GPT replayed B6.3, passed the profile cleanup, and produced a concrete falsifier against E-1b — plus a better answer to the E-8 question than either option the B6.3 brief offered.
366. THE ARTIFACT IDENTITY WAS NOT THE WHOLE BYTE-PRODUCING CLOSURE. cema- hashed emit, church and ADD_COMBINATOR. emit() also read a module-level enum table that it did not hash. GPT flipped one boolean in that table — "operands-then-node": false → true, profile untouched — and reproduced here exactly:
emitted bytes differ 5/9 (E-5, E-6, E-7, E-8, E-9)
target_term_sem_id move 0/9
CANONICAL_EMITTER_PROFILE_ID SAME
CANONICAL_EMITTER_ARTIFACT_ID SAME
xenc · esem · tenc · lsem · isem all SAMEadd(2,3) went from operand labels &0,&1,&2 with the add at &3 to the add at &0 with operands &1,&2,&3 — GPT's example, byte for byte. All three of E-1b's stated preconditions held across a change to its own conclusion. A theorem refuted by a change to something it claimed to bound.
One count differs and it is worth naming rather than smoothing. GPT reported 4/9; the measurement here is 5/9. The five are E-5, E-6, E-7, E-8 and E-9 — and E-8 is byte-identical to E-5, because add(const 2, const 3) and add(input x, input y) closed with {x:2,y:3} are the same closed template, which is E-3's standing finding. Four distinct terms change; five fixtures do. Both numbers are right about different things, and the fixture count is the one this battery reports.
367. THE RULE, AND WHY REPAIRING THE TABLE WOULD NOT HAVE BEEN THE FIX. GPT's formulation, taken verbatim: if changing a piece of implementation can change emitted bytes while the template and profile stay fixed, that piece belongs to the emitter artifact identity. Fixing the one table satisfies the falsifier and leaves the next helper exactly as exposed. So:
the table moved inside labelAllocPreOrder, which is bundled — the dependency is **eliminated
rather than tracked**;
EMITTER_ARTIFACT_MEMBERS names the bundle and EMITTER_ARTIFACT_INERT declares, with reasons,
what a member may reference without being bundled;
and `E-2e` derives completeness rather than trusting it: it reads lowering.mjs's own source,
derives its 66 module-level bindings, and requires every one a bundled member references to be bundled, be the profile, or be declared inert. Today: 4 members, 4 references, 0 escaping.
A hand-kept bundle is the shape `artifact_versions` was in when three of its six entries had no reader — correct the day it was written and silently short afterwards. That is the defect one artifact down, and it is why this is derived.
368. THE NEW INSTRUMENT IS PROVED NON-VACUOUS BEFORE IT IS TRUSTED. Restoring the B6.3 shape — a module-level table referenced by a bundled member — makes E-2e FAIL and name the escaping binding (LABEL_ALLOC_ORDERS←labelAllocPreOrder). An instrument that has never been seen to fail is a claim, not a measurement.
369. `E-2d` PERTURBS EVERY MEMBER, WHICH IS THE PROCEDURAL HALF OF THE SAME FINDING. B6.3 perturbed emit() alone, concluded the bundle was covered, and the member it never perturbed was the one that was missing. Each of the four members' source is now separately shown to move the id, and so is a rename — the member name is hashed beside its body, so two arrangements of the same code are two artifacts.
370. E-8's CAVEAT IS RETIRED BEFORE THE OPERATOR THAT WOULD HAVE TRIGGERED IT EXISTS. The B6.3 brief posed a choice: make E-8 explicitly add-only, or invent an equivalence for sub before sub exists. GPT's answer is neither, and it is better than both. E-8's theorem is that target_term_sem_id identifies this emitter's output rather than the term's meaning. That needs a meaning-preserving structural alternate; it does not need an algebraic law of any source operator. A beta redex supplies one generically:
T → (λz.z T)
different target_term_sem_id 9/9 (operand swap: applicable 6/9)
same decoded outcome 9/9Measured over the whole family before adopting it. It is independent of commutativity, reaches the const and input fixtures the swap could never reach, and survives sub and mul by construction. The binder is chosen against the profile rather than hard-coded — binder_names is a knob now, and a wrapper spelled z would capture under a profile that named a binder z; an adversary that silently stops being alpha-safe when a knob moves is the same species as an instrument that stops measuring while still reporting.
The operand swap is KEPT as `E-8b`, an add-specific algebraic measurement, demoted rather than deleted. When sub arrives it stops being an equivalence and becomes a falsifier for the opposite property — that operand order is semantic. Its applicability count is worth watching for exactly that reason. Still-open item 6 is closed by measurement rather than deferred again.
371. AND MY OWN NEW ASSERTION HAD THE DEFECT ITS SITE CONVENTION EXISTS TO PREVENT. The grid check for the generic alternate tested /betaEmit/, which matches the call site — so renaming the declaration left it green. §353 wrote that convention down at B6.2 ("the adversary check tests `const driftEmit = ` rather than `driftEmit`") and this assertion was written without applying it. Its own negative case reported it on the first run. Not a new species and not an argument for M-11 — the convention existed and was not followed; the fix is the site, anchored on the declaration.
372. Gate at B6.3.1. grid v1.49.0 — 96 entries / 388 citations · negative battery 333/333 · film 45/45 · lowering 23/23 · emission 15/15 over 9 fixtures · bridge 48/48 · derive 45/45 · realm 24/24 · harness 14/14 · runner 3/3. cert_id a08ee15d… byte-identical — forty-fourth consecutive round.
373. THE IDENTITY LINE IS CLOSED. `sub` IS NEXT, AND GPT HAS RULED ITS SEMANTICS. Recorded here so the next pass does not re-open the question:
The source core ALREADY HAS `sub`, defined as true subtraction over finite IEEE-754 numbers, so
2 - 3 already means -1. Compiling it as Church monus to 0 would be a miscompilation, not a design choice. Preserve source semantics; do not saturate; do not move `CORE_SEM_ID`.
The refusal is not `lower-negative` and not at lowering. sub(input x, input y) has no
underflow fact at lowering time — no phase before instantiation can decide it. The refusal belongs after the ports are bound, as a new named compiler-domain refusal (instantiate-sub-underflow, or a general instantiate-target-domain carrying its reason).
The compiler's codomain is NARROWER than the source language, and that is the correct shape for
a partial compiler — source language ⊃ currently representable target fragment. The unsupported case must be refused by name, never silently mapped to another result.
`E-9` splits accordingly and must NOT be written as source-refusal ↔ target-refusal. For
sub(2,3) the source does not refuse — it evaluates to -1; the compiler refuses. Writing the two as a correspondence would state something false about the source.
Signed integers are NOT the next round. Making sub(2,3) → -1 compile changes the target
numeric representation and reaches add, constants, inputs, decode and the proofs. First make sub honest over the representable natural domain.