The coverage tool's id arm cannot reach five of its nine prefixes - 225 entries - and the prose arm covers for it in 338 of 338 cases, so the reported false clear happens only under a flag that disables the prose arm
also called is-symbol-covered id arm blind spot · 225 entries unreachable by the id arm · --no-prose is unsafe for absence · 338 of 338 caught by prose
An agent reported that the project's safeguard against re-writing things it had already decided against was silently failing. The mechanism it described is real, but the safeguard has a second check that catches every affected case - unless you run it with a speed flag that switches that check off, which is what the agent had done.
The same thing, for a scientist
q160 reported that `scripts/is-symbol-covered.py` *"gives a FALSE CLEAR on the entire interface-entry population"* and that **BRAF, AKT1, EGFR, CD226 and IRF2 all return 'no match' while all five have a declined entry**. That tool is the guard this project relies on against re-opening reasoned refusals, and I have cited it in three separate cycles, so it was checked before anything else. 🔵 **The mechanism is real, and broader than reported.** The id arm matches `prefix + symbol` exactly. Counted live, entries reachable by that arm [E1]: | prefix | entries | reachable | |---|---:|---:| | `gene-` | 530 | **524** | | `mark-` | 22 | 21 | | `drug-` | 55 | 45 | | `noncoding-` | 9 | 3 | | `protein-` | 51 | **0** | | `target-` | 38 | **0** | | `variant-` | 2 | **0** | | `interface-` | 84 | **0** | | `localization-` | 50 | **0** | **225 entries across five prefixes are structurally invisible to the id arm**, because they are named descriptively - `interface-menin-akt1-declined-single-bret-neoppi`, not `interface-akt1`. q160 flagged `interface-` only; `protein-`, `target-`, `variant-` and `localization-` are equally unreachable. 🔴 **But the effect is null, and that is the finding.** All **338** symbol-like tokens appearing in `interface-`/`localization-` ids were run through the tool. **Zero returned 'no match'.** The prose scan catches every one, and returns the *more* useful verdict - `DECLINED ON PURPOSE` rather than `COVERED` [E2]. The five named symbols return `DECLINED ON PURPOSE` on a plain invocation. 🟠 **The reported behaviour is reproducible under one flag.** With `--no-prose`, BRAF, AKT1 and IRF2 return *"no match under N candidate names"* - **verbatim the string q160 quoted**. That flag disables the prose scan, which is the only arm that can see those 225 entries. **The agent was right about what it saw and wrong about what the tool does** [E2]. 🟢 **So the fix is not to widen the match.** Token-matching the id would add false positives to a verdict whose entire value is that `COVERED` is reliable, in exchange for catching zero additional cases. What was changed instead: the measurement is recorded as a comment at the match loop so it is not re-discovered, and **`--no-prose` now prints a warning to stderr** naming the 225 unreachable entries and stating that a no-match under that flag is not evidence of absence. Default behaviour is byte-identical [E1]. 🟠 **This is the fifth agent alarm about this corpus to be right about a mechanism and wrong about the harm**, and the second in three cycles. The pattern is now specific enough to name: **an agent finds a real structural weakness, does not measure its effect, and reports the weakness as a live defect.** The coordinator's job is the measurement, not the re-derivation.
What remains genuinely missing
🟠 The redundancy, not the coverage. For 225 entries there is exactly one arm that can see them. If a future interface- entry declines a partner without using any of the phrases find_declines looks for, nothing will catch it - the id arm cannot, and the prose arm will not. Today that set is empty; it is not guaranteed to stay empty.
🔵 The cheap guard is a regression test, not a matcher change. Re-running the 338-token sweep and asserting zero no-matches would catch the day that changes, costs one pass, and risks nothing. It is not written.
🔴 And --no-prose should probably not exist for this purpose at all. Its help text now says it is unsafe for coverage decisions, but a flag whose only effect is to make the tool confidently wrong about absence is a footgun with a speed justification - the prose scan is the slow part precisely because it is the part that works.
| reported by | q160 |
|---|---|
| verdict | mechanism real, effect null on the current corpus; the reported output is reproducible only under --no-prose |
| entries unreachable by ID arm | 225 |
| tokens tested | 338 |
| no match returned | 0 |
| five named symbols plain invocation | DECLINED ON PURPOSE (all five) |
| five named symbols no prose | no match under N candidate names (reproduces the report) |
| default behaviour changed | false |
| agent alarm instance | 5 |
reachable by ID arm
{
"gene-": [
524,
530
],
"mark-": [
21,
22
],
"drug-": [
45,
55
],
"noncoding-": [
3,
9
],
"protein-": [
0,
51
],
"target-": [
0,
38
],
"variant-": [
0,
2
],
"interface-": [
0,
84
],
"localization-": [
0,
50
]
}change made
[ "measurement recorded as a comment at the match loop", "--no-prose now warns on stderr and its help text says it is unsafe for coverage decisions", "matcher NOT widened - zero cases would be gained and COVERED would lose reliability" ]
measurements
[
{
"quantity": "entries the coverage tool's id arm cannot reach",
"value": 225,
"unit": "entries across five prefixes",
"subject": "protein- 51, target- 38, variant- 2, interface- 84, localization- 50, all named descriptively rather than prefix+symbol",
"universe": "the id arm reaches 524 of 530 gene- entries; q160 reported interface- only"
},
{
"quantity": "symbol tokens from those entries that the tool fails to resolve on a plain invocation",
"value": 0,
"unit": "of 338 tested",
"subject": "the prose scan compensates fully; the reported false clear reproduces only under --no-prose",
"universe": "338 tokens extracted from 134 interface-/localization- ids"
}
]- Write the regression test: re-run the 338-token sweep and assert zero no-matches. It would catch the day an entry declines a partner in words find_declines does not know, which is the only failure this analysis leaves open.
- Decide whether --no-prose should exist. Its only effect is to make the tool confidently wrong about absence, and the arm it disables is the one that works.
- Check whether the 225 unreachable entries mostly concern molecules that also have a gene- entry. If so the coverage gap is smaller than the count suggests, and the count in E1 overstates it.
- Reproduce q160's claim that a token match found 30 already-decided items among its 135 gaps. It was not reproduced here, and if true under a plain invocation it would contradict the 338-of-338 result.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.source-code
scripts/is-symbol-covered.py matches prefix+symbol exactly, and counted live against the graph's node ids the entries reachable by that arm are: gene- 524 of 530, mark- 21 of 22, drug- 45 of 55, noncoding- 3 of 9, and protein- 0 of 51, target- 0 of 38, variant- 0 of 2, interface- 0 of 84, localization- 0 of 50 - so 225 entries across five prefixes cannot be reached by the id arm, because those entries are named descriptively rather than as prefix plus symbol.
What this evidence cannot show
'Reachable' is computed as an id having no further hyphen after its prefix, which is the condition for prefix+symbol to match it; a handful of genuine two-token symbols would be counted unreachable by this test even though a symbol containing a hyphen could match. The counts move as entries are added and were taken at one moment during an active beat.
Evidence that argues against this (1)
- Counting reachable IDS overstates the problem if the unreachable entries are mostly about molecules that ALSO have a gene- entry, which the id arm does reach. An interface- entry declining a partner that already has gene- coverage is not a coverage gap at all, and this count does not distinguish those cases.
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.source-code
Running all 338 symbol-like hyphen-delimited tokens extracted from interface- and localization- entry ids through scripts/is-symbol-covered.py on a plain invocation returns zero 'no match' verdicts; BRAF, AKT1, EGFR, CD226 and IRF2 each return DECLINED ON PURPOSE and each has a dedicated interface- entry. Under the --no-prose flag the same symbols return 'no match under N candidate names', reproducing verbatim the output q160 reported as the tool's behaviour.
What this evidence cannot show
The 338 tokens were extracted by splitting ids on hyphens and filtering to lowercase alphanumeric strings of 3 to 8 characters, excluding a hand-written bait/stopword list - so the set includes non-symbols and may exclude genuine symbols that are longer, shorter, or that appear only inside a longer token. 'Zero no-match' is therefore a statement about that extracted set, not a proof that no interface- entry is unreachable. q160's own claim that a token match found 30 already-decided items among its 135 gaps was not independently reproduced here.
Evidence that argues against this (2)
- That the prose scan catches all 338 today is a property of how entries happen to be written, not a designed guarantee. The decline phrases find_declines looks for are a fixed list, and an entry that declines a partner in different words would be missed by both arms - so 'the effect is null' is a measurement of the current corpus, not a property of the tool.
- Declining to widen the matcher because zero cases are currently missed is the same reasoning that left the SEM collision in prose for a cycle: correct about the present, and silent about the next entry. A regression test would settle it, and is not written.
the same tool, relied on there as the guard; its id arm is narrower than assumed
cited the tool as reliable; that still holds for a plain invocation