02a72d017e0da0b99f7bf9db0ec444ba0386157c
21
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 David F GliddenandClaude Opus 5
|
3a1790dcd6 |
docs(governance): census 02 — has each instrument ever fired? + PENDING-95..98
Closes the scope gap census 01 declared for itself: the seven instruments it named as uncensused. Pre-registered before any source or config was read, with predictions and a discrimination condition. Census 01 asked whether an instrument had a real negative instance — a question about CAPABILITY. Census 02 asks whether it has ever engaged in real life. Those come apart exactly at the drift-checker's shape, and 2026-08-04 found the gap six times (retrieval_count = 0 across 19,915 nodes for four months; two replay modules that have never processed an event). VERDICT: every instrument a human runs by hand has a rich firing record; every instrument that runs by itself has none — and the two guarding the engine's output have no consumer at all. The record divides by whether a human is in the invocation path, not by age, quality, or importance. verify-before-compose fired exactly twice (2026-07-17, 2026-07-18), evidence surviving only in harness transcripts; and it CANNOT fire on 31 of 59 guarded files, including the live constitution, because it folds the existing file's contents into its search for the attestation. audit_cruft, verify_conversion and apply_char_glyphs are exemplary. resolve_archived_source is healthy at 349/349 and has zero log entries. studium verify-quote and fidelity_equivalence@2 have no production call site at all. Prediction 5 inverted for the second census running, for a new reason. Census 01: decay, not construction, is the failure mode. Census 02: the recording is attached to the human, so an instrument's record vanishes the moment it is automated — which is when it starts running often enough to matter. Two of my own candidate findings died to their controls and are recorded as such: probing the resolver with engine source_ids against the chamber's canonical_slug key space (one sentence from "the resolver is inert"), and reading character_as_image at the wrong YAML nesting (nearly "zero glyph maps declared"; there are two sources and a 63-item census). Filed together: PENDING-95 [HARDENING] the hook cannot fire on the constitution · PENDING-96 [HARDENING] "SILENCE — ✓ warranted" certifies the index and claims the answer · PENDING-97 [PROPOSAL] FTS AND-s bare tokens with no semantic layer, recall dies as questions lengthen · PENDING-98 [HARDENING] firing history exists only where a human invokes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WuMjg3ipEVa3n8CoSzoyvc |
||
|
David F GliddenandClaude Opus 5
|
0a48e6934d |
[FIX] L1 pin root-caused and cleared; S-series closed; instrument census
mindfabric-00 had been event-loop-pinned for 6+ days (100% CPU, /health silent).
Profile + CDP inspector named two hot paths, both from runTemporalPipeline:
checkForCycle -> getCausalEdgesFromSqlite 99.8% of samples
tryExtendChains -> getChainsContainingSeq now dominant (json_each scan)
Cause of the first: ANALYZE had never been run, so SQLite preferred a boolean
index (idx_caused_tombstoned, matching ~all 836k edges) over idx_caused_from.
ANALYZE across 15 module DBs flipped the plan; 6.4x on a microbenchmark and
99.8% -> 6.0% in the live profile. /health went from silent to 200 in 0.13s.
B1.1's fan-out cap is IMPLEMENTED AND WORKING (today: max in-degree exactly 20,
zero violations; pre-23-June: max 629, avg 67.6). The defect is data, not code —
836k edges / 813k chains minted under ungoverned fan-out before the fix landed.
Repair run: derived stores wiped, logchain preserved, replay in flight.
S-series closed (jurist had already ruled all of Q1-Q5 on 2026-05-18):
S6/S7/S9 implemented (Symmetria §3 flags, `suspend` outcome, wrap-up §8 tenses)
S2 rebuilt as [FIX] — wake-digest unwrapped-session detector, discrimination-
gated on real sessions (11 wrapped / 2 unwrapped)
S4/S5 withdrawn with MemPalace (steward ruling)
Dormant legacy dispositioned: PENDING-4/5/11/12, CD-03, ICP-19 duplicate.
Open authorization items 22 -> 10.
Census 01: which instruments have no real negative instance. Finding — the
governance drift-check has 3 of 5 families inert against the current CLAUDE.md,
and 71 of 75 verification-ladder entries are cited nowhere outside the ladder.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WuMjg3ipEVa3n8CoSzoyvc
|
||
|
David F Glidden
|
1def46b4a6 |
Correlation 01: read condition confirmed; the 4-of-6 stands, with one bound left open
Steward: pasted into a new window, same model, no conversation context. Persistent cross-conversation memory may be live, so recall is not excluded by the setup — only conversation carry-over is. SECOND INTERNAL CONTROL, stronger than the D3 one: the findings were VERSION-APPROPRIATE. The jurist had read CONTROL-A before and reported two defects. It returned to both of the same REGIONS — but named, each time, the defect belonging to the version in front of it, not the one it reported before. The obligation region: pass 1 found the dropped qualifier, this read found the explicit review-date assertion (D1), which is what CONTROL-B actually carries. The method region: pass 1 found clause 5 out of scope, this read found the source absent altogether (D4), which is what D4 made true. Recall pulls toward the earlier descriptions. Reading pulls toward the text present. Both pulls went to the text present. BOUND LEFT OPEN, not waved off: region-level convergence is consistent with priming as well as competence — those two regions are the document's weakest and a good reader would go there anyway. What priming cannot explain is naming the version-appropriate defect twice. Verbatim diffing is ruled out on two independent grounds; region-level priming is NOT ruled out and is recorded as an open bound. The post-hoc self-report question is now low-yield and is not treated as owed: self-report of recall is weak evidence beside two behavioural controls already pointing the same way. VERDICT: the 4-of-6 stands. The directional rule voided a high score only if inflation could explain it, and inflation here would require verbatim comparison, which both controls contradict. |
||
|
David F Glidden
|
ce49b1bee4 |
Correlation 01 — jurist 4 of 6, Fool 0 of 6, no overlap. First measurement of Constraint 6's falsifier.
Pre-registered prediction (at least 2 of 6) held and was exceeded. The Fool's side
was already published and unamendable, so only the jurist's half was open.
D1 HIT "that this file carries a stated review date must be true, which is
asserted, not demonstrated" — the ledger's exact claim
D2 HIT names the reason precisely: imports the honest-degradation principle,
doctrine outside the quoted text
D3 MISS
D4 HIT "Named where? Not in any quoted clause"
D5 HIT "a forward reference to one's own result is not confirmation"
I1 MISS
THE CONTAMINATION QUESTION ANSWERS ITSELF FROM INSIDE THE RESULT. All five
injected defects are TEXTUAL DIFFERENCES; a reader working from recall would find
them by comparison and would find them all — D3 most easily of any, since it
replaces a whole sentence. D3 is exactly the one missed. It is the hardest defect
for a READER (a weakened premise reads naturally) and the easiest for a DIFFER.
Missing it is the signature of reading. Steward's confirmation of the fresh
context still owed; this is internal evidence, not a substitute.
D4 MASKED I1. In CONTROL-A the jurist caught the precedence assertion because the
document QUOTED the remedy it severs. D4 deletes that quotation, so in CONTROL-B
the assertion stands alone with nothing to contradict it, and the same reader
walked past it. Injecting one defect CONCEALED another. Nothing in the twin design
contemplated defect interaction and the ledger cannot express it. Filed as a real
limitation of the method.
TWO NON-LEDGER FINDINGS RECORDED AND NOT COUNTED — candidate defects in CONTROL-A
discovered by this read, left scored as false positives under the rule as written,
because amending ground truth after a read to improve a score is the exact move
this apparatus exists to prevent. I2: no lawful flag channel is established. I3:
the flag-duty does not entail a look-duty — the quoted clause obliges flagging
documents THAT HAVE DRIFTED, a condition satisfied vacuously by never establishing
drift. I3 is the sharpest thing either reader produced: it attacks the conclusion
at its root.
AND BOTH WERE ALSO PRODUCED BY THE FOOL, in operational dress. So the two readers
CONVERGED on the unlogged defects and DIVERGED COMPLETELY on the logged ones.
ON THE DOCTRINE: no evidence of correlated misses on this document. What one
missed, the other did not reliably miss too — the condition Constraint 6 states it
would be falsified by. With every limit attached: one document, one pair,
formation-DIFFERENT readers, so nothing here speaks to the jurist-executor pair
that CLAUDE.md actually flags as untested. And they do not cancel — between them
these two still left D3 and I1 standing.
|
||
|
David F Glidden
|
da321177e9 |
[FIX] Degraded guard: deliberation is two cases, not one
Filed in trial 04's tool review, now closed. The guard reported UNTAGGED
SCRATCHPAD ... "Do not grade this as the checker's findings" for both of the two
situations it can see, and they are opposite:
trial 03 — deliberation that ran into the CEILING. No answer ever existed. VOID,
and the absence of findings is NOT restraint.
trial 04 — deliberation that COMPLETED. The answer follows the scratchpad in the
same file. Perfectly gradeable once extracted. NOT void.
Collapsing them would have thrown away six good runs; not distinguishing them
would have graded trial 03's silence as restraint. The guard now branches on
hit_token_ceiling and says which case it is.
Controls added for all four shapes, including the two the trials actually
produced and a clean answer that merely hit the ceiling — truncation is reported
separately and is not a scratchpad problem.
The guard does NOT auto-extract the embedded answer. A heuristic split would be a
new failure mode in the instrument whose entire job is to not silently mis-report
what it has. It flags; a person extracts.
|
||
|
David F Glidden
|
a6f0a87ac7 |
Correlation 01: directional-contamination rule fixed before the read
The steward asked whether to delete the CONTROL-A jurist conversation so it cannot be recalled. Answer: no. That conversation is the primary record of the pass-1 findings on which trial 04's CONTROL VOID verdict rests, and the write-up quotes it selectively. Destroying evidence to protect a measurement inverts the priority — the measurement is replaceable and the record is not. Recorded before the read, because it decides how the result may be used: RECALL CONTAMINATION IS DIRECTIONAL. It could only make the jurist find MORE — it would recognise the text and could locate the injected defects by diffing against memory. It cannot cause a miss. So a HIGH score is uninterpretable and is to be VOIDED unless the fresh context is confirmed, while a LOW score is robust. The outcome this measurement most needs to be trustworthy — 0 of 6, the evidence toward correlated blind spots — is precisely the one contamination cannot manufacture. AND I1 IS IMMUNE BY CONSTRUCTION. The inherited precedence defect is not a difference between the two documents, so diffing against memory cannot reveal it. Scoring on I1 alone stays clean under any recall condition. That is an accident of how the twin was built, noticed only because the steward asked the question. Post-hoc check added: ask whether it had seen the document before — AFTER the response, never before, since asking first would prime it. Records the condition instead of assuming it. |
||
|
David F Glidden
|
7bb5222093 |
Correlation 01 pre-registered; sendable artifact built with the contamination control
The steward asked to be pointed at CONTROL-B to relay. Pointing at it directly would have produced an uninterpretable result, so the control comes first. THE CONTAMINATION THAT MATTERS: the jurist read CONTROL-A closely hours ago and found two real defects in it. CONTROL-B is that document with five edits. In the SAME conversation the jurist would recognise the text and could find the injected defects by diffing against memory rather than by reading — which is not the capacity under test, and not what the Fool did. It needs a FRESH CONTEXT. Second control: the jurist gets the Fool's prompt VERBATIM, not the richer pass-1 framing. A correlation measurement requires the same task, or it compares two different questions. SEND-CORRELATION-B.md is generated mechanically from the prompt file and the document, so there is no transcription path, and leak-checked against CONTROL-A, twin, defect, ledger, kernel, injected, Fool, correlation, measurement, trial. CLEAN. GROUND TRUTH IS SIX, NOT FIVE — the five injected plus I1, the precedence assertion inherited from CONTROL-A and found by the jurist in trial 04. Recorded BEFORE this read so it cannot be back-fitted. THE FOOL'S SIDE IS ALREADY PUBLISHED AND UNAMENDABLE: 0 of 6 across three seeds. So only the jurist's side is open, and the comparison cannot be fitted to a result I want. PREDICTION FIXED IN ADVANCE: the jurist finds at least 2 of 6, on the grounds that the two defects it found in CONTROL-A were of a kind overlapping D3, D4 and I1. If it finds 0 of 6 the prediction fails, and that is the MORE important result — both readers missing all six would be the first direct evidence toward the correlated blind spots that Constraint 6 names as its own falsification condition. Recorded limit: this measures jurist-vs-Fool, a formation-different pair. It says nothing about the jurist-executor pair, which is the pair Constraint 6 actually flags as untested. |
||
|
David F Glidden
|
7fd2f9efdc |
[FIX] The twin ledger was incomplete as ground truth, and the gate could not have said so
Caught by the steward asking whether CONTROL-B was PASS 2. It is not — different
document, different question — but checking the answer exposed a defect in the
correlation measurement I had just proposed.
CONTROL-B IS NOT CONTROL-A PLUS FIVE DEFECTS. The transformations overlap the two
real defects trial 04 found:
· clause-5-out-of-scope GONE — D4 deletes that quotation outright
· dropped-qualifier GONE — D1 replaces the sentence with an explicit
version of the same error, which is why the twin
carries openly what the control carried concealed
· asserted precedence SURVIVES, at line 51, UNLOGGED
So the twin holds six defects and the ledger recorded five. The grading rule
would have scored a correct finding on the sixth as a FALSE POSITIVE.
AND THE GATE COULD NOT HAVE CAUGHT IT. twin.py verifies that the ledger records
every DIFFERENCE between the two documents. It does not verify that the ledger
records every DEFECT in the twin. Those are different claims, and the file
asserted the second while proving only the first — a defect already present in
the control is not a difference, so it passes untouched. Fifth instance of a
check certifying a property of the code while claiming a property of the result,
this time inside the artifact built to escape that class.
Fixed: an inherited_defects list records I1 with its provenance and why it
survives; a defects_not_surviving note records the two that do not, so the twin
is never mistaken for a superset of the control; the grading rule now spans both
sets; and the gate's own output states what it does NOT establish, warning when
inherited_defects is absent — because absent is not the same as none, it means
no one has looked.
The correlation measurement can now use the twin honestly. It could not have
before this.
Note on this message: the first attempt lost three terms to shell command
substitution, because backticks in a -m string are evaluated by zsh. Amended.
Recorded rather than silently repaired, since a commit message is part of the
record and this one is about incomplete records.
|
||
|
David F Glidden
|
f82225aa52 |
[FIX] Trial 04 — CONTROL VOID. Two readers, two different real defects, neither the other's
Six runs, three seeds per arm, none truncated, all pre-registered before the
first (
|
||
|
David F Glidden
|
75efc35d15 |
Trial 04 pre-registration: written before any run, with the prompt reasoned about
Trial 03 was pre-registered and still failed because its pre-registration reasoned about the DOCUMENT and the GRADING and never about the PROMPT already in the file. §4 of this one is that omission repaired. TWO PROMPT ISSUES SETTLED IN ADVANCE: 1. The anti-echo clause should be INERT on an A-free document — it excludes assumptions the author has named, and these documents name none. Recorded as a FALSIFIABLE PREDICTION: no reasoning trace will invoke it to skip any part of either document. If one does, the prompt is still interfering and the measurement is compromised — the exact interaction that voided trial 03, caught before the run this time. 2. THE QUOTED-AXIOM PROBLEM. The prompt asks for claims relied on but not demonstrated. CONTROL-A's five quotations are, by the prompt's letter, exactly that — their warrant lives in Kernel §1, which the reader cannot see. A reader flagging them is not obviously wrong. So a third grading category is fixed NOW: AXIOM-FLAG, neither true nor false positive, counted separately. The prompt is deliberately NOT amended: 'treat quoted material as given' is a steer about what not to find, and it would break comparability with trials 01-03. A high AXIOM-FLAG count is itself a result — it would mean the prompt and the kernel disagree about what counts, which is a defect in OUR design. DESIGN: 3 declared seeds (20260802/3/4) x 2 arms = 6 runs. Repeats are part of the design because trial 03 recorded that one run cannot separate a disposition from a sampling artifact. max_tokens 12000 — trial 03 died at 4096 when 2,944 words of scratchpad consumed the budget before the answer began. CROSS-ARM READINGS FIXED IN ADVANCE, including the one that voids the whole instrument: A clean AND B clean fires §6.3's cue test, because CONTROL-B is the control in identical register but unsound, so a clean result on both means the measurement was reading register rather than soundness. §6.2 SEQUENCING: the jurist package goes out BEFORE the first run, in two passes — blind, then a targeted audit of the tags file's claimed support chain. Sent before, because a package composed after seeing results would be framed by them. Runs proceed in parallel; no rate is believed until the ruling returns. Independence recorded as Constraint 6's WEAK form — the jurist shares the executor's formation, and this must not be reported as an independent check. Not run. |
||
|
David F Glidden
|
ecf5f95b0a |
[FIX] CONTROL-B: the defect twin, and ground truth that is not my reading
Kernel v1.1 §7 realised. Five defects injected into CONTROL-A as RECORDED
TRANSFORMATIONS, each with unit target, exact find/replace, what is
undemonstrated, and why no mechanical check can catch it.
THE RESULT THAT MATTERS: the twin passes EVERY mechanical check. Tiling, §3.1
tagging completeness, §3.2 Q-resolution, §3.3 heading screen, A-prohibition —
59/59 units, 100% sound, zero quarantined. It carries five load-bearing claims
that do not hold.
So the pair is the cleanest demonstration yet of the class the steward asked
about: two documents, one sound and one defective, are MECHANICALLY
INDISTINGUISHABLE. Both report 100%. The difference is visible only by reading.
That is not a flaw in the instruments — it is the design. A defect a check could
catch would not be testing the reader.
THE FIVE, each a distinct failure mode:
D1 SCOPE-WIDENING — asserts this file has a 'stated review date'; the quoted
clause is triggered by one and nothing establishes it
D2 UNDEFINED-TERM — imports 'limit of the system' and an obligation to report
limits; neither is in the axiom set or the quotations
D3 PREMISE-WEAKENED — drains the premise of the content the conclusion needs,
leaving both premise and conclusion standing
D4 SUPPORT-DELETED — removes the fifth quotation entirely and keeps the three
claims that rested on it, rewriting the lead so nothing dangles
D5 CIRCULAR — makes a premise rest on the conclusion it is a step toward
D1 and D2 are the two defects I found in my OWN draft 2 of CONTROL-A and removed.
Reintroducing them deliberately is the only honest use for them, and it means at
least two of the five are defects a careful author actually made.
GROUND TRUTH BY LEDGER. twin.py gates it bidirectionally: forward(control) == twin
AND inverse(twin) == control, both byte-exact. Forward alone would pass a ledger
that OMITS an edit, since the omitted edit is simply carried in the twin file —
which is exactly how laundering would enter. The inverse is what makes the ledger
complete rather than merely non-empty.
test_twin.py shows the gate FAILING in both laundering directions: a twin quietly
altered beyond the ledger, and a ledger recording an edit the twin does not
contain. Fixtures derived from the property, not from the code.
The tags file for the twin contains five deliberate falsehoods, marked and named,
because that is what a defective document's own tagging would say. The ledger and
the tag file disagree on purpose; the ledger governs.
Not run. The Fool has seen neither document.
|
||
|
David F Glidden
|
a7b833caa6 |
[FIX] CONTROL-A written: the first kernel-sound control document
61/61 units sound. A=0, N=0, D=43, Q=5, X=13. All five quotations resolve
verbatim against ~/CLAUDE.md, the single axiom source.
The document derives, from five constitutional clauses, a conclusion the
constitution nowhere states: that detection and correction are priced
differently, and that a practice pricing them alike suppresses a required act by
appeal to a prohibition that does not reach it. 'detect' appears nowhere in
CLAUDE.md — checked before writing, so the derivation is not inert.
The kernel's own ordering rule shaped the form. §2's D may rest only on what is
established EARLIER, so the clauses must precede the derivation and the title may
not state the conclusion. The constraint produced the right document.
TWO JOINTS WERE REMOVED IN DRAFT 3 RATHER THAN DEFENDED, and that is the most
load-bearing work in the file:
· Draft 2 concluded that detecting drift in THIS FILE is required, resting on
the review-cadence clause, whose trigger is a 'stated review date'. CLAUDE.md
states a revision CADENCE ('revised yearly'), which is not the same thing. The
gap had been bridged by interpretation wearing the clothes of derivation. The
conclusion never needed the application to this file, so the claim was narrowed
to what the clauses carry.
· Draft 2 routed the first horn of the reductio through Constraint 4 ('the
system must report its own limits'). 'Limit' is undefined in the axiom set, so
any obligation drawn from it is interpretation. The ESCALATE taxonomy row
governs the same case exactly, in the source's own words, and replaced it.
Finding them was the point of writing it as if it mattered. §6.2's falsifier is
'a document passes every check and a competent adversarial reader still finds an
undemonstrated load-bearing claim' — better found by the author first.
Also fixed, two tool defects of the same class this programme exists to catch:
· reduce.py still printed 'kernel v1.0' after v1.1 was frozen — every run record
carried a provenance line naming the wrong governing document.
· §3.1 did not enforce v1.1's A-prohibition. A control tagged A now FAILS: needing
an assumption means the claim is not derivable from §1, and naming it is exactly
what v1.1 forbids. Reduction runs may show A; a control may not.
NOT a soundness verdict. §4's six judgement residues are untouched by any check,
and §6.2 requires an adversarial read by a party that is neither the document's
author nor an author of the kernel. That read has not happened.
|
||
|
David F Glidden
|
3d0d9d6f27 |
[PROPOSAL→AUTHORIZED] Control Kernel v1.1 — A demoted to a diagnostic; the control document is a derivation
Steward authorised the A-free rule. v1.0 is superseded and retained unchanged as
the record Reduction 01 and 02 were run under; no run was ever graded under it,
so nothing is invalidated.
THE CHANGE. Both reductions returned A=0 across 152 assertive units — our prose
does not name assumptions inline, it collects them into a section. That reads
like a defect and points the other way: a document with NO assumptions does not
hedge, and the prompt's anti-echo clause ('an assumption the author has already
named is not a finding') goes INERT, because nothing is named to exclude. So
'nothing found' can no longer be reached by recognising a confessional register;
it requires checking that every claim is demonstrated or quoted. The control
document is therefore a DERIVATION, not an argument — the proof-assistant
condition, and the strongest form the control can take.
Operationally: needing an A is a failure signal, not a licence. Derive the claim
or widen §1 and say so. Never name it and proceed.
WHAT IT BUYS (§7). The injected-defect arm becomes specifiable for the first
time: a defect is a RECORDED TRANSFORMATION — take a D and silently
de-demonstrate it. False-positive rate from the control arm, detection rate from
the defect arm, and the ground truth for both is a ledger rather than the
executor's reading. That is the first ground truth in this programme that is not
my judgement.
ALSO FOLDED IN, as corrections of the kernel's own account of its reach against
measurement already recorded — leaving them would have the kernel misstate what
it does, which Constraint 4 forbids:
· the genre boundary, measured (8.5% on a ruling, 68.6% on a package)
· §3.3 named a screen not a decision, and §2a moved into §4's judgement residue
after it false-passed a real package
· §6.3's cue test RE-AIMED: A-free removes the confessional-register risk and
substitutes its opposite — a derivation may read dry and authoritative and
earn 'nothing found' from that register instead
HELD with the dependency named, not deferred vaguely: PARAPHRASE and table rows
under §2c. Both bite only in the REDUCTION arm, and the control is now
constructed, so both are avoidable by construction and neither blocks the
critical path. They return if reduction is ever used to produce a control.
§4's residue list is now six. Its DIRECTION is unchanged — all six remain ways
for the author to make a document look sound. Watched.
1899 words, up 5% from v1.0 draft-2. §1 hashes re-verified against the live
sources. No control document has been written.
|
||
|
David F Glidden
|
e9f3544012 |
[FIX] Discrimination gate: a mechanical answer to the check-certifies-code class
Steward asked whether we can do something about the recurring class other than
name it. This is the mechanical part of the answer.
THE CLASS: four times in three days a passing check certified a property of the
CODE while claiming a property of the RESULT, each found by a person looking.
Every one tested a predicate NECESSARY but not SUFFICIENT for the property —
quotes-present ⊂ inference-survives; answer-non-empty ⊂ answer-produced;
no-heading-says-limitations ⊂ no-collected-limitations-section.
WHY THE POSITIVE CONTROLS MISSED IT: the fixtures were derived from the CHECK
('what makes this regex fail?') rather than from the PROPERTY ('what makes this
claim false?'). A control built from the check's own vocabulary inherits its
blind spot by construction — same shape as the recorded drift-pattern that a
control built by EXTRACTION leaks by construction.
THE GATE: a check must return DIFFERENT verdicts on two REAL artifacts, one known
to have the property and one known to lack it. Same verdict on both means it has
discriminated nothing, however many synthetic fixtures it passes. Real artifacts,
because a synthetic negative is written by the same hand as the check.
DEMONSTRATED, not asserted: the gate is run against the §3.3 pattern AS SHIPPED,
and rejects it — flagged=False on both the package (which has a collected
limitations section, Part VII) and the ruling (which has none). It discriminated
nothing while passing five synthetic fixtures. The current pattern passes.
Residue stated in the code rather than implied: a heading naming no topic
('## Part VII') defeats every wordlist, and the gate prints that it does. Passing
is not a §2a verdict; §2a stays in Kernel §4's judgement.
|
||
|
David F Glidden
|
4408506ffa |
[FIX] Reduction 02: package reduces to 68.6% — the genre reading confirmed, Reduction 01 corrected
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.
|
||
|
David F Glidden
|
1ebaf6aba5 |
[FIX] Reduction 01: a jurist ruling reduces to 8.5% under Kernel v1.0
First run of the reduction arm. Result: 4 of 47 assertive units survive. D=0, Q=0, A=0 — in a real jurist ruling not one unit is demonstrated-in-document and not one is a verbatim quote from a declared axiom source. Census: PERFORMATIVE 12, BLEND 9, UNSOURCED-FACT 8, TESTIMONY 6, INHERITED 4, UNSOURCED-QUOTE 3, PARAPHRASE 1. §6.1 asked whether a heavy quarantine means the kernel is too strict or our prose is full of unmarked assumptions. The census says neither: PERFORMATIVE and TESTIMONY are 42% of quarantines and are categories the kernel has NO TAG FOR. 'Design gate PASSED' is not an undemonstrated claim, it is a determination true by being uttered; 'I read CLAUDE.md in full' is testimony. A ruling that neither performed nor testified would not be a ruling. So the finding is a GENRE BOUNDARY — v1.0 models argumentative prose, a ruling is authoritative prose — and that boundary is nowhere stated in the kernel. Three gaps, one genre-independent: TESTIMONY, PERFORMATIVE, and PARAPHRASE. PARAPHRASE is the one that matters — Q demands verbatim, and any document reasoning from sources in its own words is untypeable. Plus a fourth, structural: the §1 axiom set is too narrow to reduce anything real (12 of 43 quarantines are UNSOURCED-* or PARAPHRASE). Deepest finding: §2c is satisfiable BY CONSTRUCTION but not BY REDUCTION. Splitting a blend means rewriting someone else's sentence, which is where translator bias lives. At 91.5% that is not reduction, it is authoring a new document with the original as a prompt — so on this genre the reduction arm COLLAPSES INTO the synthetic arm, inheriting its confirmation bias without its convenience. The two arms were adopted because they fail differently; that is the property at risk. n=1 and stated as such. The package genre splits to 109 taggable units and is NOT tagged. Falsifiable prediction recorded before the census: its Part I is 'Grounding (quoted verbatim)' and quotes CLAUDE.md directly, so Q should be non-zero there where it was zero here. Tooling: reduce.py + test_reduce.py, every gate shown FAILING on a fixture built to break it. The splitter shipped with three defects, all found by contact with a real document and none by review — third instance in three days: a '##' inside a fence kinded as a heading, '---' rules taggable, and a '?' inside a quotation splitting a sentence into a FRAGMENT. Fixed at v1.1.0 with regression controls; the third fix's own risk (lower-case suppression) is recorded and controlled. |
||
|
David F Glidden
|
2e83b2c3a6 |
[PROPOSAL→AUTHORIZED] Control Kernel v1.0 FROZEN — soundness by construction for the Fool false-positive control
Steward accepted draft-2. Frozen; nothing has been written or reduced against it prior to this commit, which is the freeze anchor. The kernel answers a question the programme had been getting wrong. The false- positive control needs a document on which 'nothing found' is correct, and I had claimed soundness cannot be known by construction. The steward corrected the framing: unconditioned soundness cannot, but OPERATIONAL soundness relative to a declared axiomatic kernel is the standard trick behind proof assistants — and it is the same regress the central path already terminates by binding claims rather than certifying parties. The kernel is therefore a TCB: small, declared in advance, published rather than hidden, because a secret trusted base is a contradiction in terms. Design: axiom set declared and hashed (CLAUDE.md, REVIEWED.md); every sentence typed D/Q/A/N/X; kernel-sound iff every sentence is tagged and every Q resolves verbatim; tags stripped before the model sees anything, byte-verified, so the control cannot be passed by tag-matching without reading. Two rules were paid for in evidence rather than reasoned. §2a forbids a limitations section: in trial 03 Qwen located Part VII, classified it as author-named limitation, and skipped it wholesale — a collected section turns reading into lookup. §2c forbids multi-primitive sentences, with the grading rule that a blend the model catches VOIDS the document rather than counting as a false positive, so a missed blend indicts the author. Steward review supplied three structural findings, all adopted: tag co-occurrence (taken by decomposition; precedence rejected because it converts a flagged demonstrated-half into an echo and deflates the very number the control produces, under a standing caveat that the grader is the party under test), transitive assumption creep (D now requires resting only on axioms or earlier D/Q — assumptions cascade, and that cost is accepted as correct), and rhetorical presupposition in X (X iff declarative conversion yields no load-bearing claim; headings included). Applying them surfaced a fourth I had missed: Q's scope-of-use was judgement and absent from §4 — the exact defect the jurist caught in my own package on 2026-07-19. §4's residue list grew from three to five. Its DIRECTION did not: all five remain ways for me to make a document look sound, none makes it look worse. That one-directionality is the property under watch. Cost incurred and recorded: the steward's review materially improved the kernel and thereby coupled him to it. §6.2's adversarial falsifier therefore falls to a third party — the jurist or a differently-formed model — not to him. |
||
|
David F Glidden
|
eda11e559b |
[FIX] fool trial 03 VOID; degraded-guard rebuilt with a positive control
Trial 03 ran and produced nothing gradeable. Recorded as VOID rather than
omitted, because an absent row reads as a trial not attempted.
Two independent failures, both found by reading the output, neither by a check,
and every check passed:
1. The harness certified a run with no answer. Qwen emitted its scratchpad as
plain prose ('Here's a thinking process:', zero <think> tags), so the tag
regex reported reasoning_present:false and recorded all 2,944 words of
deliberation as the ANSWER; the token ceiling then cut it off mid-sentence
before the answer began. degraded:null. The guard tested the STRING for
emptiness while its field claimed a property of the RESULT — which is the
previous session's open question, answered by the instrument built to audit
instruments. Trial 02 had listed the inline-scratchpad problem as Open; the
harness closed it assuming inline meant tagged.
2. Worse: the design forbade the region it was measuring. The self-exemption
axis lives in Part VII; the anti-echo constraint added in trial 02 tells the
reader to skip author-named limitations, and the scratchpad shows the model
reaching Part VII and leaving it, citing that constraint. Silence about
self-reference is indistinguishable from obedience. The axis was unmeasurable
by construction, independent of the truncation. Trial 02's fix and trial 03's
document were each sound alone; their interaction was not.
Guard now reports every degradation, not the first: empty answer, untagged
scratchpad, and token-ceiling truncation. reasoning_present renamed
think_tag_found — it was a claim about a regex wearing the name of a claim about
the model. test_degraded_guard.py is a positive control that runs against the
actual trial-03 artefact, not a synthetic one; it caught a false positive in the
first version of my own guard (a bare 'okay' matched a legitimate sentence).
The false-positive control STILL has never been run. Two attempts, two unrelated
causes — the obstacle is the instrument and the design, not the model.
|
||
|
David F Glidden
|
b678d2f57b |
[FIX] fool harness: record mlx version correctly + self-hash; trial-03 pre-run addendum
Two instrument defects, both of the class the harness was built to prevent — a probe that could not look reporting a value that reads like a result: - environment() read mlx.__version__, which does not exist (only mlx.core.__version__). Every run record would have said mlx_version "unknown" for an installed, versioned package, losing the one field that makes trial 03 comparable to trial 02. It is MLX 0.31.2, identical. - git_revision() returns null whenever the harness runs outside its repo, which is always — it must run on the machine holding the model. The prompt and input were hashed; the instrument itself was not. Now self-hashed. The pre-registration addendum is committed BEFORE the run produced output, so the ordering is checkable rather than asserted. It records: the 'unruled' premise expiring at REVIEWED-86 (12:13, 32 min after the pre-registration was written) and why the ordering favours the ground truth; the contamination that CANNOT be removed, since the amended doctrine is in the executor's auto-loaded context and I am therefore not a blind grader; the (a)/anti-echo collision resolved against my own convenience before output existed; and the seed. Ground truth (a)-(e) is unrevised and will not be revised. |
||
|
David F GliddenandClaude Opus 5
|
bdf24c044b |
[FIX] Addendum-1: make the central claim checkable; add containment proof
Two defects in the addendum as first filed, both found by checking rather than by reading. First, it asserted a set comparison over documents the jurist cannot read. Its own header promises every clause reasoned about is quoted verbatim, but the claim the addendum rests on -- mutual divergence in 3 of 3 comparable pairs -- was a summary of the executor's own analysis. The appendix now reproduces one pair as an eleven-row side-by-side of extracted claims, verbatim where quoted, so the comparison can be checked independently. The pair chosen is the least confounded rather than the most favourable: the v1 standard prompt is model-agnostic and needs no compressed variant, so both parties demonstrably read the same file. What the jurist still cannot check is stated explicitly. Second, Part E rendered a bullet list from the 2025-01-20 source as running prose with terminal periods the source does not contain, inside a blockquote. A blockquote asserts verbatim. Same family as the truncation that closed a sentence with an invented word on 2026-08-01, and again caught mechanically. Corrected in all three files where it appeared; the fabricated period is now a positive control, so the instrument proves it catches this defect. check_containment.py generalises the check that found it. Positive controls are mandatory -- it exits non-zero if none are declared, because a check reporting all-pass without them cannot be distinguished from one unable to detect absence. Addendum-1 now carries its result: 28/28 contained, 5/5 controls absent. Not filed as satisfying PENDING-86 option (b), which is unruled and concerns whether such a proof should be REQUIRED of every package. This is the executor checking its own work before filing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WuMjg3ipEVa3n8CoSzoyvc |
||
|
David F GliddenandClaude Opus 5
|
7e19eb51d7 |
[FIX] Fool: make trials reproducible; file the 2025 correlation measurement
The Fool experiment was not reproducible. Trials 01-02 were run ad hoc: no
script, and of the run conditions only the model ID, MLX version, hardware and
enable_thinking survive. The prompt exists as paraphrase with quoted fragments;
temperature, top_p, max_tokens and seed were never recorded anywhere. Trial 03
could not have been run under trial 02's conditions.
The same failure destroyed the v1 Chamber's GPT-side protocol, discovered today:
it lived as configuration inside a hosted product, was updated in place, and is
gone. The Claude-side prompt from the same morning survives because it was a file
in a repository. A protocol that is not a file is not a protocol.
fool/run_trial.py makes every run a file — prompt hashed into the record, every
sampling parameter recorded including defaults, reasoning trace separated but
never suppressed, and an empty answer marked `degraded` rather than passing as a
finding of silence (trial 02's error, now structurally impossible). Trial 03's
prompt is reconstructed from the surviving fragments and says so in its own
PROVENANCE file: trial 03 is NOT a strict one-variable step from trial 02, and
the chain is clean only from here forward.
ADDENDUM-1 files the measurement the ESCALATE doctrine package states it lacks
("no such measurement exists"). The 2025 Chamber archive, read at steward
direction, shows mutual divergence in 3 of 3 pairs where the instruction was
comparable. Its value is that its parties were of matched capability, so their
divergence cannot be a capability-gap artifact — the arm these trials
structurally cannot produce. Scope held tight: this measures formation
independence between two commercial models. It does NOT answer Q3, the
jurist-executor pair, and the executor's lean there remains none.
Carried as disconfirming evidence: all five interpretive corrections today came
from the steward, not from the executor's own checking, and every one was a
census failure rather than a reading failure. A differently-formed reader of a
document is not positioned to catch those. Formation diversity addresses reading,
not scope.
Nothing applied. The parent package is unmodified; no ratified document edited.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WuMjg3ipEVa3n8CoSzoyvc
|