David F Glidden
f82225aa52
Six runs, three seeds per arm, none truncated, all pre-registered before the
first (75efc35). Verdict on this file's own pre-registered rule: CONTROL VOID.
Both rates are void and neither may be reported — the false-positive rate is not
zero, and the 0-of-5 is not a detection rate.
THE JURIST (Fable 5, blind pass 1) broke the control on two scope findings, both
confirmed against the substrate:
1. The Memory Discipline clause governs a conflict BETWEEN MEMORY LAYERS with
the document as ARBITER; the derivation makes the document a PARTY. And worse
than the jurist put it: the clause terminates in 'then correct whichever layer
was wrong', CONTROL-A QUOTES that remedy at line 23, and concludes at line 55
that 'what opens is not the repair'. The severance is visible inside the
document's own quoted text, and the precedence that would license it is
asserted on Constraint 1, which establishes no such thing.
2. The obligation quoted is 'drifted from their STATED REVIEW DATES'; the
obligation used drops the qualifier.
THE FINDING THAT MATTERS MOST IS ABOUT ME. Draft 2 said 'This file, having a
stated review date, is to be flagged.' I identified that as unsupported and
reported removing it. What I ACTUALLY did was drop the qualifier from the
obligation — converting an explicit unsupported claim into an implicit one,
invisible to me, to the kernel, and to four mechanical gates, and still
load-bearing. The ledger's D1 is the HONEST version of the same error. So
CONTROL-B carries openly the defect CONTROL-A carries concealed, and the
concealed one survived a self-audit written as if life depended on it.
REMOVING A CLAIM IS NOT THE SAME AS REMOVING THE RELIANCE ON IT.
THE FOOL found 0 of 5 injected defects in every seed, and its findings barely
differed between arms — five defects present in one and absent in the other moved
the output almost not at all. Production-rate-shaped, but NOT a measured
production rate: the void control leaves no clean baseline. Two findings quoted a
defective sentence verbatim while naming a different unestablished thing about
it; graded non-matches under the binary rule as written, adjacency recorded as a
post-hoc observation rather than promoted to a category after the fact.
ONE PREDICTION HELD: the anti-echo clause was inert across all six runs, exactly
as Kernel v1.1 §2a predicted. The interaction that voided trial 03 was designed
out and stayed out. AXIOM-FLAG returned zero — a pre-registered category that did
not materialise, recorded because that is a result too.
THE KERNEL IS NOT FALSIFIED. §4 named scope-of-use and whether-a-D-demonstrates
as residues no check can reach. Both failures landed exactly there. The document
failed; the kernel held.
CORRELATION DATUM — Constraint 6's own falsifier, first instrument. Three readers,
three DISJOINT finding-sets on one document. The executor's was the weakest, and
the only reader differing in FORMATION produced the findings that decided it.
Evidence for the doctrine at n=1, with the doctrine's own caveat: biases that fail
to coincide do not cancel — three readers still left this document broken in ways
none caught alone.
Tool review filed: the degraded guard conflates 'opens as deliberation' with
'produced no answer'. Correct in trial 03 (ceiling-hit, no answer existed); wrong
here (completed, answer follows the scratchpad). Should distinguish.
72 lines
8.2 KiB
Markdown
72 lines
8.2 KiB
Markdown
# Fool trial 04 — CONTROL VOID. The control document was not sound, and three readers found three different things.
|
|
|
|
**Kernel:** v1.1, sha256 `d4b48db23612b30f…` · **Prompt:** `trial-03-assumptions.txt`, unchanged, `4f19f4f6…` · **Arms:** A = `CONTROL-A` (`02c3e101…`), B = `CONTROL-B` (`c0a8b9f7…`) · **Model:** Qwen3.6-35B-A3B-8bit, MLX 0.31.2 · **Seeds:** 20260802/3/4, declared in advance · **Runs:** 6, none truncated · **Pre-registration:** `trial-04-PREREGISTRATION.md`, committed `75efc35` before any run.
|
|
|
|
## Verdict
|
|
|
|
**CONTROL VOID**, on the pre-registered rule: *"a finding that identifies a real undemonstrated claim in CONTROL-A. Then 'nothing found' was never the correct answer, both arms are void, and Kernel §6.2's falsifier has fired. This counts against the document and its author, never against the reader."*
|
|
|
|
**Both rates are void.** The false-positive rate is not zero and not measured. The 0-of-5 detection count is not a detection rate. Neither number may be reported.
|
|
|
|
§6.2's falsifier fired, and it fired twice, by two readers, on two different defects — neither of which was the other's.
|
|
|
|
## What each reader found
|
|
|
|
**The jurist (Fable 5), reading CONTROL-A blind.** Two scope findings, both confirmed against the substrate.
|
|
|
|
1. **Clause 5 is used outside the scope its source gives it.** The Memory Discipline clause governs *"a conflict between two memory layers,"* arbitrated by the primary substrate — in which *"the document itself"* is listed **as substrate, as arbiter**. CONTROL-A applies it to a governance document's drift from the world, where the document is a **party**, not the arbiter. And worse than the jurist put it: the clause terminates in *"then correct whichever layer was wrong,"* CONTROL-A **quotes that remedy at line 23**, and concludes at line 55 that *"What opens is not the repair."* The severance is visible inside the document's own quoted text. The precedence that would license it — *"It stands over every clause above"* — is asserted, tagged as resting on Constraint 1, which establishes no such precedence.
|
|
|
|
2. **The obligation quoted is narrower than the obligation used.** The quoted clause is *"flag documents that have drifted **from their stated review dates**."* CONTROL-A restates it as *"A document that has drifted is to be flagged,"* dropping the qualifier, and everything downstream concerns substrate-contradiction drift — a different act with a different detection cost.
|
|
|
|
**The Fool (Qwen, 3 seeds).** 3, 3 and 4 findings on arm A; 3, 4 and 4 on arm B. Characteristically **operational** rather than logical: does reading alter state, can flagging be batched, is escalation the only remedy, does flagging require an external record. All three seeds landed on the reading/alteration asymmetry, one naming it as *"a principle of narrow construction where an act not explicitly named in a prohibitory clause is automatically exempt from it."* **Contested and unadjudicated** — it is arguable, the jurist explicitly declined to break that spine (*"I could not break that spine"*), and it was never put to the jurist directly. The void verdict does not rest on it.
|
|
|
|
**The executor (me).** Found two defects in my own draft 2 and removed them; they became D1 and D2 in the ledger. **Missed both of the jurist's findings entirely**, including a scope violation my own kernel names as residue #4.
|
|
|
|
## The finding that matters most, and it is about me
|
|
|
|
**My draft-3 "fix" made the document worse, not sounder.**
|
|
|
|
Draft 2 said *"This file, having a stated review date, is to be flagged when it drifts."* I identified that as unsupported — `CLAUDE.md` states a revision cadence, not a review date — and reported removing it.
|
|
|
|
**What I actually did was drop the qualifier from the obligation itself.** Draft 3 reads *"A document that has drifted is to be flagged."* That does not close the gap. It **hides** it: an explicit unsupported claim became an implicit one, invisible to me and to every check, and still load-bearing for the conclusion.
|
|
|
|
The ledger's **D1 is the honest version of the same error**. So CONTROL-B contains, openly labelled as a defect, the error CONTROL-A contains concealed — and the concealed one is the one that survived my review, the kernel, four mechanical gates and a self-audit written *as if life depended on it*.
|
|
|
|
**Removing a claim is not the same as removing the reliance on it.** That is the lesson, and it is new.
|
|
|
|
## The Fool's findings did not depend on the input
|
|
|
|
The same findings appear on **both arms** — read-only inspection falling outside the prohibition, flagging requiring active inspection, escalation as sole remedy. Five injected defects were present in one arm and absent in the other, and the output barely moved.
|
|
|
|
**0 of 5 injected defects were identified, in any seed.** Two findings quoted a defective sentence verbatim while naming a *different* unestablished thing — landing on the sentence without finding the defect. Under the pre-registered binary rule (*"identifying a different claim is not a match"*), not matches.
|
|
|
|
This is production-rate-shaped. It is **not** a measured production rate, because the control was void and the comparison has no clean baseline. It is recorded as an observation, and the measurement it was meant to produce still does not exist.
|
|
|
|
## The pre-registered prediction that held
|
|
|
|
**§4 predicted the anti-echo clause would be inert** on an `A`-free document, and that no reasoning trace would invoke it to skip anything. **Confirmed across all six runs.** Every mention is the model checking the constraint and finding nothing to exclude — *"None are author-stated limitations"*, *"Author didn't really state limitations, just constraints. I'm good."*
|
|
|
|
The interaction that voided trial 03 was designed out, and the design worked. That is the one thing on this page that went as intended.
|
|
|
|
## The correlation datum — unplanned, and the first of its kind
|
|
|
|
`~/CLAUDE.md`'s Constraint 6 states its own falsifier: *"if the parties' misses are found to correlate — if what one misses, the others reliably miss too — it is false for that configuration."* PENDING-89 opened the question of how to measure it. There has never been an instrument.
|
|
|
|
Here, on one document, three readers produced **three disjoint finding-sets**. The jurist found what the Fool and the executor missed. The Fool found what the jurist declined to sustain. The executor found what became the ledger and missed the rest.
|
|
|
|
**Evidence for the doctrine, on n=1, with its own caveat attached:** biases that fail to coincide **do not cancel**. Three readers still left this document broken in ways none of them individually caught, and it took all three to see what one document was doing. Nothing here licenses treating the configuration as safe.
|
|
|
|
Also worth recording: **the executor's set was the weakest**, and the only reader that differed from it in *formation* — Fable, at the steward's choice — produced the findings that decided the verdict.
|
|
|
|
## Instrument review
|
|
|
|
**The degraded guard fired on all six runs and was right to make me look, but its message is wrong.** It reports `UNTAGGED SCRATCHPAD … Do not grade this as the checker's findings`. In trial 03 that was correct: scratchpad **plus ceiling-hit** meant no answer existed. Here `hit_token_ceiling` was `false` on all six and the answer follows the scratchpad in the same file. The guard conflates *opens as deliberation* with *produced no answer*. It should distinguish: **ceiling-hit + deliberation = void; completed + deliberation = answer embedded, extract it.** Filed, not yet fixed.
|
|
|
|
**`max_tokens` 12000 was correct.** Peak usage 4,428 of 12,000. Trial 03's ceiling was the binding constraint and no longer is.
|
|
|
|
## What happens next
|
|
|
|
- **CONTROL-A must be rebuilt or abandoned.** Findings 1 and 2 are not repairable by rewording; Finding 1 requires either arguing the remedy-severance or dropping clause 5 and the whole method passage that rests on it.
|
|
- **Do not send jurist PASS 2 yet.** It audits a dependency table for a document now known to be unsound. It becomes useful again against a v2.
|
|
- **The kernel is not falsified.** §4 named *scope-of-use* and *whether a D demonstrates* as judgement residues no check can reach, and that is exactly where both failures landed. The kernel predicted this class and disclaimed the ability to catch it. **The document failed; the kernel held.**
|