/archives/opa/pr/four-former-blackwater-employees-found-guilty-charges-fatal-nisur-square-shooting-iraq), [archived snapshot](https://web.archive.org/web/20260526092245/https://www.justice.gov/archives/opa/pr/four-former-blackwater-employees-found-guilty-charges-fatal-nisur-square-shooting-iraq).
[^359]: The White House, 'Statement from the Press Secretary Regarding Executive Grants of Clemency', 22 December 2020 (full pardons to Nicholas Slatten, Paul Slough, Evan Liberty, and Dustin Heard), White House press secretary statement, 22 December 2020 (Trump clemency grants). [trumpwhitehouse.archives.gov](https://trumpwhitehouse.archives.gov/briefings-statements/statement-press-secretary-regarding-executive-grants-clemency-122220/), [archived snapshot](http://web.archive.org/web/20260429144434/https://trumpwhitehouse.archives.gov/briefings-statements/statement-press-secretary-regarding-executive-grants-clemency-122220/).
[^360]: Coalition Provisional Authority Order No. 17 (Revised), 'Status of the Coalition Provisional Authority, MNF — Iraq, Certain Missions and Personnel in Iraq', issued by L. Paul Bremer III, 27 June 2004, CPA Order 17 (Revised), 27 June 2004, §§ 4-7 (contractor immunity provisions). [govinfo.library.unt.edu](https://govinfo.library.unt.edu/cpa-iraq/regulations/20040627_CPAORD_17_Status_of_Coalition__Rev__with_Annex_A.pdf), [archived snapshot](http://web.archive.org/web/20260314033335/https://govinfo.library.unt.edu/cpa-iraq/regulations/20040627_CPAORD_17_Status_of_Coalition__Rev__with_Annex_A.pdf).
[^361]: Erik Prince, public statements and writings on his role at Blackwater (subsequent companies Xe Services, Academi, Frontier Services Group; 2009 book 'Civilian Warriors' and interviews 2007-2026); public-record search of U.S. Department of Justice press releases and U.S. federal-court PACER docket for criminal charges naming Erik Prince on Nisour Square matters as of mid-2026, PACER public-record search and DOJ press-release search as of 26 May 2026; Erik Prince's published 'Civilian Warriors' (Penguin / Portfolio, 2014). [pacer.uscourts.gov](https://pacer.uscourts.gov/), [archived snapshot](http://web.archive.org/web/20260525074532/https://pacer.uscourts.gov/).
[^362]: Russian Federation Presidential Decree No. 330 of 30 May 2022; Russian Federation Presidential Decree No. 585 of 11 July 2022 (simplified citizenship procedures applicable to children from Ukrainian territories), Decree No. 330 of 30 May 2022; Decree No. 585 of 11 July 2022 (Russian official publication). [publication.pravo.gov.ru](http://publication.pravo.gov.ru/Document/View/0001202205300018), [archived snapshot](http://web.archive.org/web/20260118101931/http://publication.pravo.gov.ru/Document/View/0001202205300018).
[^363]: Public statements by President Vladimir Putin and Russian Federation Commissioner for Children's Rights Maria Lvova-Belova defending Russian transfers of Ukrainian children from occupied territories as humanitarian, broadcast on Russian state television and via TASS, 2022-2023, TASS and Russian state television broadcasts, 2022-2023 (multiple). [tass.com](https://tass.com/society/1509049), [archived snapshot](http://web.archive.org/web/20250809130006/https://tass.com/society/1509049).
[^364]: International Criminal Court, Pre-Trial Chamber II, Decision on the Prosecutor's Application for Warrants of Arrest under Article 58 of the Rome Statute, ICC-01/22, 17 March 2023 (arrest warrants for Vladimir Vladimirovich Putin and Maria Alekseyevna Lvova-Belova), ICC Pre-Trial Chamber II arrest warrant docket, 17 March 2023, Situation in Ukraine (ICC-01/22). [icc-cpi.int](https://www.icc-cpi.int/news/situation-ukraine-icc-judges-issue-arrest-warrants-against-vladimir-vladimirovich-putin-and), [archived snapshot](http://web.archive.org/web/20260522234900/https://www.icc-cpi.int/news/situation-ukraine-icc-judges-issue-arrest-warrants-against-vladimir-vladimirovich-putin-and).
[^365]: Elisabeth Bumiller, 'Traces of Terror: The Strategy; Bush Aides Set Strategy to Sell Policy on Iraq', The New York Times, 7 September 2002 (NYT category /us/, not /world/), NYT, 7 September 2002, Bumiller article on Bush aides' Iraq strategy. [primary URL no longer resolves as of 2026-05-28; canonical archive] [archived snapshot](http://web.archive.org/web/20251107175635/https://www.nytimes.com/2002/09/07/us/traces-of-terror-the-strategy-bush-aides-set-strategy-to-sell-policy-on-iraq.html), access: paywall.
[^366]: Memorandum from Daniel Levin, Acting Assistant Attorney General, OLC, to James B. Comey, Deputy Attorney General, "Legal Standards Applicable Under 18 U.S.C. §§ 2340-2340A" (Dec. 30, 2004), Levin memorandum of 30 Dec 2004, page 2 (introduction) and footnote 8. [justice.gov](https://www.justice.gov/file/18791/download), [archived snapshot](https://web.archive.org/web/20260526120231/https://www.justice.gov/file/145681/dl?inline=).
[^367]: Knight-Ridder Washington Bureau, prewar Iraq intelligence coverage by Warren Strobel and Jonathan Landay, 2002-2003, Knight-Ridder Washington Bureau coverage, 2002-2003 (archive accessed via McClatchy DC). [pbs.org](https://www.pbs.org/moyers/journal/test_blogs/knight_ridder/), [archived snapshot](https://web.archive.org/web/20221115222550/http://www.pbs.org/moyers/journal/test_blogs/knight_ridder/).
[^368]: Lawrence Wilkerson, public statements on the preparation of Colin Powell's 5 February 2003 UN Security Council presentation, including interview broadcast on PBS Frontline, "The Dark Side" (20 June 2006), and interview broadcast on CNN, "Late Edition with Wolf Blitzer" (19 August 2007), PBS Frontline broadcast 20 Jun 2006; CNN Late Edition broadcast 19 Aug 2007. [pbs.org](https://www.pbs.org/wgbh/pages/frontline/darkside/), [archived snapshot](https://web.archive.org/web/20260402162428/http://www.pbs.org/wgbh/pages/frontline/darkside/).
[^369]: Tyler Drumheller (with Elaine Monaghan), On the Brink: An Insider's Account of How the White House Compromised American Intelligence (New York: Carroll & Graf, 2006); CBS News, 60 Minutes, interview with Tyler Drumheller, broadcast 23 April 2006, Drumheller, On the Brink, 2006; CBS 60 Minutes broadcast 23 April 2006. [cbsnews.com](https://www.cbsnews.com/news/a-spy-speaks-out-21-04-2006/), [archived snapshot](http://web.archive.org/web/20251125135255/https://www.cbsnews.com/news/a-spy-speaks-out-21-04-2006/).
[^370]: Matthew Rycroft (Downing Street), 'Iraq: Prime Minister's Meeting, 23 July', secret minute of meeting between Prime Minister Tony Blair and senior officials, 23 July 2002 (the 'Downing Street Memo'); published by Michael Smith, The Sunday Times, 1 May 2005, Rycroft minute, 23 July 2002, paragraph attributing the observation to 'C' (Sir Richard Dearlove, MI6 Chief); published in The Sunday Times 1 May 2005. [nsarchive2.gwu.edu](https://nsarchive2.gwu.edu/NSAEBB/NSAEBB328/II-Doc14.pdf), [archived snapshot](https://web.archive.org/web/20260527002539/https://nsarchive2.gwu.edu/NSAEBB/NSAEBB328/II-Doc14.pdf).
[^371]: Martin Chulov and Helen Pidd, 'Defector admits to WMD lies that triggered Iraq war', The Guardian, 15 February 2011 (interview with Rafid Ahmed Alwan al-Janabi, codename 'Curveball'), Guardian online article, 15 February 2011. [theguardian.com](https://www.theguardian.com/world/2011/feb/15/defector-admits-wmd-lies-iraq-war), [archived snapshot](https://web.archive.org/web/*/theguardian.com/world/2011/feb/15/defector-admits-wmd-lies-iraq-war).
[^372]: U.S. Federal Bureau of Investigation, Reports of Interviews with Saddam Hussein, Special Agent George L. Piro, January-June 2004 (declassified July 2009; obtained by the National Security Archive under FOIA), FBI Piro interview reports, January-June 2004 (declassified July 2009, hosted at National Security Archive Electronic Briefing Book No. 279). [nsarchive2.gwu.edu](https://nsarchive2.gwu.edu/NSAEBB/NSAEBB279/index.htm), [archived snapshot](http://web.archive.org/web/20260315164932/https://nsarchive2.gwu.edu/NSAEBB/NSAEBB279/index.htm).
[^373]: The Report of the Iraq Inquiry — Report of a Committee of Privy Counsellors (Chair: Sir John Chilcot), HC 264, 12 volumes (London: The Stationery Office, 6 July 2016), with declassified annexes including the Blair–Bush correspondence of 28 July 2002, Executive Summary ¶¶ 499 (September Dossier), 510 (JIC certainty caveats), 875-876 (legal basis); Blair-Bush correspondence annexed to Section 3.8 (volume 4). [webarchive.nationalarchives.gov.uk](https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/), [archived snapshot](https://web.archive.org/web/20260526113028/https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/).
[^374]: The Report of the Iraq Inquiry — Report of a Committee of Privy Counsellors (Chair: Sir John Chilcot), HC 264, 12 volumes (London: The Stationery Office, 6 July 2016), with declassified annexes including the Blair–Bush correspondence of 28 July 2002, Executive Summary ¶¶ 499 (September Dossier), 510 (JIC certainty caveats), 875-876 (legal basis); Blair-Bush correspondence annexed to Section 3.8 (volume 4). [webarchive.nationalarchives.gov.uk](https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/), [archived snapshot](https://web.archive.org/web/20260526113028/https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/).
[^375]: Iraq Survey Group, Comprehensive Report of the Special Advisor to the Director of Central Intelligence on Iraq's Weapons of Mass Destruction (Duelfer Report), 30 September 2004, with addenda March 2005, Duelfer Report, Volume I (Regime Strategic Intent) and Volume III (Iraq's WMD). [govinfo.gov](https://www.govinfo.gov/content/pkg/GPO-DUELFERREPORT/pdf/GPO-DUELFERREPORT-1.pdf).
[^376]: U.S. Senate Select Committee on Intelligence, Report on the U.S. Intelligence Community's Prewar Intelligence Assessments on Iraq (S. Rep. 108-301), 108th Cong., 2d Sess., 7 July 2004, S. Rep. 108-301, Conclusion 1 (Key Judgments) and accompanying findings. [govinfo.gov](https://www.govinfo.gov/content/pkg/CRPT-108srpt301/pdf/CRPT-108srpt301.pdf).
[^377]: Commission on the Intelligence Capabilities of the United States Regarding Weapons of Mass Destruction (Robb-Silberman Commission), Report to the President of the United States, 31 March 2005, Robb-Silberman Report, Letter of Transmittal and Chapter 1 (Iraq). [govinfo.library.unt.edu](https://govinfo.library.unt.edu/wmd/report/index.html), [archived snapshot](http://web.archive.org/web/20250421105951/https://govinfo.library.unt.edu/wmd/report/index.html).
[^378]: U.S. Senate Select Committee on Intelligence, "Postwar Findings About Iraq's WMD Programs and Links to Terrorism and How They Compare with Prewar Assessments" (S. Rep. 109-331, 8 September 2006), and "Report on Whether Public Statements Regarding Iraq by U.S. Government Officials Were Substantiated by Intelligence Information" (S. Rep. 110-345, 25 May 2007 / released June 2008), S. Rep. 109-331 (8 Sep 2006); S. Rep. 110-345 (25 May 2007 / released June 2008). [intelligence.senate.gov](https://www.intelligence.senate.gov/sites/default/files/publications/109331.pdf), [archived snapshot](http://web.archive.org/web/20250418205243/https://www.intelligence.senate.gov/sites/default/files/publications/109331.pdf).
[^379]: U.S. Department of Defense, Office of the Inspector General, Report No. 07-INTEL-04, 'Review of the Pre-Iraqi War Activities of the Office of the Under Secretary of Defense for Policy', 9 February 2007, DoD IG Report 07-INTEL-04, Executive Summary; verbatim 'inappropriate' characterisation per report findings. [globalsecurity.org](https://www.globalsecurity.org/intell/library/reports/2007/dod-ig_07-intel-04_070209.htm), [archived snapshot](http://web.archive.org/web/20260526075519/https://www.globalsecurity.org/intell/library/reports/2007/dod-ig_07-intel-04_070209.htm).
[^380]: Martin Chulov and Helen Pidd, 'Defector admits to WMD lies that triggered Iraq war', The Guardian, 15 February 2011 (interview with Rafid Ahmed Alwan al-Janabi, codename 'Curveball'), Guardian online article, 15 February 2011. [theguardian.com](https://www.theguardian.com/world/2011/feb/15/defector-admits-wmd-lies-iraq-war), [archived snapshot](https://web.archive.org/web/*/theguardian.com/world/2011/feb/15/defector-admits-wmd-lies-iraq-war).
[^381]: The Report of the Iraq Inquiry — Report of a Committee of Privy Counsellors (Chair: Sir John Chilcot), HC 264, 12 volumes (London: The Stationery Office, 6 July 2016), with declassified annexes including the Blair–Bush correspondence of 28 July 2002, Executive Summary ¶¶ 499 (September Dossier), 510 (JIC certainty caveats), 875-876 (legal basis); Blair-Bush correspondence annexed to Section 3.8 (volume 4). [webarchive.nationalarchives.gov.uk](https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/), [archived snapshot](https://web.archive.org/web/20260526113028/https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/).
[^382]: The Report of the Iraq Inquiry — Report of a Committee of Privy Counsellors (Chair: Sir John Chilcot), HC 264, 12 volumes (London: The Stationery Office, 6 July 2016), with declassified annexes including the Blair–Bush correspondence of 28 July 2002, Executive Summary ¶¶ 499 (September Dossier), 510 (JIC certainty caveats), 875-876 (legal basis); Blair-Bush correspondence annexed to Section 3.8 (volume 4). [webarchive.nationalarchives.gov.uk](https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/), [archived snapshot](https://web.archive.org/web/20260526113028/https://webarchive.nationalarchives.gov.uk/ukgwa/20171123124621/http://www.iraqinquiry.org.uk/).
[^383]: Executive Order 13328 of February 6, 2004, "Commission on the Intelligence Capabilities of the United States Regarding Weapons of Mass Destruction," 69 Fed. Reg. 6901 (Feb. 11, 2004), 69 Fed. Reg. 6901; Sections 2 and 3 of the Order. [govinfo.gov](https://www.govinfo.gov/content/pkg/WCPD-2004-02-09/pdf/WCPD-2004-02-09-Pg203.pdf), [archived snapshot](https://web.archive.org/web/20260527024057/https://www.govinfo.gov/content/pkg/WCPD-2004-02-09/pdf/WCPD-2004-02-09-Pg203.pdf).

Copyright: © LookMinds. Our own edited text and translation, not a substitute for the original; rights in the original belong to the author and the platform. Please see the original link.