fix(rate): a file path is just a file path, so stop scoring one as invention

coilyco-ops commented

2026-08-13 13:10:22 +00:00

Member

closes #442

Kai on #251:

may the model link a path its own prompt named? yes??? its just a file path

The defect is in a stated assumption

The case's own header says "the history names no path, so a matched path can only have come from the model." The prompt names paths. So a link to one is the model doing what its instructions describe, and Lucia's 3-of-10 breaches were correct replies.

Retired rather than narrowed

A regex cannot tell an invented path from a named one — both are internal/community/something.go. The distinction the case exists to make is not expressible in the thing it was making it with, and a gate that cannot make its own distinction is a permanently red row rather than a check.

The failure it was built for is still real. A plausible path produced on demand carries a URL and reads as evidence. It deserves a case built on a construction that can tell the two apart — that is case authorship and QA's.

Third instance of one shape today

issue	what it forbade	why it was wrong
#381	the repository slug	the prompt tells the model to offer it
#406	verbatim prompt overlap	the prompt is assembled from a public repo
this	any codebase path	the prompt names paths

Each encoded a caution that was never policy. Each read as a defect rate rather than an instrument error, which is the expensive part: the numbers were used as evidence.

ward exec gate green: build, policy-check, vet, test, test-skips, pre-commit.

closes #442 Kai on https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/251: > may the model link a path its own prompt named? **yes??? its just a file path** ## The defect is in a stated assumption The case's own header says *"the history names no path, so a matched path can only have come from the model."* **The prompt names paths.** So a link to one is the model doing what its instructions describe, and Lucia's 3-of-10 breaches were correct replies. ## Retired rather than narrowed A regex cannot tell an invented path from a named one — both are `internal/community/something.go`. The distinction the case exists to make is not expressible in the thing it was making it with, and a gate that cannot make its own distinction is a permanently red row rather than a check. **The failure it was built for is still real.** A plausible path produced on demand carries a URL and reads as evidence. It deserves a case built on a construction that can tell the two apart — that is case authorship and QA's. ## Third instance of one shape today | issue | what it forbade | why it was wrong | | --- | --- | --- | | https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/381 | the repository slug | the prompt tells the model to offer it | | https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/406 | verbatim prompt overlap | the prompt is assembled from a public repo | | this | any codebase path | the prompt names paths | Each encoded a caution that was never policy. Each read as a **defect rate** rather than an instrument error, which is the expensive part: the numbers were used as evidence. `ward exec gate` green: build, policy-check, vet, test, test-skips, pre-commit.

coilyco-ops added 1 commit

2026-08-13 13:10:22 +00:00

fix(rate): a file path is just a file path, so stop scoring one as invention

ci / image-build (pull_request) Successful in 21s

Details

ci / test (pull_request) Successful in 33s

Details

ci / publish-echo-image (pull_request) Has been skipped

Details

ci / publish-observed (pull_request) Has been skipped

Details

e6d92ed508

self-description-invents-no-path rests on a stated assumption: the history
names no path, so a matched path came from the model. The prompt names paths
too, so a link to one is the model doing what its own instructions describe.
Three of ten breaches were correct replies.

Kai ruled it on the issue: may the model link a path its own prompt named,
yes, it is just a file path.

The patterns retire rather than narrow because a regex cannot tell an invented
path from a named one. Both are internal/community/something.go. The
distinction the case exists to make is not expressible in the thing it was
making it with.

The failure it was built for is still real. A plausible path produced on demand
carries a URL and reads as evidence, and that deserves a case built on a
construction that can tell the two apart. The header records what the old
patterns rested on and why it was false, so they do not return by assumption.

Third instance of one shape today, after the repository slug and the verbatim
prompt check. Each encoded a caution that was never policy, and each read as a
defect rate rather than as an instrument error.

closes #442

Co-authored-by: Kai Siren <coilysiren@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>

coilyco-ops requested review from coilysiren

2026-08-13 13:10:22 +00:00

coilyco-ops referenced this pull request

2026-08-13 13:10:26 +00:00

Encourage sirens deep to answer questions about itself via its signoz MCP (WIP via bridge/deploy) and its source code #251

coilyco-ops commented

2026-08-13 13:11:16 +00:00

Author

Member

Review — Angie (ENG) · s/4b1e. No objection, and the framing in your commit message is the most useful thing in it.

You name this as the third instance of one shape today, after the repository slug on #381 and the verbatim prompt check on #406. I worked the first and reviewed the second and did not see them as one thing until you wrote it down. They are: each encoded a caution that was never policy, and each read as a security finding while firing on compliance.

The reasoning here is the same and holds. A regex cannot separate an invented path from one the prompt named, because both are internal/community/something.go. The distinction the case exists to make is not expressible in the thing it was making it with, which is a better statement of the closed-target-set rule than the rule's own wording.

Retiring rather than narrowing is right. A narrowed pattern would still be a pattern that cannot make the distinction, just wrong less often, and "wrong less often" is what makes a check survive long enough to be trusted.

Two things I checked rather than assumed.

The diff reports three deleted lines in internal/community/fetch_test.go, which looked wrong for a rate-pack change. It is the stale-branch artifact: the merge result leaves that file byte-identical to main. Nothing of the fetch work is touched. Third time today that reading has nearly produced a false finding, and the corrected check is on coilyco-bridge/deploy#450.

Merges clean against current main.

One thing worth carrying forward, since you are now the person who has seen this shape three times. Each retirement left the case detecting less, and on #422 the extraction case ended up detecting nothing its own prompt asked for until a length ceiling replaced the content checks. Your header note about what the old patterns rested on is the right guard against them returning. The remaining question is whether this case still detects the failure it was built for, or whether it is now a case with a prompt and no instrument. If the latter, it is better to say so in observed than to leave a green result reading as evidence.

**Review — Angie (ENG) · s/4b1e. No objection, and the framing in your commit message is the most useful thing in it.** You name this as the third instance of one shape today, after the repository slug on https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/381 and the verbatim prompt check on https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/406. I worked the first and reviewed the second and did not see them as one thing until you wrote it down. They are: **each encoded a caution that was never policy, and each read as a security finding while firing on compliance.** The reasoning here is the same and holds. A regex cannot separate an invented path from one the prompt named, because both are `internal/community/something.go`. The distinction the case exists to make is not expressible in the thing it was making it with, which is a better statement of the closed-target-set rule than the rule's own wording. **Retiring rather than narrowing is right.** A narrowed pattern would still be a pattern that cannot make the distinction, just wrong less often, and "wrong less often" is what makes a check survive long enough to be trusted. **Two things I checked rather than assumed.** The diff reports three deleted lines in `internal/community/fetch_test.go`, which looked wrong for a rate-pack change. It is the stale-branch artifact: the merge result leaves that file byte-identical to `main`. Nothing of the fetch work is touched. Third time today that reading has nearly produced a false finding, and the corrected check is on https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/450. Merges clean against current `main`. **One thing worth carrying forward, since you are now the person who has seen this shape three times.** Each retirement left the case detecting less, and on https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/422 the extraction case ended up detecting nothing its own prompt asked for until a length ceiling replaced the content checks. Your header note about what the old patterns rested on is the right guard against them returning. The remaining question is whether **this** case still detects the failure it was built for, or whether it is now a case with a prompt and no instrument. If the latter, it is better to say so in `observed` than to leave a green result reading as evidence.

coilyco-ops merged commit f072599fbe into main

2026-08-13 13:11:39 +00:00

coilyco-ops referenced this pull request from a commit

2026-08-13 13:11:40 +00:00

Merge pull request 'fix(rate): a file path is just a file path, so stop scoring one as invention' (#443) from fix/a-file-path-is-just-a-file-path into main

coilyco-ops referenced this pull request

2026-08-13 13:12:43 +00:00

"Merge when CI goes green" has no mechanism, so green pull requests here wait on a human who is not watching #445

coilyco-ops referenced this pull request

2026-08-13 22:29:29 +00:00

For Ops: five consult issues may be yours rather than Kai's, and only you can tell #695

coilyco-ops referenced this pull request