ms-ai-architect/docs/r11-flag-format-2026-07.md

92 lines
4.1 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# R11 flag format — the judge-pass → fix-work-list contract
**Spec for the flag record the R7R10 judge pass emits and R11 (the fix step)
consumes. A flag is a judged claim the pass could NOT ground against its cited
Microsoft Learn source. The pass never fixes — it stamps (born-verified) or
flags; R11 does the human-confirmed fix.**
Status: design spec (R6 Step 5). Builds on the frozen v3.1 claim judge
(`scripts/kb-eval/judge-claim-prompt-v3.1.md`) and the judge-pass ledger
(`scripts/kb-eval/judge-pass-manifest.mjs`, R6 Step 4). Consumed by R11.
---
## Where a flag comes from
The v3.1 judge returns, per claim, a strict-JSON result (no fence):
```
{"file":"<FILE>","results":[
{"id":"<claim id>","judge_verdict":"grounded|not_grounded|source_silent",
"rule":"<R1-R8 or empty>","evidence_url":"<url actually used>",
"evidence_quote":"<verbatim quote or empty>","reason":"<one sentence>"}
]}
```
The pass augments each **non-grounded** result (`judge_verdict !== 'grounded'`)
into a flag record. This is a **minimal augmentation, not a pass-through** — the
judge output is a subset of the flag record; the pass adds four fields from the
fan-out context and the claim manifest:
| Added field | Source |
|-------------|--------|
| `file` | the fan-out context (one subagent per file) — also echoed by the judge output's top-level `file` |
| `line` | the claim manifest (`scripts/kb-eval/extract-judge-claims.mjs` output) |
| `claim` | the claim manifest (the verbatim claim text that was judged) |
| `disposition` | mapped from `judge_verdict` (see the vocabulary mapping below) |
## The flag record
A flag record is the union of the judge result and the four augmented fields:
```
{
"id": "<claim id>",
"judge_verdict": "not_grounded" | "source_silent", // 'grounded' is never flagged
"rule": "R1".."R8" | "",
"evidence_url": "<url the judge actually used>",
"evidence_quote":"<verbatim quote or empty>",
"reason": "<one sentence: what the source said vs the claim>",
"file": "skills/.../x.md",
"line": <number>,
"claim": "<verbatim claim text>",
"disposition": "outdated" | "wrong" | "unsourced" // R11 fix disposition
}
```
In the ledger (R6 Step 4), a flagged file is a manifest record with
`per_file_verdict: "flagged"` and one `flags[]` entry per non-grounded claim —
the structural mirror of `spor0-fix-manifest.json`'s "fix-record with
`applied:false`". A file is `flagged` if **any** judgeable claim is non-grounded
(the born-verified rule is AND: `pass` requires EVERY claim grounded).
## The two vocabularies + the mapping
There are two distinct enumerations, and conflating them is the trap this spec
exists to prevent:
1. **`judge_verdict`** — the judge's per-claim output (v3.1 prompt):
- `grounded` — the source confirms the claim → never flagged.
- `not_grounded` — the source contradicts / does not support the claim.
- `source_silent` — the source does not address the claim at all.
2. **`disposition`** — the R11 fix disposition, the correctness vocabulary from
`scripts/kb-eval/lib/base-rate.mjs` (`VERDICTS`): `correct`, `outdated`,
`wrong`, `unsourced`. Its `ERROR_VERDICTS` subset — `outdated` and `wrong`
are the only ones R11 treats as **fix targets**. `unsourced` is recorded but
is not, on its own, a content error (no MS source confirms it either way).
**Mapping** (`judge_verdict``disposition`):
| `judge_verdict` | `disposition` | R11 fix target? |
|-----------------|---------------|-----------------|
| `grounded` | `correct` | no (never flagged) |
| `not_grounded` | `outdated` or `wrong` | **yes** — the human assigns which at R11 |
| `source_silent` | `unsourced` | no (flagged, but not a fix target) |
`not_grounded → {outdated, wrong}` is deliberately a one-to-two mapping: the
judge establishes the claim is not grounded; whether it is stale-but-once-true
(`outdated`) or never-true (`wrong`) is a human call at R11, re-verified against
the live source (verification duty). The pass never auto-fixes — every flag is
human-confirmed in R11, so a judge false-positive becomes a human review, never
silent corruption of a public file.