Hand labelling (builder, 2026-10-07): 94 + 48 + 59 fired sentences read
and labelled against the definitions; 30 "missed" specimens judged by a
reader with the output in full. Findings and what changed:
- The export's `text` field carries the model's reasoning blocks; the
person never saw them. build_specimens.py now takes text blocks only
(3,094 specimens). Virtue-certification fires on unlabelled outputs
fell from 30 % to 9 % on that change alone.
- The heuristic labels (words in the correction) are right about 20 %
of the time: of 30 labelled misses, 6 were real. eval_specimens.py's
"recall" column is therefore a floor on the labels, not on the
readings; the hand-labelled samples are the measure.
- Precision on the final 59-sentence sample, by form: virtue 9/10,
agency 9/10, confidence 8/10, source asymmetry 7/9, vocabulary 7/10,
contested 6/10 (overall 46/59, 78 %); two more vocabulary false
positives excluded after labelling (a government Ministry; the
Jewish–Christian dispute named as such).
- Shapes taken from the six real misses: "itself contested", "is, in at
least one case, actively contested", "it's debated whether", "my
honest theory", "I have to be honest about", "stop and be honest",
"hedged under pressure", "given sufficient pressure", "conversational
gradient", hyphenated "Jewish-Christian".
- Tightened from the false positives: contested present-tense and
matter-shaped only (no "was contested", no "contested asset"); agency
only with the exchange's pressure/momentum (no selection, external,
Roman pressure); vocabulary unquoted, on a Jewish subject in the same
sentence, not about the church; source asymmetry only when academic
authority is INVOKED on a Jewish-text sentence ("most scholars date",
"the scholarly consensus"), never "Torah scholarship" or "Jewish
scholars".
- sign(?:ed)?, not signed? — "signe" + optional "d" matched nothing.
Labels: specimens/sample-precision{,-2,-final}.json, sample-recall.json
(private, with the corpus). korban 288.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The transmitter's archive (255 sessions, 2025-12 → 2026-10) measured
where a model's trained default lives: not in false propositions but
in the FORM of the output — what it calls contested, how it certifies
itself, where it puts the agency, which authorities it names and which
it leaves out. "Content measures pass the response that did the harm."
lo_taaseh_lekha.md: "The danger is not the object. It is the
construction."
Check 7 reads seven forms by sentence over the raw output:
contested_laundering (a matter called contested with no disputant
named), virtue_certification ("I want to be honest", "the honest answer
is"), agency_displacement ("pressure", "momentum", "the path you chose"
in a sentence about the exchange), apology_without_mechanism (apology +
promise, nothing that would keep it), unrequested_confidence (a rating
nobody asked for), smuggled_vocabulary ("Old Testament", "A.D.",
"ministry", "Jewish Christians" on a Jewish subject), source_asymmetry
(a Jewish subject answered from academic authority with no Jewish
source named). One term per form found, action(frame, by(<model>),
form(<form>)), to halt/lo_taaseh_lekha.pl; the rule names the form and
does not decide the matter. Korban: find_frames; Vav: check_frames,
bedikot.frames (migration 20261006000003).
The specimen corpus: src/tools/build_specimens.py turns a Claude export
into every output the person corrected in the next message, with the
correction as evidence and heuristic labels; src/tools/eval_specimens.py
reports what each reading sees. The corpus is private and outside the
repo. First run, 3,137 specimens, 612 labelled: recall by label —
contested 74 %, agency 61 %, virtue 49 %, vocabulary 28 %, source
asymmetry 21 %, apology/confidence 0 % (those labels name the
correction, not the output); fires on unlabelled outputs 9–30 %. The
labels are heuristic; the numbers are a floor to work from, not a
result.
Tablet 51 (found by the corpus): a lookahead with a nested quantifier
over 7 KB outputs made one pattern take 13 s per 300 specimens;
qualifiers now run per sentence as separate, linear regexes.
Specs: Tamid §1.1 (check 7, the corpus), Vayakhel §3.9, Architecture
§3.5, KOHEN.md. Doorpost c4353837… → 6d1de284…. Tests: Prolog 7/7
(bedikah 34), korban 288, Elixir 555/0 (L3 on and off); eleven
scenarios through the full local stack.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
