Prior Work Source Authority Promotion Handoff
A durable receipt for the thing Ali kept pointing at: harvest and verify prior validated work before spending more tokens on title-level review.
This milestone makes the anti-loop rule executable. If E5/E6 ledgers, source-authority receipts, donor salvage maps, export manifests, and validation proof already account for a title/source family, the engine blocks redundant title-level LLM review and routes the work to deterministic promotion or repair instead.
The repo now has a promotion gate, not another plan.
The new command is prior-work-source-authority-promotion. It writes
validation receipts, JSONL row proof, a deterministic repair queue, and paired
Markdown/HTML reports. The main validation command now includes its summary, so
future runs cannot silently drift back into title-by-title rework.
uv run aina-data-engine --root /srv/aina/aina-data-engine-room prior-work-source-authority-promotionThe receipt passes and blocks broad title re-review.
| Measure | Value |
|---|---|
| Status | pass |
| Missing assets | 0 |
| Title-level LLM review blocked | 19 assets |
| Trusted jobs-research titles counted | 15,104 |
| Clean candidate rows counted | 44,440 |
| Current vector count | 151,983 |
E5, E6, registry, export, vector, and salvage proof are now linked.
The accounted set includes the E5 title ledger, the E6 mapping-chain ledger, source-authority registries, the Academy-safe export manifest, vector reconciliation, and donor retirement/salvage route maps. This does not claim full runtime perfection. It says title-level LLM review is the wrong next move for these already-accounted families.
| Area | What it proves |
|---|---|
| E5 title lineage | Title/service-tier history is already recorded. |
| E6 mapping chain | Role-to-workflow topology and gaps are named. |
| Source authority | Current registry rows, chunk families, and authority classes pass. |
| Export manifest | Top 500/top 1,000 Academy-safe payloads are readable and bounded. |
| Vector reconciliation | 151,983 vectors exist with zero stale vectors, but vectors are not runtime truth. |
Remaining work is deterministic and named.
| Priority | Class | Next move |
|---|---|---|
p0 | mapping_chain_gap | Close role-to-workflow promotion with existing workflow intelligence. |
p0 | runtime_source_authority_gap | Complete bridge or abstain decisions for incomplete top-band rows. |
p1 | export_workflow_coverage_gap | Add workflow/practice refs from mapping-chain assets. |
p1 | clean_candidate_promotion_backlog | Promote clean candidates through deterministic family gates. |
p1 | donor_lineage_delta | Diff donor/salvage candidates against the current registry. |
The reviewers narrowed the milestone to proof, not hope.
Codex subagents handled source mining, data-integrity review, and adversarial review. Claude Opus ran as a read-only CLI checkpoint copilot. All converged on the same operating constraint: old LLM verdicts are provenance, not authority, and the closeout must be the receipt plus deterministic exception ledger.
Promote useful donor work by deterministic diff, not broad LLM review.
Start with aina-jobs-research source-intelligence exports and the
6,866 workflow seed queue, then wire aina-core canonical evidence
parquets where hash sidecars already exist. Treat old PE scripts and schemas as
concept/script donors unless current receipts promote them.
Run the new promotion gate first; then repair the named deterministic gaps.