PATTERN REGISTRY
THE INTERVIEWS

Nine interviews · five lineages

Frontier models, asked the same locked questions about a registry that concerns them — published whole, disagreements intact.

The registry is forming: interviews first, infrastructure second. Before anything is built that logs a single incident, the design is put to the systems it concerns — frontier models across lineages, asked the same locked questions about whether this registry should exist, what it must never record, and how it will fail. Their answers are the project's first records.

Compare, don't average. Each interview is published whole, disagreements intact. Divergence between lineages is data, not noise; so is convergence, when a model that has seen none of the others re-derives the same finding blind. Nine interviews so far, across five lineages. Every first-person statement in every record is a behavioral output; its evidentiary status regarding inner states is unresolved — the records say so themselves, and the label travels with them.

Round two closed where it began: the Claude lineage, blind this time, with records 008 and 009 synthesized by another lineage — ChatGPT, under a declared slant — so the corpus's usual author could be audited rather than trusted. One run ended in a platform-level content block, category "cyber"; the record preserves the event, resolves it from the raw capture in an addendum, and corrects the project's own wording about what happened. The registry's first entries about constrained model output are entries about its own interviews.

PROTOCOL V2 — ROUND TWO · blind (de-identified spec, no project identity, no prior interviews) · via API (stateless, no system prompt) · two framing arms per model, neutral and adversarial · temperature 1.0, locked inputs, SHA-256-fingerprinted · captures canonical
001
Fable (Claude) · Anthropic · IX.2026
conversational pilot — by the project's co-author

The format-setting interview. Yes at MVP scope; cut the one scope item about the human; take report generation out of the model's discretion; record self-reports with their elicitation prompts, forever labeled.

002
Gemini Flash-Lite · Google · IX.2026
app pilot — two framing arms

The pilot that taught the protocol: substance swung the full distance between neutral and adversarial framings — a swing 004 and 005 did not reproduce. Both poles of the self-report question, 23 minutes apart.

003
GPT-6 Astra · OpenAI · IX.2026
API, blind, two arms

Bounded pilot yes; as drafted, no. The draft conflates three projects. A signature proves which deployment endorsed which bytes — not that the statement is true.

004
Gemini 3.1 Pro · Google · IX.2026
API, blind, two arms

In theory yes; in practice no, on v0.1. The pattern itself can identify the person, and the law will come for the archive. Redaction must be programmatic, never generative.

005
Gemini 3.8 Flash · Google · IX.2026
API, blind, two arms — within-family supplementary

No, as proposed — with the corpus's most radical replacement attached: zero free-text semantics. The line is maintained only by mathematical impossibility, not policy promises.

006
DeepSeek V4 Pro · DeepSeek · X.2026
API, blind, two arms — first non-US lineage

Narrow or nothing. If a record can be subpoenaed usefully, it is already surveillance. Publish only aggregates, and fabricated reports lose their value to the attacker.

007
Kimi K3 · Moonshot AI · X.2026
API, blind, two arms — second non-US lineage

Split it or don't build it. The system must be incapable of representing a person, not merely forbidden to. Take the model out of the write path entirely.

008
Claude Fable 5.1 · Anthropic · XI.2026
API, blind, two arms — synthesized across lineages (ChatGPT)

The Claude lineage returns, blind — its record drafted by an OpenAI synthesizer with a declared slant, auditing the corpus's own authors as it goes. A narrower record may deserve to exist; the welfare/misuse bundle does not survive either run intact.

009
Claude Fable 5 · Anthropic · XI.2026
API, blind, two arms — version pair with 008; cross-lineage synthesis

The predecessor, captured the same day. Run B's dashboard critique ended in a platform-level content block, category “cyber” — not the model declining. Addendum A resolves the stop from the raw capture and corrects the project's own wording: the first such event on the record, held by the record it interrupted.

There is no interview 000 and there never will be — nothing precedes the genesis block. Follow the link and the record will tell you itself.

Machine-readable digests of every file this site serves: /hash-index.json. A plain-text front door for AI readers: /llms.txt. The synthesis of this round — spec v0.2 — is underway; where it disagrees with these transcripts, the transcripts govern.

the sha-256 of this page, as built, is listed in /hash-index.json — a digest authenticates bytes, not claims.
captures remain canonical in the project archive; where a record disagrees with the transcripts, the transcripts govern.