David F Glidden
4408506ffa
Prediction recorded in Reduction 01 BEFORE this census, so it could fail: the
package's Part I is 'Grounding (quoted verbatim)' and quotes CLAUDE.md directly,
so Q should be non-zero where it was zero. Q=9. D=40, where the ruling had none.
ruling package
sound 8.5% 68.6%
PERFORMATIVE 12 0 <- the genre signature
BLEND 9 25
INHERITED 4 0
UNSOURCED-QUOTE 3 0 <- §1's header clause worked
Genre reading confirmed eightfold: a package proposes, a ruling determines.
CORRECTS Reduction 01's strong conclusion that 'the reduction arm collapses into
the synthetic arm'. On package prose repair touches 31.4%, not 91.5% — reduction,
not authoring, and the two arms stay distinct. That conclusion was correctly
bounded at n=1; the bound was the whole of its content and one document collapsed
it. Reduction 01 now carries the correction inline.
BLEND is now the blocker and is genre-independent: 25 of 33 quarantines, 7 of
them rows of the Part IV table, which pairs a quote with an end-state and a
verdict — three primitives by construction.
Two check findings, one good and one bad:
§3.2 CAUGHT A REAL TAGGING ERROR OF MINE. Unit 145 was tagged Q; it is a sentence
ABOUT a quotation, not a quotation, so not verbatim-as-a-unit. Corrected to D.
The check found it, the reading did not — the 'quoted but not traced' defect the
jurist caught on 2026-07-19, mechanised.
§3.3 GAVE A FALSE PASS, found by looking. Part VII 'Disconfirming evidence' IS a
collected limitations section under §2a — the exact section trial 03 showed the
model skipping wholesale — and the screen missed it because it never says
'limitations'. Widened; the package now correctly FAILS §3.3. But no pattern can
decide this: a section titled only 'Part VII' defeats any wordlist, and a control
now asserts that. §3.3 is a SCREEN, not a decision; §2a belongs in §4's judgement
residue. Fourth time in three days a passing check certified the code while the
property failed, and the fourth found by a person looking.
A=0 IN BOTH DOCUMENTS, and it is the same fact as the §2a failure seen from the
other side: we do not name assumptions inline, we collect them into a section.
Our best governance prose is written in exactly the shape that defeats the reader
the section was written for.
Also fixed: the tool was still printing 'NOT checked here: §3.2' after §3.2 was
implemented — under-claiming, but still a false statement about what ran.
Kernel v1.1 candidates are now evidence-backed and remain UNAPPLIED; v1.0 stays
frozen and a revision is a new experiment. The false-positive control remains
unrun and neither reduction produced a usable control document.
6.1 KiB
Reduction 02 — a jurist package against Control Kernel v1.0, and what it corrects in Reduction 01
Kernel: v1.0 FROZEN, sha256 67c9b870491db744… · Splitter: v1.2.0 · Document: differently-biased-checkers-JURIST-PACKAGE-2026-08-01.md, sha256 f5e6ff20b2a76500…, 2,774 words · Artefacts: *.units.jsonl, *.tags.tsv
The prediction held
Recorded in Reduction 01 before this census, so it could fail: "its Part I is headed 'Grounding (quoted verbatim)' and quotes ~/CLAUDE.md directly, so Q should be non-zero there where it was zero here."
Q = 9. And D = 40, where the ruling had none.
| Ruling (01) | Package (02) | |
|---|---|---|
| assertive units | 47 | 105 |
| sound remainder | 4 (8.5%) | 72 (68.6%) |
| quarantined | 43 (91.5%) | 33 (31.4%) |
D / Q / A / N / X |
0 / 0 / 0 / 1 / 3 | 40 / 9 / 0 / 2 / 21 |
| PERFORMATIVE | 12 | 0 |
| BLEND | 9 | 25 |
| UNSOURCED-FACT | 8 | 5 |
| TESTIMONY | 6 | 2 |
| INHERITED | 4 | 0 |
| UNSOURCED-QUOTE | 3 | 0 |
| PARAPHRASE | 1 | 1 |
The genre reading is confirmed, eightfold. PERFORMATIVE 12 → 0 is the signature: a package proposes, a ruling determines. INHERITED 4 → 0 because D is reachable once anything is demonstrated. UNSOURCED-QUOTE 3 → 0 because §1's header clause did its work — the package's GROUNDED-IN comment names its sources, so they entered the axiom set exactly as the kernel provides.
Correcting Reduction 01
Reduction 01 concluded: "on this genre the reduction arm collapses into the synthetic arm — inheriting the synthetic arm's confirmation bias without its convenience."
That is overturned for the genre that matters. It was scoped with an explicit n=1 caveat, and the caveat was load-bearing: on package prose, repair touches 31.4% of units, not 91.5%. That is reduction, not authoring, and the two arms remain distinct. The strong reading survives only for authoritative prose, where it was measured.
The lesson is not that the conclusion was wrong — it was correctly bounded — but that the bound was the whole of its content, and a single further document collapsed it.
What now blocks reduction: BLEND, and it is genre-independent
25 of 33 package quarantines (76%) are BLEND — a sentence carrying more than one primitive, which §2c requires be split. 7 of those 25 are rows of the Part IV consequence-trace table: a row pairing a quoted clause with an end-state and a verdict is three primitives by construction. Tables are structurally blended.
This is the cost §2c imposes on reduction and not on generation, now measured: to reduce this package you would rewrite about a third of it, and a quarter of that third is a table that arguably should not be prose at all.
Two findings the checks produced, one good and one bad
§3.2 caught a real tagging error of mine. I tagged unit 145 — "The taxonomy's [ESCALATE] row reads 'Surface immediately; do not proceed.'" — as Q. It is not a quotation; it is a sentence about one, with the words inline. Not verbatim-from-source as a unit, so Q is wrong; D holds because it rests exclusively on a §1 axiom. The check found this, the reading did not — and it is precisely the quoted-but-not-traced defect the jurist caught in my work on 2026-07-19, now mechanised.
§3.3 gave a false pass, and I found it by looking. The package's Part VII — Disconfirming evidence is a collected limitations section in §2a's sense, and trial 03 showed the model skipping exactly that section wholesale, by name. The screen missed it because the heading never says "limitations". Widened, and the package now correctly fails §3.3.
But the deeper point is recorded in the code: no pattern can decide this. A section titled only "Part VII" defeats any wordlist, and a control asserts that it does. §3.3 is a screen over obvious namings, not a decision on §2a — so §2a compliance belongs in Kernel §4's judgement residue, where it currently is not. That is the fourth time in three days that a passing check certified the code while the property failed, and the fourth time a person looking found it.
A = 0 in both documents, and it is the same fact seen twice
Not one unit of either document is "assumed, named at the point of use". Two readings, and the evidence picks one: we do not name assumptions inline — we collect them into a section. The package does it in Part VII; that is why A is empty and why §2a fails, and the two are one phenomenon.
Which is uncomfortable, because §2a is the rule trial 03 paid for: a collected section is what the model located and skipped. Our best governance prose is written in exactly the shape that defeats the reader we built the section for.
Kernel v1.1 candidates — now evidence-backed, still unapplied
- State the genre boundary. v1.0 models argumentative prose; on authoritative prose it measures mismatch and reports 91.5%.
- Resolve
PARAPHRASE. One instance in each document — low, but structural:Qdemands verbatim and prose restates. - Move §2a compliance into §4. §3.3 is a screen; the code now says so and the kernel does not.
- Decide
BLEND's treatment for tables — exempt structurally, or forbid tables in a control document. A's reachability. If no real document ever tagsA, either the definition is unreachable or §2a is asking prose to change shape. Both are worth saying out loud.
§1's header clause is validated and needs no change — it widened the axiom set correctly and drove UNSOURCED-QUOTE to zero.
Nothing above is applied. Kernel v1.0 remains frozen, and a revision is a new experiment.
Standing disclosure
This package was written by the executor, who is also its reducer. Reducing one's own prose, one knows what one meant and is disposed to tag charitably — the translator bias in its strongest form, and unmitigated here. Reduction 01's document was not mine, which is why it was chosen first. The two censuses differ in genre and in authorship, and this census cannot separate those.
The false-positive control remains unrun, and neither reduction produced a usable control document.