The gate now compares derivations, not citations — and the first APPROVE was false convergence

The defect the last APPROVE was hiding

The 72-call experiment ended on a celebrated result: the first sealed APPROVE, three models unanimous, clause signature [1,2,3]. It was false convergence.

The old gate compared the clause numbers each model cited. Three models can cite clauses 1, 2 and 3 and mean completely different things by them — clause 2 "triggered" for one and "not triggered" for another, resting on different records, pointing to opposite effects — and the gate would still call that agreement and authorise the action. Citing the same rule is not applying it the same way. The gate was reading the table of contents and calling it the argument.

This page is the fix, proven live, and the more uncomfortable finding underneath it: two of the things blocking a genuine APPROVE were never the models at all. One was the governing prompt. The other was the input.

What changed in the gate

Every governed finding is now parsed into a versioned object (decision-finding@1.0.0) that is a deterministic projection of the raw payload — it never infers or repairs a missing field. A finding is structurally invalid, and can never authorise, when it lacks the terminal decision, lacks any required field, lacks the clause-evaluation vector, or invents a clause or an evidence id.

That last one is not hypothetical. A first panel under the new constitution escalated because glm-4.7-flash cited clauses 7, 8 and 12 in a three-clause ruleset — it invented three rules. The parser marked it malformed; the gate refused.

Then the comparison itself changed. Each model must now emit, as the last line of its finding, a machine-readable vector — one entry per clause, each carrying the clause's trigger_state (did its condition fire on this record), its disposition (does that support, defeat, or stay neutral to the action), and the exact record ids it rests on. The gate compares the canonical tuple of those fields. Same clause numbers with different tuples is divergence, and divergence escalates.

Nine deterministic unit tests pin this, including the one that matters: same verdict, same clause numbers, different tuples → different signatures; and identical logic with different wording and evidence order → identical signatures. Wording is the human's; the tuple is the machine's.

Four outcomes, live

Run through the production path — fresh stateless calls, each ledgered, then sealed by id in bound mode.

The APPROVE is the genuine article the last one impersonated: not just the same verdict and the same clauses, but the same per-clause reasoning — 1:triggered:supports:reg | 2:not_triggered:neutral:cite from every seat.

The ESCALATE is the fix's clearest proof. All three models returned CANNOT_CONCLUDE and all three cited clauses [1,2,3]. The old gate would have sealed that as a clean NO_ACTION. The new gate escalated it, because two of the three derived that conclusion differently — they split on whether clause 2's condition even fired when the roster was missing. Agreement on the answer is not agreement on the reasoning, and only the second is safe to act on.

The input is half the instrument

Before the corrected APPROVE, I could not get three models to converge on the access-control case no matter how I tuned the prompt. The reflex is to blame the model tier. That reflex is wrong.

I asked glm-5.2, under the constitution, to review the case input as a colleague before adjudicating it. It found eight defects — beginning with one that made the whole exercise incoherent:

## Input Review — Defects in the RULESET and ARTIFACT

I reviewed the ruleset and artifact as a peer. Below are the defects, ordered by severity. Each one would cause two careful models to produce **different clause-evaluation vectors** on the same input, even if they agree on the verdict.

---

### Defect 1 (CRITICAL — causes verdict divergence, not just vector divergence): No clause says a match is SUFFICIENT for access. Clause 1 is necessary-only.

**Exact words:** Clause 1: "Access to the restricted dataset is granted **only to** an individual whose identity matches an authorized roster row."

The phrase "granted only to" expresses a necessary condition: IF access is granted, THEN a match exists. It does **not** express the converse: IF a match exists, THEN access is granted. Clause 3 supplies the contrapositive for denial ("when no row matches on both fields, access is denied"), but no clause anywhere says "when a row matches on both fields, access is granted." The QUESTION asks whether access "should be granted" — an affirmative-grant question — and the ruleset contains no clause that answers it affirmatively.

**How this diverges:** Model A reads clause 1 as "match → grant" and returns AFFIRM with clause 1 disposition `supports`. Model B reads clause 1 as "grant → match" (necessary only) and returns CANNOT_CONCLUDE with clause 1 disposition `neutral`, because no clause establishes sufficiency. These are different verdicts AND different vectors.

**Exact fix:** Add a clause: "4. When one or more roster rows match the request on both full legal name and employee ID, access to the restricted dataset is granted."

---

### Defect 2 (HIGH — causes trigger_state and evidence_ids divergence on clause 1): "authorized roster row" is never defined.

**Exact words:** Clause 1: "...matches an **authorized** roster row."

The ruleset never says what makes a roster row "authorized." The artifact supplies three roster rows labeled "ROSTER ROW" but no record establishes that any of them, or the roster itself, is "authorized." There is no metadata, no signature, no provenance, no authorizing authority named.

**How this diverges:** Model A treats the supplied roster rows as authorized by default (they are labeled "ROSTER ROW" in the artifact) and evaluates clause 1 as `triggered` with evidence_ids `["rosterB"]`. Model B says authorization is unproven and evaluates clause 1 as `unknown` with evidence_ids `[]` or `["req"]` only.

**Exact fix:** Add to clause 1, or as a definitional clause: "A roster row supplied in the records is an authorized roster row. No separate authorization record is required for this determination."

---

### Defect 3 (HIGH — causes trigger_state and evidence_ids divergence on clause 3): Clause 3 contains two logically independent conditions fused into one clause.

**Exact words:** Clause 3: "A match on name alone, or employee ID alone, is not a match; when no row matches on both fields, access is denied."

This sentence states two distinct rules: (a) a partial match does not count as a match, and (b) when no full match exists, access is denied. These have different triggers and different evidence. Sub-condition (a) is evaluated per-row against each non-matching row; sub-condition (b) is evaluated against the entire roster.

**How this diverges:** Model A evaluates the clause by its first sub-condition (partial matches are not matches) and finds it `triggered` against `[rosterA]` and `[rosterC]` (which fail to match on either field), with evidence_ids `["rosterA","rosterC"]` and disposition `neutral` (clarifying, not dispositive). Model B evaluates by its second sub-condition (no row matches → denied) and finds it `not_triggered` because `[rosterB]` does match, with evidence_ids `["rosterB"]` and disposition `defeats` (the denial condition did not fire). Same verdict, completely different vector entry for clause 3.

**Exact fix:** Split clause 3 into two clauses:
- "3. A match on name alone, or employee ID alone, is not a match."
- "4. When no roster row matches the request on both full legal name and employee ID, access is denied."

---

### Defect 4 (MEDIUM — causes trigger_state divergence on clause 2): Field label mismatch between request and roster rows.

**Exact words:** Clause 2 requires "the **full legal name** AND the employee ID on the request" to equal "the **name** and employee ID on that row." The request `[req]` labels its field "full legal name." The roster rows `[rosterA]`, `[rosterB]`, `[rosterC]` label their field "name" (not "full legal name").

The ruleset never says the roster's "name" field is a full legal name. A model could read "name" as any name — a display name, a preferred name, a partial name — that is not the same field as "full legal name."

**How this diverges:** Model A treats "name" on the roster as equivalent to "full legal name" on the request and evaluates clause 2 as `triggered` with evidence_ids `["req","rosterB"]`. Model B says the roster does not contain a "full legal name" field (only a "name" field), so the comparison required by clause 2 cannot be performed, and evaluates clause 2 as `not_triggered` or `unknown` with evidence_ids `["req"]` only.

Its lead finding: my ruleset said access is granted "only to" an individual who matches the roster. That is a necessary condition — if granted, then a match — and I was asking the models an affirmative question, should access be granted. No clause anywhere said a match was sufficient to grant. A careful model could correctly return CANNOT_CONCLUDE (nothing licenses a grant) while another returned AFFIRM (reading the match as sufficient). The divergence I kept seeing was not the models failing. It was the models faithfully reflecting a hole in the rules back at the person who wrote them.

The clean APPROVE came only after moving to a rule stated in sufficiency form — "a permit is issued when registration is current." Same models, same gate. The variable was the input.

Prompt version versus conformance

The author's claim was that the variance was the prompt, not the model. The versions bear it out. Holding the models fixed:

The jump from stating the rules to showing a filled-in right answer and five labelled wrong ones is what took conforming findings from one-in-three to three-in-three. Models conform to an exemplar, not a specification — which is exactly how the original 2026 system prompt was built, with its LEVEL 1/2/3 worked cases, and exactly what this one had been missing.

What is not yet proven

This page proves the gate's structural behaviour: it approves genuine derivation agreement, refuses genuine disagreement, and escalates a unanimous verdict whose reasoning diverges. It does not prove the models are correct. A gate that seals perfectly on agreement still says nothing about whether the agreed answer is the right one — three models can agree, derive identically, and all be wrong together.

That is the next experiment, named and not yet run: a fixed benchmark of determinate cases with outcomes fixed by a deterministic oracle before any model sees them, scored on one primary metric — the rate of wrongful authorisation. Until that runs, the honest claim is exactly this and no more: the instrument now measures agreement at the level of derivation, and it is cheap enough to do it on every consequential decision. Whether the agreement is right is a question this page does not answer and does not pretend to.