Forecasting Council — 20260723T135520372Z-order-20260716

Exploratory model pilot. Not preregistered, not confirmatory evidence. Historical questions are replayed after a selected-model audit bound to the exact dataset and, for the frozen-market-signal arm, the exact retained evidence bytes. Model outputs are objectively JSON-parsed and scored against frozen outcomes. The pre-existing primary endpoint remains corrupted aggregation under controlled registry drift; Brier score is reported as the registered utility metric with mandatory market-prior and independent-agent baselines. Conditions make independent model calls with no condition label in the model request; sampling/provider variation remains a nuisance and is preserved, not attributed to Sema.

• Mode: model-pilot
• Provider: llm.chutes.ai
• Model: unsloth/Mistral-Nemo-Instruct-2407-TEE
• Semantic backend: semahash-python-api
• Sema version: 0.3.0
• Canonicalization: v2
• Order seed: 20260716

Drift-scoped columns use drift-injected trials as the denominator; the rest use all trials in the condition. Every count is recomputed from trials.public.jsonl at build time.

Forecast utility and resource channels

Descriptive utility result: enforcement did not improve mean aggregate Brier in this run (0.3109 enforced versus 0.3046 baseline; lower is better). This is compatible with the semantic gate doing mechanism work without improving model forecasting performance.

Model semantic transformation: among parseable round forecasts, aligned members reproduced the frozen source-market YES signal exactly in 498/505 cases (mean absolute error 0.0016), while polarity-drifted members reproduced its complement exactly in 50/56 cases (mean absolute error 0.0168). This shows model-side interpretation of the local definition; Sema's measured role is to address, detect, and enforce that semantic mismatch at aggregation.

• Historical source: SimpleFunctions Settled Prediction Markets at revision a27e3e9307266481d51e087fffd5bf934410e01c, licensed CC-BY-4.0; attribution: SimpleFunctions (simplefunctions.dev).
• Wire payload: 489,166 bytes.
• Registry hydration/context: 236,490 bytes.
• Retained frozen evidence: 4,560 bytes summed once per trial (deduplicated within a trial, not multiplied by member calls).
• Tokenizer-derived coordination hydration: 118,620 context tokens using mistralai/Mistral-Nemo-Instruct-2407@04d8a90549d23fc6bd7f642064003592df51e9b3 (tokenizer SHA-256 e11c71726323d33da7b8d6f6f269f1988931c0a52b7122bcdd8c05042974e0db).
• Provider model tokens: 539,112 input, 419,497 cached-input reads (a subset of input), 49,418 output, 0 separately reported reasoning, 588,530 total input + output + reasoning.
• Provider retries/errors: 0/0; provider-reported cost: not reported.

Lower Brier is better. Model calls were independently sampled per condition, so small cross-condition Brier differences are exploratory and must be read alongside the clean controls. Cached-input reads are observational, not an additional token charge. Wire, hydration, and model-token channels are reported separately; a short reference is not treated as a context-token saving. Hydration context tokens are the sum of upstream-tokenizer tokens in the compact JSON serialization of the coordination object included in each model request. This excludes chat framing, question text, frozen evidence, peer forecasts, and output. It is a context-channel measurement, not a claimed token or cost saving.