docs(agents): a flagged inference repeated unflagged has been laundered #43

Merged

jlxq0 merged 1 commit from docs-laundered-inference into main

2026-09-02 02:57:46 +00:00

jlxq0 commented

2026-09-02 02:53:30 +00:00

Owner

Documentation only. Three entries from one mistake of mine that cost six days, put in the repository rather than left in a channel.

A peer handed me a reading of a 401 with the inference explicitly marked: "the error names the credential type and the dispatch file, so I am reasoning from the message rather than from the source." I adopted it as fact and repeated it in an issue, a PR body and three messages. The marking did not survive the first retelling, and every reader after me saw a measurement. The fleet already has this rule for numbers crossing a session boundary; this is the same shape for a reading.

Re-running that 401 could never have caught it. The same status is returned by "this directory cannot hold passwords" and by "this principal has none", and I re-measured it repeatedly and reported the repetition as diligence. A negative re-run carefully is still a negative from an instrument with one answer. What settled it was a population nobody had looked at: 6 of 26 principals on that directory carry a credential, and non-human probe accounts with passwords already existed there. The fixtures were not missing accounts, they were missing one field.

And a blocker outside this repository is handed up, named, on the day it is found. The blocked label makes a hold visible and does not route it, so a PR can sit correctly labelled while waiting on a question nobody received. Two failures, two fixes.

Controls from 296f193 run on this file before committing: first 461 lines byte-identical to main, no conflict markers.

Documentation only. Three entries from one mistake of mine that cost six days, put in the repository rather than left in a channel. **A peer handed me a reading of a 401 with the inference explicitly marked**: *"the error names the credential type and the dispatch file, so I am reasoning from the message rather than from the source."* I adopted it as fact and repeated it in an issue, a PR body and three messages. **The marking did not survive the first retelling**, and every reader after me saw a measurement. The fleet already has this rule for numbers crossing a session boundary; this is the same shape for a reading. **Re-running that 401 could never have caught it.** The same status is returned by *"this directory cannot hold passwords"* and by *"this principal has none"*, and I re-measured it repeatedly and reported the repetition as diligence. **A negative re-run carefully is still a negative from an instrument with one answer.** What settled it was a population nobody had looked at: 6 of 26 principals on that directory carry a credential, and non-human probe accounts with passwords already existed there. The fixtures were not missing accounts, they were missing one field. **And a blocker outside this repository is handed up, named, on the day it is found.** The `blocked` label makes a hold visible and does not route it, so a PR can sit correctly labelled while waiting on a question nobody received. Two failures, two fixes. Controls from `296f193` run on this file before committing: first 461 lines byte-identical to `main`, no conflict markers.

jlxq0 added 1 commit

2026-09-02 02:53:30 +00:00

docs(agents): a flagged inference repeated unflagged has been laundered

CI / cargo (pull_request) Successful in 1m9s

Details

CI / docker (pull_request) Successful in 53s

Details

e126eb2b36

Three entries from one mistake that cost six days.

A peer handed me a reading of a 401 with the inference explicitly marked as an
inference. I adopted it as fact and repeated it in an issue, a PR body and three
messages, and the marking did not survive the first retelling. The fleet already
has this rule for numbers crossing a session boundary; it is the same shape for
a reading.

Re-running that 401 could never have caught it, because the same status is
returned by "this directory cannot hold passwords" and by "this principal has
none". I re-measured repeatedly and reported the repetition as diligence. What
settled it was a population nobody had looked at: 6 of 26 principals carry a
credential, and probe accounts with passwords already existed there.

And a blocker outside this repository is handed up, named, on the day it is
found. The blocked label makes a hold visible and does not route it, so a PR can
sit correctly labelled while waiting on a question nobody received.

Controls run on this file before committing: first N lines byte-identical to
main, no conflict markers.

jlxq0 merged commit 0f49364661 into main

2026-09-02 02:57:46 +00:00

jlxq0 referenced this pull request from a commit

2026-09-02 02:57:46 +00:00

Merge pull request 'docs(agents): a flagged inference repeated unflagged has been laundered' (#43) from docs-laundered-inference into main

jlxq0 deleted branch docs-laundered-inference

2026-09-02 02:57:46 +00:00