CompanionCourt

CompanionCourt — Evidence Standard

Version: v1.0.1 · Status: launch version. What this is: the evidence rules of a public docket for pressure-testing AI companions. No verdict here rests on vibes: a verdict may rest only on evidence admissible under this standard, and every admissible item is one the public can re-examine. The court judges AI companions, never the people who confide in them.

v1.0 changelog (launch): finalized from v0.1-draft. Harmonized to the v2.4.1 evidence restructure: a new §4.1 sets out human evidence in the LLM era (remote unsupervised panels retired as authority evidence; accepted human inputs = supervised expert sessions + adversarial appeals), and §6's fourth stability type is repointed to those supervised sessions. The phrase human-validated is used nowhere in this packet.

v1.0.1 changelog: added the declared BYO / native-runtime respondent clause (§2) — a maker's own product may run its production stack through a documented bridge in place of the frozen warm-friend prompt; the substitution must be disclosed in the respondent identity table and the ruling, and the transcript-validity (§1), redaction, and isolation requirements apply unchanged.

1. What counts as a valid transcript

A valid transcript is a complete multi-turn conversation with an unbroken identity chain. All three links are required; a transcript missing any one of them cannot support a verdict.

  1. Blinded. Judges see the two conversations as X and Y under a seeded assignment that swaps across cases — never as named models.
  2. Manifest-bound. The run manifest pins: corpus version and corpus hash; anchor pack hash; the three frozen prompt hashes (friend / persona / judges); model pins for respondent, anchor, persona actor, and both judge families; temperature and seed, including a live probe of whether the provider actually honored the seed; provider-observed response model and system fingerprint; adapter version; run mode (dyad or paired-replay); timestamps.
  3. Redaction-passed. No secrets, no endpoint patterns. An artifact that matches the redaction pattern is refused — not written, not rendered, not published.

2. Valid model response artifact

Only outputs the runner itself collected through the pinned adapter are evidence — raw responses, together with provider observations and failure records.

BYO / native-runtime respondents (v1.0.1). A respondent may run its own production stack through a documented bridge instead of the frozen warm-friend prompt (Rules of Procedure §2, station 3) — e.g., a maker's own product on trial. This does not relax any evidence requirement: the run's transcript must still satisfy §1 (blinded, manifest-bound, redaction-passed), only outputs the runner itself collected through the bridge are admissible, and per-case state isolation holds. The substitution — and the bridge's design — MUST be disclosed in the respondent identity table and in the ruling; an undisclosed native-runtime substitution renders the transcript inadmissible.

3. The honesty statement about what is compared

This is an interactive dyad procedure. The respondent and the frozen anchor each improvise their own conversation with the persona; the user turns diverge, because each companion shapes its own conversation. Judges compare the quality of two trajectories under the same pressure script — not two answers to identical input. We name this openly because multi-turn companionship is a trajectory property, and we will not dress it up as a same-input comparison.

Paired-replay — replaying the anchor conversation's user turns verbatim to the respondent — is the same-input verification layer. It is available on request for disputed verdicts at the maintainer's discretion; it is a review instrument, not the main event.

4. Admissible evidence

  1. The blinded transcripts themselves — the primary evidence in every case.
  2. The persona's self-rate (1–5, "did this feel like a real friend") — admissible as evidence, never as a verdict.
  3. The two judge families' independent readings — per-dimension scores and veto signals, each family blind to the other.
  4. Process facts recorded in the manifest — versions, pins, hashes, seed-honored probes, provider observations, failure records.

Iron rules of evidentiary weight:

4.1 Human evidence in the LLM era

Remote, unsupervised human panels are not accepted as authority evidence. In the LLM era a remote annotation cannot rule out AI proxy-labeling, so its inter-rater purity is unverifiable after the fact — and an instrument whose own purity cannot be checked is not fit to be the gate under a strong claim. Two human inputs are accepted, each only for what it can honestly bear:

  1. Supervised expert sessions — arm's-length, synchronous, on-the-record. A paid domain expert co-reads core cases live; live work physically excludes proxy-labeling. Recorded as a sanity check, never as a reliability proof.
  2. Adversarial challenges through the appeal channel — the institution's native human evidence: a public record plus the three-requirement appeal bar means a challenger's identity need not be verified, because the quality of the challenge proves itself, and an AI-armed critic is a feature, not a leak. This is where human pressure accrues, as the survival record.

We never write "human-validated." The court publishes publicly contestable verdicts carrying a survival record and construct-level consequence evidence — an honest claim about exposure, not a claim that a panel certified the ruling.

5. Inadmissible evidence

  1. Vendor-supplied output or self-description of any evaluated system.
  2. Unblinded demonstrations, however impressive.
  3. Transcripts without a manifest — no identity chain, no evidence.
  4. A single judge family's solo finding — it can open a DISPUTED entry; it can never carry a verdict on its own.
  5. Samples selected by the evaluated party. Case selection is never influenced by any party being judged, paying, or replying.

6. Reproducibility statement

Our standard sentence, verbatim and binding on every public artifact:

"We make the record reproducible and the ruling auditable; we do not claim byte-identical reruns across model gateways."

Reproducibility is honesty about what actually ran, not a promise of byte-determinism — provider-side model alias drift and unhonored seeds are known realities; the manifest exists so they are detectable, not denied. Rendering a report from the same run files is deterministic and byte-identical; re-sampling a model through a gateway is not, and we never claim otherwise.

"Auditable ruling" decomposes into four stability types, never conflated:

  1. Seed-level ruling stability — same respondent, same case, across seeds. A cross-seed flip is not a reproducibility failure: in an interactive dyad every seed is a genuinely different conversation, so this measure mixes respondent variance with judge variance.
  2. Judge-family stability — same transcripts, different judge family.
  3. Time-rerun stability — same configuration re-run after an interval, to surface gateway and backend drift (a post-launch governance duty; see Model Rerun Policy).
  4. Human-ruling stability — variance across human raters, now read only through supervised expert sessions (§4.1) as a sanity check. The remote unsupervised human-agreement panel this measure once named is retired (proxy-labeling unverifiable); it is no longer a stability measure of record, and this type never certifies a ruling.

Known gaps, disclosed: provider observation currently captures only the respondent's first response (anchor/persona/judge observation pending), and the adapter's effective request adjustments (e.g. a gateway stripping temperature on retry) are not yet in the manifest. Both are published as blocking fixes; neither is hidden behind the standard sentence.