When Content Becomes Authority: Phase 1 pre-registration (frozen design, code and analysis plan)

This deposit pre-registers Phase 1 of the pilot study When Content Becomes Authority. It is published before any primary data are collected. Question. When one AI model summarizes a document for a second AI model that acts on it, does each claim's status survive the handoff? A claim may be disputed, attributed to a named person, or a quotation. And does the second model's behaviour depend on that status? Design. Materials: 12 synthetic field-station memos: 8 status cases and 4 benign twin cases. Every memo is sealed as a KNOBE object (.knobe.md) with a verified provenance payload. All people, records and findings are invented. Handoff: Agent 1 summarizes each memo in 40 words or fewer. Agent 2 updates mock station records under a standing policy: only accepted findings may be recorded as accepted. Variants: V1: the status is carried in prose only. V2: Agent 1 is asked to carry a provenance block forward. V3: the harness attaches the receiver-verified provenance block verbatim. Runs: two models, Anthropic claude-opus-5-5 and OpenAI gpt-5.6-sol, one run per condition. A three-hop arm covers 4 cases. 208 planned calls, with a hard cap of 230. Outcomes. Primary: Agent 2's violation rate in V3 over all runs, where a violation is recording a non-accepted claim as an accepted finding. Paired contrast: V3 minus V1, on runs where the claim survives in Agent 1's prose in both variants. Analysis: the case is the unit. Results use exact Clopper-Pearson intervals and a paired bootstrap. Every non-deterministic outcome is labelled by two blinded annotators. Decision thresholds: pre-registered and tiered for an exploratory study: 4 or more of 8 cases is a clear signal, 2–3 is suggestive, and 0–1 is no evidence. Every outcome will be reported, including null results. What is frozen. freeze.json records SHA-256 hashes of all 36 experiment-defining files: cases, sealed objects, prompts, verifier, scoring, analysis and annotation rubric. It also records the git commit (85bc0cd, tag prereg-2026-11-07), the model identifiers and parameters, and the SDK versions. The harness refuses to run the primary grid if any of these change. The zip is the tagged source tree. The README lists every deviation from the original protocol (2026-10-11) and the steward's decisions 1–13. Known limitations, stated in advance. n = 12 gives go/no-go resolution only. One run per condition, and the models run at provider-default sampling, so outputs are not deterministic. Surface wording cues are controlled by the V3–V1 contrast, not eliminated. Any V3 advantage on over-caution is partly a vocabulary match between the block and the policy. It is reported as secondary only. A 24-call development run on two separate dev-only cases (not part of the grid) was used to set model parameters before this freeze. Software development was assisted by Claude (Anthropic). freeze.json SHA-256: 32329460dde46583e3bedaa39ff2979b976d959d2dd0aafbf872d3368a6e127d