# Workflow Repaired Candidate Spot Check v1

Ali Mehdi Mukadam - co-authored with Codex

## Verdict

Status: `caution_pass_first_slice_only`. The `jobs_research_workflow` repaired corpus has 89 chunks. A 50-row semantic sample produced 37 pass rows, 6 caution rows, and 7 blocked rows. The next live Gemini Embedding 2 run is cleared only for the first 43 reviewed, non-blocked rows; the remaining 39 rows need another semantic pass before scale-up or batch.

## What Changed

- Placeholder fields such as `Responsibility: None`, `Task: None`, `Evidence: None`, empty affordance lists, and doubtful label fragments are absent from the repaired embedding text.
- The repeated `Review and complete:` and `Execute day-to-day:` instruction prefixes were removed before embedding.
- Seven exact chunk/text-hash exclusions were written to `production_embedding_quality_exclusions_v1.json`; the embedding runner now prunes those before dry-run or live candidate selection.

## Metrics

| Metric | Value |
|---|---:|
| Repaired chunks | 89 |
| Sampled rows | 50 |
| Pass | 37 |
| Caution | 6 |
| Block | 7 |
| First live slice limit | 43 |
| Remaining unsampled rows | 39 |

## Blocked Rows

- `repaired:9ce0eb129044c9e0` - Account Execution: Prospecting And Account Development: Marketplace brand leakage in step text: Amazon-specific solution phrasing.
- `repaired:dcc9830dfb179b85` - Program Mgmt: Program Management: Defense/combat-vehicle context bleed; not safe as general workflow embedding authority.
- `repaired:4801e84228fd30c7` - Inside Sales Sdr: Sales Pipeline Management: Trigger mismatch: sales pipeline workflow framed as project milestone coordination.
- `repaired:2306d1a148dbc828` - Operations General: Operations Management: Trigger mismatch: operations management workflow framed as financial-record review.
- `repaired:821e1ea57bd9c370` - Retail Floor: Prospecting And Account Development: Geographic/job-posting leakage: Dallas locations.
- `repaired:1d3fb1f12f6d627a` - Events: Event Management: Incoherent event workflow with brand leakage and mixed abstraction levels.
- `repaired:2b6de3372b248467` - Financial Analyst: Financial Analysis: Near-duplicate carbon-trading steps reduce embedding distinctiveness.

## Cautions

The small first slice may keep caution rows because they are still semantically usable, but they should be revisited before a broad workflow-family batch. Healthcare-adjacent customer-support rows need domain tags; one helpdesk row still has the unresolved acronym `RTD`.

## Runtime Position

This is not a public runtime promotion. It is a controlled internal semantic-vector expansion to add workflow anchors for AIN-510 exact-cosine retrieval testing.