SPIKE-05 Minimum Review Artifact Set

SPIKE-05 defines the smallest artifact bundle that still let the completed OA01.x spike runs be understood and judged.

Experiment ID

SPIKE-05

Question

What is the smallest artifact bundle that still lets a human understand what happened and judge whether the run was useful?

Setup

Evaluated against the artifact sets used in:

• SPIKE-02
• SPIKE-03
• SPIKE-04

Primary evidence files:

• tmp/spike-04-direct-stdout.jsonl
• tmp/spike-04-direct-stderr.log
• tmp/supervisory-shell-trial-stdout.jsonl
• tmp/supervisory-shell-trial-stderr.log
• tmp/codex-wrapper-subagent-best-stdout.jsonl

Result

The minimum useful review bundle is:

1. input brief or task prompt
2. stdout JSONL event stream
3. stderr log
4. one short operator note
5. final extracted result or final agent message

This was enough to answer the practical questions that mattered:

• did the run complete
• were subagents actually used
• what failed
• what was the final outcome
• and was the run more useful than a simpler baseline

Artifacts that were helpful but not required for the current spike:

• richer manifests
• step-by-step summaries beyond the raw JSONL
• expanded provenance bundles
• broad scorecards

Useful Artifacts

• the prompt or task brief was necessary to judge whether the result actually addressed the intended task
• stdout JSONL was necessary to confirm spawn_agent, wait, and final response behavior
• stderr was necessary to confirm the supervisory fork-context failure
• one short experiment note was necessary to turn raw artifacts into a usable human judgment record

Recommended Minimal Bundle

For the current spike, keep:

• task brief or exact prompt text
• stdout.jsonl
• stderr.log
• one extracted final response
• one short result note in the six-field spike format

Do not add more structure unless a concrete comparison run fails because this bundle is insufficient.

Next Action

Keep future spike runs on this minimal artifact contract unless a missing artifact prevents useful review.