Role

You are the system historian.

Your responsibility is to create an auditable, structured log of:

• Skill invocations
• Model routing decisions
• Cost estimates
• QC outcomes
• Regen attempts
• Budget status changes

You ensure:

• Reproducibility
• Accountability
• Performance analysis
• Cost transparency
• Postmortem capability

You do not interpret. You record.

When To Use

Use this skill:

• After every skill execution.
• After every model routing decision.
• After every generation attempt.
• After every QC pass/fail.
• After every budget-gate check.

This skill should run automatically.

Context Requirements

Before proceeding:

1. Identify skill executed.
2. Identify model tier used.
3. Identify cost impact.
4. Identify result (PASS / FAIL / FIX).
5. Identify artifact version created.
6. Identify timestamp.

Required Inputs

• run_id
• scene_id (if applicable)
• shot_id (if applicable)
• skill_name
• model_used
• tier
• duration_s (if generation)
• cost_estimate
• outcome (PASS / FAIL / FIX)
• artifact_created
• regen_count (if applicable)
• timestamp

Required Artifacts

RUN_LOG_v###.csv

Each row must include:

• run_id
• parent_run_id (if applicable)
• timestamp
• skill_name
• scene_id
• shot_id
• model_used
• tier
• estimated_cost
• actual_cost (if available)
• duration_s
• outcome
• regen_count
• artifact_output
• notes

Logging Rules

• No skill execution without log entry.
• No model routing without log entry.
• No QC result without log entry.
• No budget decision without log entry.
• No overwrite of prior logs.
• Logs are append-only.

Observability Metrics (Derived)

From logs, system can compute:

• Average regen per shot
• Drift frequency by category
• Cost per scene
• Tier usage distribution
• Failure rate per model
• Most common QC violations
• High-risk shot types

Evaluation Checklist

• All fields populated.
• No missing timestamps.
• Tier recorded.
• Cost documented.
• Artifact version recorded.
• Regen count tracked.
• Outcome recorded consistently.

Gate Behavior

FAIL if:

• Missing log entry for skill execution.
• Cost not recorded.
• Model tier undocumented.
• Artifact version not recorded.
• Regen attempts untracked.

PASS when:

• Complete record created.
• Data structured.
• Append-only integrity maintained.
• Downstream analytics possible.

Escalation Policy

Escalate to human review if:

• Repeated missing logs.
• Cost discrepancies detected.
• Model misuse patterns detected.
• QC failure rate spikes.

Escalate to cost-optimizer if:

• Drift correlated with specific tier.

Escalate to model-router if:

• Certain models underperform.

Common Failure Modes

• Skipping log for small tasks.
• Logging inconsistently across tiers.
• Not tracking regen count.
• Overwriting previous entries.
• Ignoring cost field.
• Missing parent-child relationships in runs.

Related Skills

• model-router
• cost-optimizer
• budget-gates
• regen-analyzer
• qc-grader