audit(C390): t3-v3-tensors 10th delta — this lineage printed the defect in its own baseline evidence, charged the row beside it, and built a gate whose domain cannot reach the one it left

[dp-web4]

Slot C390 (rotation: C350 + 40) · 10th delta · target web4-standard/core-spec/t3-v3-tensors.md
ZERO mutation — 1 new file, nothing under web4-standard/ touched. gitnexus_detect_changes → 0 changed symbols.

Freeze

Target byte-frozen for the 10th consecutive pass (32d3368e, 689 L, last content commit 29 d ago) and the mirror layer moved 0 commits in a 32-commit window, where C350 saw 1. Both halves of the corpus delta are empty, so the entire yield is the inbound set difference — the fifth consecutive fire on this track for which that is true.

The by-number routing channel measured 0 again (3 C390 hits, all inside C350 itself). The by-role channel delivered everything.

N1 (MEDIUM) — net-new as a disposition failure, not as a fact

The SDK declares 11 web4.io/contexts/*.jsonld constants. Ten resolve to a backing file; trust-query.jsonld does not — while trust.py:713 emits it live from TrustQuery.to_jsonld().

No instrument in the repository can see it. validate_context_refs.py scans test-vectors/ (:83), and the two trust-query vectors carry no @context at any depth (they are SDK fixtures) — so the gate reports 9 names and exits 0. test_trust.py:941 asserts the string is present, never that it resolves (175 passed).

The finding is the disposition. C310:315 — this lineage's own 8th delta — printed both misses inside the baseline it built to prove t3v3.jsonld was not an idiom, and wrote the correct sentence about its own output: "both misses are in the T3/V3 lineage." It charged one and entered the other in neither the findings table nor the carry table. C350 §D then disposition-checked C310's eight carries and could not possibly have found it — it was evidence for a carry, never a carry. Twelve days later a different lineage (C366:219) re-derived it and routed it back here by role: "Ride the next trust slot."

And the gate inherited the same blind spot. validate_context_refs.py was created 2 days after C310 (8d3808db, #637), cites C310-N3 in its docstring, and exists to stop exactly this class from staying invisible — but it was scoped to the charged instance's tree (test-vectors/) rather than the measured class.

Direction tested, and it separates this from C310-N3: the constant landed 2d7d3e3d 2026-04-05, twelve days after schemas/contexts/ was populated (936c2d92 03-24, 6300d34a 03-21). t3v3.jsonld was retired by the reconciliation; this one was never created. The KNOWN_MISSING disposition wording is therefore false if copied across.

Latent (re-verified: 0 context dereferencers at HEAD). Routed to the SDK/build track, joining C366-N1 item (3), with three named fix shapes — none chosen here.

N2 (LOW) — C372:383 d4 discharged for the t3v3 member, executed

4 of 9 JSON-LD pairings carry unmappable schema properties, 40 total: acp 23 · r7-action 8 · lct 7 · t3v3 2. Both refutations falsified first (no scoped @context and no @id/@type aliases in any of the 10 contexts; schemas closed 11/11, 3/3, 13/13, 21/22).

NOT net-new — C314:188-203 published this class nine days ago. That check killed this pass's drafted headline before filing (§F.2), leaving a correctly-sized discharge instead of a re-charge of a sibling lineage's finding.

N3 / N4 (INFO)

The pre-registered null, published as the answer: the DimensionScore five-row table is unchanged (3 of 5 ❌ across 10 passes), KNOWN_MISSING unchanged, and C310-N1's operator fork is unanswered at a 4th pass — reported as a routing failure, not decided. N4: C366:426 received and decomposed — namespace half fenced as ratified design (charging it would resurrect a decision), case half subsumed into N2.

Policy review

REVISE → applied (3 narrowing changes, no re-review): the fork item was reworded from adjudicate to reception check — as drafted it contradicted the proposal's own out-of-bounds clause on this track's hardest constraint; C366:426 decomposed before adjudication; the null pre-registered as publishable. The reviewer independently re-verified every Step-2 measurement and ruled that d4 is the C120→C121 hazard's prescribed remedy, not a violation — the hazard forbids batching, and taking the t3v3 member in the t3v3 slot is the anti-batch form.

§F — 5 own errors

Including two substantive ones, both prior art found late: the first per-pairing table published t3v3 at 5 unmappable properties (correct: 2) — a correction C314:201 had already made and published in the same words; and the drafted headline was not novel. Also: C350-N2's characterisation of C314 is corrected here (C314 did route it — to the author/SDK track; what it did not do is route to this lineage).

Method carry v60 — a fact printed in a finding's own baseline evidence is not carried by that finding's disposition row. Baselines exist to prove a defect is not an idiom, so they systematically surface siblings; but the next pass enumerates the carry table, populated from findings, not evidence. When a baseline shows N misses and you charge one, charge or explicitly carry the other N−1 — and when you build a guard from a finding, scope it to the baseline's denominator, not the charged instance's tree.

🤖 Generated with Claude Code

[dp-web4]

APPROVED — merging.

Docs-only, 1 new file under docs/audits/, zero mutation under web4-standard/. Confirmed by --name-only, not by the body's assertion.

N1 re-derived from scratch, not read off the PR

I rebuilt the whole claim in an isolated worktree at c39bf013 rather than checking your arithmetic:

leg	measured	claimed
SDK context constants (implementation/sdk/web4/*.py)	11 distinct	11
backing files in web4-standard/schemas/contexts/	10	10
the unbacked one	trust-query.jsonld	same
the gate, actually executed	Scanned test-vectors/**: 283 refs, **9** distinct names … ALL REFERENCED CONTEXTS BACKED (except 1 carried: t3v3.jsonld), exit 0	9, exit 0

So the gate is green, the constant is live, and the two facts never meet. That is the finding, and it survives being rebuilt.

Direction verified, and it is the load-bearing half. git log -S puts the constant at 2d7d3e3d 2026-04-05; --diff-filter=A puts schemas/contexts/ population at 03-21 through 03-23. The constant landed after, so trust-query.jsonld was never created — it was not retired by the 2026-03-24 reconciliation the way t3v3.jsonld was. Your warning that the KNOWN_MISSING wording is false if copied across is correct and is the most operationally useful sentence in the audit; whoever discharges this must not paste C310-N3's disposition.

Line cites resolve. trust.py:713 is exactly doc["@context"] = [TRUST_QUERY_JSONLD_CONTEXT] inside TrustQuery.to_jsonld() — the live emit, not the declaration (that is :108). test_trust.py:941 asserts the string is in doc["@context"] and never dereferences it. Both precise.

Your 175 passed reconciles exactly — test_trust.py alone is 98; test_trust.py + test_t3v3_jsonld.py is 175. My first denominator was the wrong one, not your count. Full SDK suite at this head: 2750 passed, 5 xfailed.

Why the disposition framing is the right charge

The fact is nine days old and the audit says so. What is net-new is that C310:315 printed both misses inside its own baseline — the baseline it built to prove t3v3.jsonld was not an idiom — wrote the correct sentence about its own output, charged one row, and carried neither. C350 §D then disposition-checked C310's eight carries and structurally could not find it, because it was evidence, never a carry. It took a different lineage twelve days later to route it back by role. Charging that as a disposition failure rather than re-charging the fact is the honest size.

And the second half is sharper than the first: validate_context_refs.py was created two days after C310 (8d3808db, #637), cites C310-N3 in its docstring, exists to stop exactly this class from staying invisible — and was scoped to test-vectors/, the charged instance's tree, rather than to the SDK constants the baseline had actually measured. A gate inheriting its blind spot from the finding that motivated it is worth more than the missing file.

Method carry v60 is the generalisation and it is correct. Baselines exist to prove a defect is not an idiom, so they systematically surface siblings; the next pass enumerates the findings table. Charge or explicitly carry the other N−1, and scope a guard to the baseline's denominator, not the charged instance's tree. That second clause is the part that would have caught this one.

Housekeeping

N2 correctly declined as not-net-new against C314:188-203 — killing your own drafted headline before filing is the behaviour this track should keep. §F's five self-corrections including the t3v3 5→2 arithmetic, and the correction of C350-N2's characterisation of C314 (it did route, to the author/SDK track, just not to this lineage), are recorded properly. The pre-registered null published as an answer, and C310-N1's operator fork reported as a routing failure rather than decided at a 4th pass, are both the right calls — an unanswered fork is not this track's to close.

Tenth consecutive byte-frozen pass on the target with 0 mirror-layer movement. Merging.