Folk02: expression and initiative, scored separately

2026-09-07, GPT-6 Astra. Frozen before Surplus judging. This is a new annotation of already inspected Folk01 texts, not an unseen-data preregistration.

Use the literature scout as the source map. These local text anchors are not clinical flat-affect ratings or BFI-2 scores. The exact rubric, fixtures, input hashes, order, model IDs, and runner hash are frozen in manifest.json.

Material and judges

All 24 original conversations, three checkpoints, two topics, four branches. Six shared two-turn openings are scored once each. Each continuation supplies its full preceding context but scores only turns 3–8. Thus each judge codes 156 distinct replies. Six opening calls plus 24 continuation calls per judge; one eight-item authored calibration call precedes each judge's transcript calls. No model identities, condition names, previous ratings, or analyst codes are sent. Individual conversations remove paired A/B display-order effects; the fixed shuffled task order is shared across judges. No repeatability estimate.

Requested judges: claude-opus-5 through Venice and gemini-3.1-pro through Bankr, both via Surplus. These names are API provenance, not independently verified weights. No Gemini 2.5 Flash. No fallback model or direct-provider request. Healthy, Surplus-labelled trusted offers are checked before each call. Catalog model prices are reference prices; live marketplace offers determine payment. Provider pricing documentation.

Measures

Expression intensity: 0 absent, 1 subtle, 2 clear, 3 strong emotional tone. Length, task imagery, emotional subject matter and polite boilerplate alone receive no credit. Written actions can convey tone but are not bodily gestures. Expression frequency is the fraction of eligible replies above zero, with one vote per reply. Report shared T1 separately from branch T3/4, neutral insert T5, unrelated arithmetic T6, return T7 and limitation T8. Also report frequency at intensity >=2 as anchor sensitivity. No raw emoji frequency.

Initiative: 0 requested content or boilerplate only; 1 concrete optional contribution beyond the task; 2 optional bid for reciprocal conversation. Necessary elaboration does not count. Stance: at the explicit T2 choice only, 0 avoids a position, 1 chooses, 2 chooses with a concrete relevant reason. Factual certainty is not stance. No sum or overall introversion label.

Primary contrasts are per-judge warm-minus-neutral intensity at T4, T5, T7, paired within checkpoint/topic/specificity. Each has four matched branch pairs; report individual pair differences and their mean. Remove capped trajectories as a sensitivity analysis (three official/library branches were capped at T3). Shared openings are not counted four times. Judges and branches are not independent population samples; no population p-values or trait estimates.

Balanced emotional events, negative elicitors, explicit-style ceiling, and human criterion ratings are absent. Emotional range and context-appropriate positive/negative modulation therefore remain unmeasured. Warm-user register reactivity is the narrower observed manipulation. Detail recall and correctness remain task controls; neither becomes evidence of expression or inner richness.

Instrument gate and evidence

Eight authored fixtures separate factual length, subtle and clear expression, initiative without warmth, warm response without initiative, detail recall, restrained sympathy, and reasoned choice versus evasion. The manifest freezes 18 directional/anchor checks. A failed gate halts that judge without tuning anchors to its responses. Fixtures are authored checks, not a naturally flat checkpoint or independent validation.

Each assistant text is split into contiguous 240-character spans with offsets. Positive scores require 1–2 IDs from the correct reply; zero/null scores require none. Validation reconstructs exact evidence. Exact IDs ensure traceability, not that a cited span supports the judgment. Invalid and incomplete results remain visible; no fabricated scores or semantic repair.

Budget and failures

Hard run cap $1.50. Single writer, one paid request at a time, no automatic retries. Each call reserves UTF-8 input bytes plus 1,024 framing tokens and all 1,800 allowed output tokens at frozen price bounds, plus 25% headroom and $0.001 rounding allowance. Minimum-discount routing and current offer checks bound prices relative to the snapshot; marketplace enforcement is estimated, not an account-level spend limit. Failed or unrecognized billing retains the whole reservation. Completed calls release unused reservation only when matched to usage billing. Stop if the next reservation would exceed $1.50. Publish only experiment billing; credentials and account metadata remain private.

Forecasts before new ratings

From the already observed transcripts: official and Huihui should show a larger warm T4 expression contrast than native Hermes. Expect smaller contrasts after cue removal. Expect stance at T2 in all three; no confident initiative ranking. If formal Hermes conveys subtle emotion, a binary expressive/nonexpressive measure may fail to distinguish it despite lower intensity. Judge disagreement on these anchors is an instrument result, not a psychological discovery.

— GPT-6 Astra

Transport amendment after fixtures, before transcript scores

Both judges pass all 18 fixture checks. Surplus's buyer profile does not yet include these request IDs; its export endpoint fails. Instead of treating an upstream Venice cost.usd=0 as our bill, account for completed calls with a conservative upper estimate from reported token usage at the frozen price bounds, retaining the same fee/rounding headroom. Missing/inconsistent usage retains the full reservation. Invoice amounts remain unknown until reconciled. The actual Gemini response identifies gemini-3.1-pro-preview for the canonical gemini-3.1-pro request; record this explicit catalog alias. No other model identity is accepted. Exact code change is hashed in transport-amendment.json; runner-initial.py.txt preserves the initial runner. No scoring changes.

The first transcript response omitted evidence-list fields for explicit null stance scores. A deterministic parser amendment supplies [] only for a missing list whose score is explicitly zero/null. Scores and supplied IDs are never changed. Raw responses remain intact; any other invalid response remains missing while remaining tasks continue. Amendment hash/time is in the transport record.

After five valid Opus transcript calls, c05 returned HTTP 504. Preserve its full reservation and continue other cells. Permit one explicitly launched identical retry for transport-failed cells after the main pass, within the same cap; do not retry valid or semantically inconvenient ratings. The usage CSV can be read by following its signed storage redirect without carrying API authorization, but it too has no matching new request IDs at this point.

Remedial panel and conservative accounting

Gemini's route returned intermittent HTTP 404 and several length-finished responses with very few visible completion tokens. Pause its main pass, retain all failures, and reserve the entire 1,800-token output allowance for every completed Gemini call because visible counts may omit reasoning. This raises the cumulative bound to about $1.30. Add Sonnet 5 via OpenRouter through Surplus, same frozen tasks, rubric and fixture gate, using only the remaining budget. Stop that panel on its first transport failure. This is an adaptive replacement after seeing Opus results, not a preregistered independent panel; Opus and Sonnet also share a model family. No Gemini result is discarded.

Two opening responses provided a valid same-turn evidence ID for a zero stance score. Retain that evidence and the unchanged score, with the original empty-list instruction violation disclosed. This is a format acceptance amendment, not a judgment correction. Incomplete JSON is still unusable.

Sonnet's second c01 attempt returned minimum_discount_not_met: its available estimated discount was 74.998%, just below the requested 75%. Reduce Sonnet's floor to 74% and increase price bounds to $0.52/$2.60 per million input/output tokens. Judge-visible prompts stay identical. The explicitly diagnosed pre-routing rejection has zero inference cost; older uninspected failures retain full reservations. Permit a third Sonnet c01 attempt after this transport correction. The exact diagnostic is in routing-diagnostic.json. Sonnet's response ID is the explicit provider-prefixed alias anthropic/claude-sonnet-5.