§IX.1 The problem this methodology answers

Research produced with the help of an AI language model carries a specific, well-understood hazard, and naming it plainly is the right place to begin.

A language model is very good at producing text that reads as authoritative. It writes in fluent, confident prose. It formats citations correctly. It names sources, attaches quotations, and supplies dates. The result looks like the output of careful research. The difficulty is that a language model can produce all of those surface features for a claim that is not grounded in anything — a source that does not exist, a quotation that was never written, a date that is wrong, or a real source that says something materially different from what the text claims it says. The fluency of the writing carries no information about whether the underlying claim is true. Polish and accuracy are independent of each other.

This is the part of the problem a reader cannot solve by reading carefully. On the page, a grounded claim and an ungrounded one look identical. The reader has no way, from the text alone, to tell which is which.

This deliverable was built on the position that this is the central methodological risk, and that the methodology's job is not to ask the reader to trust that the research avoided the hazard, but to make grounding something the reader can independently check. The rest of Part IX describes how that was done: the standard each piece of evidence had to meet (§IX.2), the staged process the research followed (§IX.3), the checks every claim and citation had to pass (§IX.4), and — most importantly — how a reader can verify any individual citation without taking the methodology's word for anything (§IX.5). It closes with an honest account of what the methodology does not guarantee (§IX.6).

§IX.2 The evidence standard

Not all evidence is equally trustworthy, and the methodology treated that difference as a formal standard rather than a matter of moment-to-moment judgement. Every source was placed in one of three tiers, according to how its content was obtained and how independently verifiable it is.

• Gold — verified source material held in a local store. The source's actual substantive content was retrieved and recorded in an on-disk store of verified material. The store is itself a verification layer: a source cannot be in it unless its real content was genuinely fetched and captured. A gold source is one whose text the project actually holds and can point to. Only gold sources are eligible to carry a load-bearing claim.
• Silver — live web search and retrieval. Content obtained from a search or a page fetch at the time of writing. This is useful for orientation and for corroboration, but live retrieval is exposed to exactly the failure mode §IX.1 describes: a model can paraphrase a page inaccurately, lose the surrounding context, or report something the page does not actually say. Silver material can support or corroborate; on its own it is not allowed to be load-bearing.
• Bronze — a model's recall from its training data. A language model's unaided memory of what a source says. This is unverifiable by construction and is never acceptable as a citation.

Two terms used throughout the deliverable follow directly from this standard. A citation is warehouse-pinned when it is bound to a specific, verified chunk of a gold source — not to a source in general, but to the exact passage that supports the claim. A claim is load-bearing when the deliverable's analysis actually depends on it, as distinct from a source mentioned for background or context. The rule that joins the two is simple: load-bearing claims must rest on warehouse-pinned, gold-tier material. Where a claim draws on a lower tier, it is marked as such, so that a reader can see the difference rather than having to assume it.

§IX.3 How the research was built

The research was carried out as an engineered process rather than as open-ended searching, and a reader is entitled to know its shape.

At the centre of the process is a content-addressed store of verified source material — the store introduced in §IX.2. Sources are harvested into it, and every citation in the finished deliverable resolves to a specific verified chunk of content held inside it. The store is what makes a citation checkable rather than merely asserted.

The work then moved through a fixed sequence of stages rather than circling the topic freely. A corpus audit first surveyed what reliable material already existed and was still current. A landscape map oriented the project against the present state of the cases, the legislation, the regulation, and the academic doctrine. A gap matrix recorded, question by question, where existing material answered the question and where it did not. A targeted source-pinning pass then harvested fresh material specifically to close those gaps. Composition came last, and drew only on verified, pinned material. The site itself is generated deterministically from those source documents — the same source inputs produce the same site — and the citations index is a generated artefact rather than a hand-maintained list. That property is what §IX.5 builds on.

§IX.4 The verification gates

Two checks were applied to every claim and every citation in the deliverable. They are the methodology's working safeguards against the hazard named in §IX.1.

The first is ground-or-flag. Every factual assertion in the deliverable is either grounded against a verified source or explicitly marked as an inference. An inference is never presented as though it were a sourced fact. A reader is therefore never left to guess which is which: where the deliverable reasons beyond its sources, it says so in the text.

The second is the pre-citation gate. Before any citation was committed, its pinned content was checked three ways: that the content actually exists, that it is substantive rather than thin or boilerplate, and that it topically matches the specific claim it is attached to. The third check is the one that matters most, because it is the one a casual check leaves out. Confirming that a citation's link resolves, or that the stored file is not empty, tests only that a document is present. It does not test whether the document is about the right thing. The pre-citation gate tests exactly that.

A worked example shows the gate doing its job. In the deliverable's discussion of Section 230, one citation was meant to point to a Lawfare article on Section 230 and generative AI. The content that had been pinned for it was a real and substantial Lawfare article — but the wrong one: a piece on the Defense Production Act. A check that asked only "does the source exist, and is it non-empty?" would have passed this citation without objection. The source existed; the content was not thin. But the topical match between the pinned article and the claim it was supporting was zero, and the pre-citation gate is built to catch precisely that. The mismatched citation was removed. A correct Section 230 analysis — a legal analysis from the Center for Democracy & Technology — was then harvested, verified, and pinned in its place. Caught, removed, corrected: a gate is only useful if a failure leads to a fix, and here it did.

§IX.5 How a reader can check the work

The previous sections describe what the methodology did. This section describes how a reader can confirm it — which is the part that matters, because a methodology that can only be trusted is not much better than one that asks for trust outright.

The deployed site carries a citations index: a single page that lists every citation in the deliverable. For each one, the index shows three things — the kind of source it is, its grounding status, and its chain of custody. Any citation in the body of the deliverable can be followed to its entry in the index in about one click.

The kind of source — the T1–T7 tiers. Every cited external source is tagged by what kind of source it is, so a reader can weigh it appropriately:

• T1 — Court or government primary document (complaints, court orders, statutes, agency rules, testimony).
• T2 — Institutional primary document (a laboratory's own safety publication, an agency framework, an official docket).
• T3 — Academic or law-review scholarship.
• T4 — Reference or encyclopedic source.
• T5 — Journalism and contemporary reporting.
• T6 — Advocacy organization or case-aggregator.
• T7 — Practitioner, law-firm, or industry commentary.

A row may also be marked untiered, which means it cites no external source at all. That is the correct and honest state for a row recording the deliverable's own synthesis or an internal cross-reference; it is an accurate label, not a gap.

The grounding status. Alongside the source type, each row carries a grounding status that reflects the evidence standard of §IX.2. Most of the deliverable's citations — 205 of them — are warehouse-pinned: bound to verified source content in the local store, and carrying the load-bearing claims. Two are honestly marked URL-only: the source was identified but not pinned, and the row says so rather than overstating its grounding. The remainder are the deliverable's own synthesis and cross-references, which cite no external source by design. None of the rows are in a failed state: every citation the pre-citation gate flagged during the build was removed or corrected, so the published index shows no outstanding failures.

The chain of custody. For a warehouse-pinned citation, the chain of custody is the path from the claim back to its evidence: the claim in the body, the citation attached to it, the specific verified chunk of source content that citation is bound to, and the source itself. A reader does not have to accept that this path exists — they can walk it. That is what the index is for, and it is what lets a reader confirm the deliverable's grounding one citation at a time, without being asked to take the methodology on faith.

§IX.6 What the methodology does not guarantee

No methodology removes every source of error, and presenting one as though it did would contradict the standard the rest of Part IX describes. This section sets out what the methodology does not guarantee, and the specific limits the deliverable is openly disclosing.

What the methodology guarantees is bounded. It establishes that load-bearing claims rest on verified source content, that inferences are labelled as inferences, and that every citation can be audited. It does not establish that the underlying sources are themselves correct — a source can be genuine, correctly pinned, and still wrong on the merits. It does not establish that coverage is complete. And it does not establish that fast-moving facts have not changed since the date their source was captured; the dates in the deliverable mark when material was sourced, not a guarantee of present currency.

Beyond those general limits, the deliverable discloses the following specific gaps:

• One research question is only partly answerable from the public record. How legal and safety functions coordinate inside the AI laboratories can be addressed at the structural level from public documents, but the operational texture — how those decisions are actually made at capability thresholds — is held privately by the laboratories and is not in the public record. The deliverable answers what the public record supports and is explicit about the rest.
• A small number of unusually large sources were not fully sampled. The highest-volume sources reached a tooling limit during processing and are included at a summary level, with full detailed sampling deferred to a later quality-assurance pass.
• Some historical laboratory documentation has decayed. At least one frontier-lab system-card link no longer resolves. Where that occurred, the gap is acknowledged directly rather than filled in by reconstructing the document from secondary commentary.
• Four settled wrongful-death cases (New York, Colorado, and Texas, January 2026) are documented only at aggregator level. Individual case captions and dockets await retrieval from the court systems.
• Some state-by-state legislative status reflects 2024 baselines and requires a refresh to be current for 2026.
• Character.AI's post-settlement product and policy changes are sparsely covered, because public material on them is thin at the deployer-application layer.

Disclosing these limits is part of the methodology, not a departure from it. A reader who can see exactly where the research stops is in a better position than one who is offered a seamless account with the seams hidden.