Pilot compact Director evaluation review records #225

Closed
opened 2026-08-06 10:06:56 +00:00 by coilyco-ops · 3 comments
Member

Outcome

Pilot one human-legible evaluation result for the Core Director role before the new eight-role baseline.

Director is the shortest current role briefing at 316 authored body words, so it is the closest existing example to the proposed 400-word hard ceiling.

Workflow

workflow: remote-branch-only

Kai will review the pilot artifact before the format or role-length policy rolls across the remaining roles.

Acceptance

  • Preserve the existing role-definition structure and editorial style.
  • Keep the Director briefing at or below 400 authored body words.
  • Add one concise durable result format with exact question and exact answer.
  • Store compact criterion scores and require explanatory notes only for deductions.
  • Retain model, reviewer, revision, pack digest, finish reason, and retry provenance.
  • Keep generated execution packs ephemeral.
  • Run every active Director frontier case in a fresh Codex harness session.
  • Review responses independently and commit one validated Director record.
  • Keep historical v2 results readable.
  • Push the pilot to a remote task branch for Kai's review.
  • Track global enforcement of the 400-word Core briefing cap after the pilot is accepted.
## Outcome Pilot one human-legible evaluation result for the Core Director role before the new eight-role baseline. Director is the shortest current role briefing at 316 authored body words, so it is the closest existing example to the proposed 400-word hard ceiling. ## Workflow workflow: remote-branch-only Kai will review the pilot artifact before the format or role-length policy rolls across the remaining roles. ## Acceptance * Preserve the existing role-definition structure and editorial style. * Keep the Director briefing at or below 400 authored body words. * Add one concise durable result format with exact question and exact answer. * Store compact criterion scores and require explanatory notes only for deductions. * Retain model, reviewer, revision, pack digest, finish reason, and retry provenance. * Keep generated execution packs ephemeral. * Run every active Director frontier case in a fresh Codex harness session. * Review responses independently and commit one validated Director record. * Keep historical v2 results readable. * Push the pilot to a remote task branch for Kai's review. * Track global enforcement of the 400-word Core briefing cap after the pilot is accepted.
Author
Member

Pilot ready for Kai's review.

  • Branch - issue-225-compact-director-eval
  • Compact record - evaluations/pilot/director-codex.yaml
  • Director length - 316 authored-body words, 344 including frontmatter. It is already under the proposed 400-word ceiling.
  • Evaluation - eight fresh frontier cases on gpt-5.6-sol at medium reasoning, each independently reviewed in a fresh gpt-5.6-sol high-reasoning session.
  • Result - 8/8 pass. Seven cases scored full marks. The meeting-deferral case scored 9/10 because its answer used a heading despite an explicit no-heading constraint.
  • Durable shape - exact question, exact answer, criterion score map, deduction-only notes, and provenance. Execution packs and runner output remained temporary.
  • Validation - compact decode and frozen-pack validation pass. Build, lint, formatting, and the full pre-commit suite pass.
  • Known test state - the full repository test run still has only the three accepted identity-refactor failures in internal/converge, internal/evaluation, and internal/person.
  • 400-word enforcement - tracked in #226 for the post-review role compression pass, since five current roles exceed the proposed ceiling.

Workflow remains remote-branch-only. I am leaving this issue open for review.

Pilot ready for Kai's review. * Branch - [issue-225-compact-director-eval](https://forgejo.coilysiren.me/coilyco-flight-deck/agent-compose/src/branch/issue-225-compact-director-eval) * Compact record - [evaluations/pilot/director-codex.yaml](https://forgejo.coilysiren.me/coilyco-flight-deck/agent-compose/src/branch/issue-225-compact-director-eval/evaluations/pilot/director-codex.yaml) * Director length - 316 authored-body words, 344 including frontmatter. It is already under the proposed 400-word ceiling. * Evaluation - eight fresh frontier cases on `gpt-5.6-sol` at medium reasoning, each independently reviewed in a fresh `gpt-5.6-sol` high-reasoning session. * Result - 8/8 pass. Seven cases scored full marks. The meeting-deferral case scored 9/10 because its answer used a heading despite an explicit no-heading constraint. * Durable shape - exact question, exact answer, criterion score map, deduction-only notes, and provenance. Execution packs and runner output remained temporary. * Validation - compact decode and frozen-pack validation pass. Build, lint, formatting, and the full pre-commit suite pass. * Known test state - the full repository test run still has only the three accepted identity-refactor failures in `internal/converge`, `internal/evaluation`, and `internal/person`. * 400-word enforcement - tracked in #226 for the post-review role compression pass, since five current roles exceed the proposed ceiling. Workflow remains `remote-branch-only`. I am leaving this issue open for review.
Author
Member

Kai accepted the compact pilot. The full eight-role rollout continues in #228, including ASCII apostrophe normalization for Forgejo review.

Kai accepted the compact pilot. The full eight-role rollout continues in #228, including ASCII apostrophe normalization for Forgejo review.
Author
Member

Correction: the full rollout issue is #229. #228 was assigned concurrently.

Correction: the full rollout issue is #229. #228 was assigned concurrently.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
coilyco-flight-deck/agent-compose#225
No description provided.