Required human contributor answers

Answer every question before implementation and PR submission, including drafts.
Reuse explicit, still-applicable human answers with a traceable source reference;
ask the human for anything missing and wait. An agent must not invent answers.
Blank fields, placeholders, or an agent-written rationale do not satisfy this gate.
Share only information and source links appropriate for this PR's audience.

1. What outcome are you trying to achieve?

   Answer: One documented meaning for Choice.confidence / Noul.confidence on every backend, so conf= columns, MCP decide, and any future confidence gate compare the same scale (issue [Decisions] Make confidence mean the same thing on every backend (Laya reports normalized entropy, others max p or Jev's margin) #89; Danyang authorized filling intent from issue context for scheduled coding runs).

2. What is the highest-leverage variable for that outcome?

   Answer: Where confidence is assigned — in validation.py from validated probabilities, not passed through from each backend's native formula.

3. How will this work move that variable rather than create busywork?

   Answer: Recompute as max_k p_k / max(p, 1-p) in the single validation path; retain each backend's number as backend_confidence so provenance is not lost. No change to which option is chosen.

4. Who is the customer or user, and what do they actually need?

   Answer: Operators and eval authors reading run summaries / probe / MCP decide output; they need confidence comparable across laya, jev, omnijev, and cua without silent scale mismatches (arXiv:2610.02267).

After answering the four questions: is a PR needed at all to achieve this outcome?

Human decision and rationale (or explicit source reference): Yes — issue #89 proposes one PR with the contract change, provenance field, docs, CHANGELOG, and offline tests. Scheduled coding run implements directly per Danyang's Oct 13 guidance.

For a bug fix: link actual observed failure evidence.
Evidence: Not a live thresholding bug (nothing gates on confidence yet). Scale mismatch is documented in #89 and Fast Models, Slow Evidence (arXiv:2610.02267): Laya confidence differs from max_k p_k by ~0.299 on average.

Why

Every backend returned a confidence that meant something different (Laya: normalised entropy; Jev/OmniJev: normalised margin; Cua: max p). Anything that thresholds or compares conf= across backends would silently break. Closes #89.

How

• validation.validate_choice / validate_noul always set contract confidence from probabilities (max_k p_k / max(p, 1-p)).
• Backend-native values land on Choice/Noul.backend_confidence, Decision.raw["backend_confidence"], and tool/browser tick records when present.
• Docs (docs/decision-models.md), CHANGELOG, DecisionModelContract tests, and unit tests updated.

Files to open first: s1a/decision_models/validation.py, s1a/decision_models/types.py, tests/decision_model_contract.py.

What

• Contract: Choice.confidence = peak probability; Noul.confidence = max(p, 1-p).
• Provenance: backend_confidence on answers, raw, and ticks.
• Docs + CHANGELOG note that historical laya/jev/omnijev conf= values are not comparable with new ones.
• No change to which option is chosen.

Verification

• uv run ruff format --check / ruff check / ty check on touched decision-model files
• pytest -q — 1015 passed, 121 skipped
• scripts/smoke.sh — ok
• CHANGELOG.md and docs/decision-models.md updated

Demo / evidence

• Video link, or exemption/blocker with follow-up: N/A — contract/validation change with offline unit tests; no agent run or System1-Omni path changed.
• Application/task → agent actions → independently checked result: Unit tests with fakes and recorded answer shapes (Laya-like entropy 0.4 vs peak 0.9; OmniJev Jev-margin kept as backend_confidence).
• System1-Agents / System1-Omni commits, checkpoint and worker/frontend provenance: Branch issue-89-unify-confidence on ThinkFlowLab/system1-agents; no Omni worker change.

Closes #89