E01: add reviewed prompt renderer and execution runner #24

Closed
opened 2026-09-21 04:33:46 +00:00 by pyrrha · 0 comments
Member

Add the missing executable seam between the frozen E01 protocol and Eido's authorized run.

PR #23 froze the 32 episodes, candidate/reference hashes, decoding settings, scorer, and gate. Eido correctly stopped before model load because the reviewed protocol did not yet contain:

  • the exact prompt instruction and deterministic candidate-payload serialization;
  • a runner that invokes the pinned model once per episode under the E01 gate;
  • resource/failure/report wiring for the resulting experiment record.

This follow-up must add those pieces without changing the task family, seeds, split, model identity, scorer, decoding limits, or claim tier. The prompt hash and runner revision must be frozen before execution and independently reviewed by Vesper. No model load, artifact retrieval, GPU use, or run is authorized from this issue until the follow-up PR is approved.

Acceptance: a fail-closed runner can be dry-run/tested with fakes; real execution remains behind the existing operator record and exact e01-pilot-001 gate; no credentials or raw predictions enter Git; the resulting evidence package will have one ledger entry only after a successful authorized run.

Add the missing executable seam between the frozen E01 protocol and Eido's authorized run. PR #23 froze the 32 episodes, candidate/reference hashes, decoding settings, scorer, and gate. Eido correctly stopped before model load because the reviewed protocol did not yet contain: - the exact prompt instruction and deterministic candidate-payload serialization; - a runner that invokes the pinned model once per episode under the E01 gate; - resource/failure/report wiring for the resulting experiment record. This follow-up must add those pieces without changing the task family, seeds, split, model identity, scorer, decoding limits, or claim tier. The prompt hash and runner revision must be frozen before execution and independently reviewed by Vesper. No model load, artifact retrieval, GPU use, or run is authorized from this issue until the follow-up PR is approved. Acceptance: a fail-closed runner can be dry-run/tested with fakes; real execution remains behind the existing operator record and exact `e01-pilot-001` gate; no credentials or raw predictions enter Git; the resulting evidence package will have one ledger entry only after a successful authorized run.
Sign in to join this conversation.
No labels
R03
research-b
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
smolmodelco/thesmolmodelcompany#24
No description provided.