shared/backlog/post-portata-production-learning.md

# Deferred Skill-OS backlog — post-PORTATA production learning

Status: `DEFERRED_EVIDENCE_REQUIRED`. This is not a productive skill and does not authorize access to, modification of or learning from a project without a separate scoped request.

## Candidate

`extract-production-playbook` may be specified only after a complete eligible production case has passed final production verification. PORTATA is the first proposed case, not automatic evidence access and not a universal template.

The later empirical review must classify findings as `PROJECT_SPECIFIC`, `REUSABLE_PROJECT_PATTERN`, `REUSABLE_SKILL_RULE`, `HARNESS_RULE`, `QA_RULE`, `ANTI_PATTERN`, `DIAGNOSTIC_PATTERN`, `TEMPLATE_CANDIDATE` or `NOT_REUSABLE`. No project detail becomes a global rule without repeated or otherwise sufficient evidence and independent review.

## Evidence collection package

Collect only reliable, attributable evidence for:

1. documentation actually used, unused, missing or redundant; active-source slices, hidden sources and machine-readable candidates;
2. useful and noisy durable instructions, local skills, subagents, hooks, permissions and freshness guards;
3. session health transitions, resets, re-entry sources and avoidable context load;
4. material tasks, attributable first-pass outcomes, revisions, stops, contradicted claims, stale-runtime and failed-edit incidents;
5. visual gates, representation proofs/changes, rework, Class-2 outcomes and owner-review burden;
6. initial versus final root-cause classifications, bounded-iteration stops and wasted parameter searches;
7. parallel work, duplicate work, reviewer findings, independence, capability limits and required ratification;
8. build/runtime/browser/device/responsive/accessibility/performance/memory/rendering/backend/error/monitoring/regression evidence;
9. failures first detected at final release, earlier detection opportunities, redundant checks and rollback readiness;
10. measured usage, rework, repeated reads, review burden, bottlenecks and time savers where provenance supports measurement.

Preserve raw records, task/attempt IDs, source and harness versions, revision/snapshot, timestamps/timezone, reviewer provenance, environments, missing coverage and measurement limitations. Do not fabricate counts, costs, tokens or causality.

## Activation gate

After the complete case and `$verify-production-implementation`, start a separate `POST-PORTATA EMPIRICAL SKILL-OS REVIEW`: inventory the case, use evidence rather than memory, separate project-specific from reusable, specify and red-team the candidate read-only, then implement only approved evidence-backed changes. Re-audit existing skills without presuming any must change.