Skip to content

AK8: the fingerprint's flags asked again of clef - #214

Merged
shinyobjectz merged 1 commit into
mainfrom
ask-double
Nov 21, 2026
Merged

shinyobjectz merged 1 commit into
mainfrom
ask-double

Conversation

@shinyobjectz

@shinyobjectz shinyobjectz commented Nov 21, 2026 •

Copy link
Copy Markdown
Owner

Stacked on #212 (base ask-calibrate). Merge #211 and #212 first.

This is AK8 in docs/ask.feature: the fingerprint's flags are asked again of clef before anyone acts on them.

raea fingerprint score FILE --ask flags|unsure (or ask: "flags" in fingerprint.ts)

  • Each flagged moment is asked as a yes/no question in its axis's own first words. A quality axis is flagged on "no". The question looks at the picture or listens to the sound, as the axis does, and is asked on that moment's window only, so the cost depends on the number of flags, never on the length of the piece.
  • unsure also asks each run of windows that an axis puts in its plausible band without flagging.
  • Moments on the same window share a call. Everything goes through raea-app-ask's runner, so the routes, the cache, the record and the person's calibration all apply. Each question carries the axis's words as its label words, so the person's notes calibrate it like any other question (AK6, AK7).
  • Each flag gets a line under the score, written the way flags are: 0:00-0:09 dark: clef doesn't see it (0.00). --json adds a clef array.
  • Each answer is written to the ledger as an eye:clef evidence row, replacing clef's last row for that moment. Such rows train the axis's head but never count as the person's (FP6d, train.rs).
  • A moment it couldn't ask (no route, the pool spent, an unreadable file) says why, and the score still prints.

Where it lives

  • A new crate, raea-app-double (apps/double, 165 lines), because raea-app-ask is at 1,841 of its 2,000 lines.
  • raea-app-fingerprint gains about 15 lines in verbs.rs (1,975 of 2,000).
  • Fingerprint gains an ask field, in taxonomy.rs and sdk/raea/src/fingerprint.ts.
  • The manual page for raea fingerprint describes --ask for score next to the existing --ask N of readings and postmortem.

Checked

  • raea-app-double: 3 tests, with no spending:
    • flags are asked, and unsure adds plausible windows, run together;
    • an axis asks as it senses, and a quality axis is flagged on "no";
    • each flag is asked on its own window, said beside it and kept as clef's row; an unreadable file asks nothing.
  • raea-sys-fingerprint and raea-app-fingerprint tests pass (8 and 11).
  • A real run on a 10 s render with five flags, on the person's own Cloudflare keys, cost $0.0007; a second run cost $0 (cache):
    • Clef saw no dark on a frame that is plainly well lit, where the fingerprint had a gating flag. That is the false positive this check is for.
    • Clef agreed the scene looks computer generated (0.96).
    • The ledger was restored afterwards.
  • The same run showed the fingerprint judging sound axes on a render with no sound track. That fix is separate, in fingerprint: a render with no sound track has no sound to judge #213.

The app side (rule_check scoring with --ask unsure and asking clef about the rules nothing measures) is shinyobjectz/raea-feed#101.

raea fingerprint score FILE --ask flags (or ask: "flags" in fingerprint.ts)
puts each flagged moment to clef as a yes/no question in its axis's own
words, looking or listening as the axis does, on that moment's window
alone, so it costs by the flags and never by the piece; --ask unsure also
asks the windows an axis puts in its plausible band without flagging them.
It goes through raea ask's runner (routes, cache, record, the person's
calibration), and the axis's words are the question's label words, so the
person's notes calibrate it (AK6, AK7).

Each flag gets a line under the score ("dark: clef doesn't see it (0.00)",
and a clef array with --json), and each answer replaces clef's last
eye:clef evidence row for that moment, which trains the head and never
counts as the person's. A moment it couldn't ask says why, and the score
goes on.

The double check is its own crate, raea-app-double, since raea-app-ask
holds 1841 of its 2000 lines. The manual page covers score --ask beside
readings' and postmortem's --ask N; docs/ask.feature marks AK8 done.
@shinyobjectz
shinyobjectz changed the base branch from ask-calibrate to main November 21, 2026 00:39
@shinyobjectz
shinyobjectz merged commit 6073b11 into main Nov 21, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant