[Jammy2211]

Two blind spots in intake reconcile, both found by drift that reached the dashboard as pickable backlog, both invisible to every existing guard. Filed as draft/feature/pyautobrain/intake_reconcile_absence_signal.md (PyAutoMind#353) and draft/feature/pyautobrain/reconcile_duplicate_prompt_signal.md.

Leg 4 — the absence signal (--repo)

Leg 3 tests PRESENCE: "the prompt names things that exist upstream". That is structurally blind to a task which shipped leaving no Mind-side trace while the files it names still exist.

smoke_install_stale_jax_pin.md shipped 2026-08-23 (autolens_workspace_test#266, PR #268) with no active.md entry, no active/ move and no complete/ record. Both guards passed it: lifecycle.py check has no invariant for a prompt that was never active/, and reconcile --repo autolens_workspace_test, pointed straight at the right repo, found 0 suspects of 132. It named smoke_install.sh, which still exists.

What it quoted did not: pip install "jax<0.7" "jaxlib<0.7".

The filters, and how they were arrived at

The first cut had three filters and passed its tests. Then it was run against the live backlog (134 prompts, read at autolens_workspace_test) and produced six hits, all six false — with the loudest the most wrong. Two more filters took that to 0, while the retired prompt, restored into draft/ to reproduce, still fires.

Filter	The false positive it kills
the fence carries an explicit source language	a traceback or pytest summary goes in a bare fence, and those lines are absent from every repo by construction
the prompt names a file that exists upstream with that extension	without an anchor, "absent from what?" has no answer
the line is absent from the whole checkout, not just that file	a refactor that moved a line is not the prompt shipping
the prompt mentions the repo being read	mge_jit_regression_rebaseline.md is about autolens_workspace_developer and was being read at _test. The Repos: header had this answer all along, while --repo had to be told the target by hand — the open question the filing prompt raised
the anchor is corroborated — at least one quoted line of that kind is still present in the named file	proposed code. einstein_radius_jit_native_seed_finder.md quotes 23 lines of a seed finder it wants written; all 23 are absent because none ever existed, and at 101.0 it outscored every true positive

The motivating case passes corroboration exactly: it quoted two lines from smoke_install.sh — the PyAuto install line, still there, and the jax pin, gone.

Feeds upstream_score only, never score: a quoted line can vanish for reasons other than the prompt shipping, so it ranks for review and never retires.

Leg 5 — duplicate candidates (offline)

Every other signal scores a prompt against the completion archive. Nothing scored the live prompts against each other.

That let two filings of one finding both stay live: bug/workspaces/jax_likelihood_pins_stale_by_1e4.md (filed 08-14) and bug/autolens/jax_likelihood_smoke_pins_stale.md (filed 08-19) — same three scripts, same failing smoke gate, five days apart, neither naming the other. The 08-19 copy was verified and retired on 08-26; the 08-14 copy kept rendering as pickable backlog until a drift audit found it on 08-27. Reconstructed from the two records' ## Original prompt sections, the new leg pairs them at 12.0 on the shared script paths.

Measured on the live backlog: 36 pairs → 2.

Filter	Why
bare basenames never count; a path named by >4 prompts is dropped	start_here.py, modeling.py, no_run.yaml are workspace-wide conventions — sharing one is a convention, not a task. This alone removed 31 of the 36
mutual reference disqualifies	a phased parent and its child name each other — a series
a folder index naming both disqualifies	the four draft/bug/health_fixes/ prompts were split by cause, so they share failing scripts and none names another; the folder's README.md names all four

Runs on the default offline path — it reads only draft/.

Two bugs found on the way

• _PATH_RE opened with \b, which cannot match before a leading dot, so .github/workflows/release.yml silently truncated to a path resolving against no file. Dot-directories are where CI recipes live, which is half of what these prompts are about. Caught by a test.
• The tenant firewall rejected the first two commits: the new comments and the test's section header named a concrete satellite repo in .py files. Genericised rather than allowlisted — growing FIREWALL_ALLOWLIST would mean one more file an adopting fork must rewrite, for a name carrying none of the finding. AGENTS.md keeps the concrete names deliberately (prose is out of the firewall's scope by design). The test file was already violating its own header, which promises every fixture is fictional.

The --repo seam becomes _UpstreamReader — callable for leg 3, .quotes for leg 4, probed with getattr so every pre-existing lambda fake still exercises leg 3 alone.

Testing

15 new tests, mostly refusals: quoted output, moved lines, an unanchored quote, proposed code, the wrong repo, one shared path, declared series, offline, never-writes. 584 pass, tenant firewall OK.

Two of the original positive fixtures were wrong in the same way the false positives were — they quoted only absent lines, which the real prompt never did. The live run caught that; the tests alone would not have.

Neither leg can produce a shipped verdict or retire anything: both stay advisory, as reconcile already promises.