Teacher round template

Copy this structure to docs/<game>/teacher-rounds/<entry-id>.md, fill it in and link it from that game's teacher log. This is a documentation template, not an executed proposal, a JSON schema or an automatically generated runtime log. Replace placeholders with actual records; unavailable values are not zero.

Identity and status

• Entry, study, round and candidate IDs; parent entry/program ID.
• Recorded/proposed/evaluated timestamps with timezone; code/optimizer revision.
• Track: direct policy, value prediction or another explicitly defined contract.
• Status: planned, proposed, evaluating, accepted, rejected or inconclusive.
• Provenance: prospective, retrospective, imported or external model invocation.

Teacher and visible evidence

• Actual teacher provider/model/version if available; requested versus returned ID.
• For an interactive assistant, record that invocation mode and known identity; never infer a historical version from the assistant's current name.
• System instructions, visible packet and public proposal/response artifact paths and hashes; sampling settings and evidence-selection rule.
• Evidence IDs, split, source trajectories and any previous development exposure.
• Context isolation method, prior memory, accepted/rejected proposals visible.
• If no exact transcript exists, say so. Do not reconstruct one as an original.

Proposed intervention

• Full parent/candidate program artifacts and hashes; exact changed fields/diff.
• Edit operator and target component; changed versus held-fixed contracts.
• Evidence-based problem description and concise stated rationale.
• Hypothesis, predicted action changes, applicable situations and regression risks.
• Required validation and the condition that would disconfirm the hypothesis.

Frozen evaluation plan

• Game profile version and study-protocol path/hash.
• Train/development/final-test separation; teacher and selector visibility.
• Primary endpoint, seed/repeat counts, episode caps and selection rule.
• Probe set/hash, old/new repeats, noise comparison and situation definitions.
• Teacher/Jev attempt caps, retry/stop rules and total environment budget.
• Record the proposal and protocol before results; do not revise the gate afterward.

Observed results and resources

• Per-seed old/new scores, paired differences, denominators and uncertainty.
• Completion, truncation, failed attempts, missing evaluations and deviations.
• Probe action flips/distribution changes and repeated-query noise, or not run.
• Teacher and executor calls/tokens/costs separately; actual versus estimated cost.
• Environment collection/branch/evaluation/replay frames separately; simulation time, API time and end-to-end time separately.
• Machine-readable results, original exchanges, videos, archive manifest and audit.

Selection and bounded lesson

• Selected program/hash; accepted/rejected/inconclusive and measured reason.
• Separate algorithmic gate outcome from a later manual research-reference choice.
• Which hypothesis survived or failed; alternative explanations and regressions.
• Evidence level: exploratory association, replicated pattern or controlled mechanism.
• Applicable conditions, contradictory cases and proposed next comparison.
• Later evidence is appended with dates/links; original run artifacts stay immutable.