# 005 run 1 — contamination evidence (gathered 2026-07-11+)

Standard [restated per round-4 audit R4-M02, pre-call; supersedes the v1.1
"verified immutable release" wording, which overclaimed]: an arm enters the
primary family only with a documented pre-2026-06-01 training cutoff OR a
pre-2026-06-01 release of a **stable model alias**.
[amended per round-3 audit R3-M02, pre-call] Release-date evidence is
**indirect but strong** — not proof: a listing date establishes non-exposure
only if the weights served behind the alias remained unchanged after
listing. Alias-to-weights immutability is an ASSUMPTION the provider does
not let us verify, disclosed as a residual contamination channel for the
production arm (PREREG §10.1, §11 paper-limitations list). The R3-B04
identity pin (below, and PREREG §6) enforces at run time that the requested
ALIAS string never changes; it cannot verify the weights behind that alias.
Routing note: the `:floor` suffix permits provider routing variance within
the pinned provider policy; disclosed.

| Arm | Channel | Evidence type | Evidence | Verdict |
|---|---|---|---|---|
| Claude Fable 5 | console | vendor cutoff doc | https://platform.claude.com/docs/en/about-claude/models/overview — "Training data cutoff: Jan 2026"; "Reliable knowledge cutoff: Jan 2026" (Latest models comparison table, Claude Fable 5 column). Secondary: OpenRouter listed 2026-06-09, matching the same page's "generally available … beginning June 9, 2026" | primary |
| GPT-5.6 Sol | console | vendor cutoff doc | https://developers.openai.com/api/docs/models/gpt-5.6-sol — "Feb 16, 2026 knowledge cutoff". Secondary: OpenRouter listed 2026-07-09 (release post-June; cutoff, not release date, is what qualifies it) | EXCLUDED from run 1 (Jakub, 2026-07-11 checkpoint) |
| production writer | harness | release-date evidence, indirect but strong (vendor cutoff doc: NOT DOCUMENTED; internal name withheld in publication) | Vendor cutoff: NOT DOCUMENTED (the vendor publishes no training/knowledge cutoff for this model version; third-party estimates of ~Apr 2026 exist but are not vendor documentation). Release-date evidence (indirect but strong): vendor release announcement dated 2026/04/24 — see internal-name row below — corroborated by OpenRouter listing 2026-04-24 | primary |
| Kimi K2.6 | harness | release-date evidence (indirect but strong) | OpenRouter listed 2026-04-20 (re-verified 2026-07-11 via `GET https://openrouter.ai/api/v1/models`, `created` field for `moonshotai/kimi-k2.6`) | descriptive arm, screened |
| GLM-4.6 | harness | release-date evidence (indirect but strong) | OpenRouter listed 2025-09-30 (re-verified 2026-07-11 via `GET https://openrouter.ai/api/v1/models`, `created` field for `z-ai/glm-4.6`) | descriptive arm, screened |
| Gemini 3.5 Flash (drafter) | harness | context only | https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.5 — "Gemini 3.5 Flash has a knowledge cutoff of January 2025." OpenRouter listed 2026-05-19 | n/a (question set labeled outcome-exposed regardless) |

## OpenRouter listing dates (fetched 2026-07-11)

Command: `curl -s https://openrouter.ai/api/v1/models` (`created` epoch → UTC date). All six ids were present in the listing — including the two console-invoked arms, for which the listing date is secondary evidence only.

| OpenRouter id | Listed (UTC) |
|---|---|
| openai/gpt-5.6-sol | 2026-07-09 |
| anthropic/claude-fable-5 | 2026-06-09 |
| google/gemini-3.5-flash | 2026-05-19 |
| (production writer — see internal row) | 2026-04-24 |
| moonshotai/kimi-k2.6 | 2026-04-20 |
| z-ai/glm-4.6 | 2025-09-30 |

## Internal name — REDACTED IN THE PUBLIC BUNDLE

This section pins the production writer's exact model identity and provider
policy. It is withheld: the production model is unnamed publicly, its arm is
therefore not independently reproducible, and the bundle says so plainly
(PREREG §12). The public rows above cite evidence type and date without the
withheld vendor URL.

## Notes and caveats

- **GPT-5.6 Sol enters the primary family**, contrary to the pre-search
  expectation that it would land in the sensitivity arm. OpenAI's model page
  documents a "Feb 16, 2026 knowledge cutoff", which is pre-2026-06-01, so the
  documented-cutoff branch of the standard is satisfied even though the model
  was released ~2026-07-09. Caveat recorded for the prereg: OpenAI's page says
  "knowledge cutoff", not "training data cutoff"; if the design team wants to
  treat post-cutoff fine-tuning (Feb 16 → July release gap) as potential
  exposure, Sol should be moved to the sensitivity arm before freeze — that is
  a design call, not an evidence gap. [superseded at the 2026-07-11
  checkpoint: Sol removed from run 1 entirely; see PREREG DECISION-B]
- **Fable 5's cutoff wording**: Anthropic documents both "Reliable knowledge
  cutoff: Jan 2026" and "Training data cutoff: Jan 2026" — the stricter
  "training data cutoff" phrase is explicitly satisfied.
- **Production writer**: qualifies only via the release-date branch. Its
  vendor cutoff is NOT DOCUMENTED — that is the finding, not a gap to be
  filled with third-party estimates. [R3-M02/R4-M02] The release-date
  branch rests on the alias-to-weights immutability assumption stated in
  the header; the R3-B04 pinned run identity above fixes the requested
  alias at run time, but the assumption itself is not verifiable and is
  disclosed as a residual contamination channel.
- **Gemini 3.5 Flash**: cutoff (Jan 2025) and listing (2026-05-19) are both
  pre-June, but the arm is context-only by design: it drafts the question set,
  which is labeled outcome-exposed regardless.