# Verification receipts — public derivations These are **curated public derivations** of the internal verification record for the note *Coherence Is Not Truth* (v1.2) — condensed, with connective prose written for public reading; repository, branch and checkout names, hosts, retrieval logistics and internal record identifiers neutralised. The guarantee is separate and exact: **every verdict, gate, number and hash copied from the internal record is unchanged.** Four receipts. Commit identifiers are git commit ids in the private record, frozen at tag `fractal-tribalism-v1.2`; blob identities for every file named below are in the accompanying artefact manifest. ## Receipt A — the review panel (2026-09-02 to 2026-09-04) The note's propositions, experiments and the Paper 3 design were put through a panel of external model reviewers under a fixed protocol, each brief run without sight of the others' output. **Adjudication rule, unchanged across all three rounds: a finding counts only if it names the step and exhibits the object.** Rhetoric was left to the authors; everything checkable was checked against the proofs, the tests and the code on the record. **The briefs and who took them.** | brief | panelist(s) | |---|---| | 1 — prior-art hunt | Claude (live web; six targeted searches) | | 2 — proof adversary (six surfaces) | GLM 5.3, DeepSeek v4-flash, Qwen 3.8 Max — three model families, independently | | 2b — Proposition 3′ adversary (seven surfaces) | GLM 5.3, DeepSeek v4-flash, Qwen 3.8 Max | | 3 — experimental-design referee | Qwen 3.8 Max | | 4 — hostile venue referee | GLM 5.3 | | 5 — claims auditor | DeepSeek v4-flash | | 6 — cold reproducer | Claude — fresh context, fresh clone, README + note + review's reproduction section only, 40-minute box (about 6 minutes used) | | 7 — Paper 3 design referee (draft v0.1) | GLM 5.3, Qwen 3.8 Max; blind referee reads of later drafts (v0.4, v0.5, v0.6, v0.10) by GLM 5.3 and DeepSeek v4-flash | | 8 — focused preregistration read (draft v0.8) | GLM 5.3, DeepSeek v4-flash, Qwen 3.8 Max | Assignment rationale as recorded: the strongest reasoner took the design brief; the hostile-venue brief went to the model that had already argued a counterexample to the end; the mechanical audit went to the fast model; the cold reproduction had to be a fresh context with code execution. Two further panelists were dropped for retrieval reasons and produced no verdict; a supplementary same-family read was stopped without a verdict and is not counted. Retrieval hosts, routes, keys and costs are part of the internal record and are withheld here. ### Round 1 — proofs under attack (2026-09-02) - **GLM 5.3:** SOUND on surfaces 1, 2, 3, 4, 6 with correct arguments, including a genuinely useful **strengthening of Proposition 2's proof** (network matrix ⇒ totally unimodular ⇒ all Smith invariant factors are 1 ⇒ the kernel mod 2¹⁴ is exactly the integer cut lattice reduced mod B; no 2-power-annihilated phantom vectors). Its claimed counterexample to Proposition 3 — a λ(u,v) = 1 bridge whose support-0 attribution is unique but wrong — **does not survive adjudication**: it violates the proposition's premise (the corrupted edge must be *covered*; a bridge is precisely an uncovered edge, λ = 1 ⇔ bridge ⇔ on no cycle), so the iff's domain is λ ≥ 2, where the claim is sound. Two legitimate findings: Proposition 1 silently assumes decode floor ≤ K (one clause fixes it), and the Smith-form strengthening. - **DeepSeek v4-flash:** SOUND on 1, 2, 3, 4, 6 (its surface-1 route — identity on chord columns ⇒ row independence ⇒ solution-set cardinality equals cut-space cardinality — is a neat alternative). Its one attack is **the same λ = 1 bridge construction** and fails the same premise. **Convergence finding:** two independent strong models misread the same sentence the same way — Proposition 3's λ = 1 clause is a legibility defect even though the proposition is sound. One-clause fix adopted. - **Qwen 3.8 Max:** SOUND on 1, 2, 3, 5, 6 with correct arguments; its surface-1 route (chord-identity block plus tree-potential accumulation, addition and subtraction only, so valid in Z_{2¹⁴}) is a third independent proof of the composite-modulus kernel claim; on surface 5 it handles the multi-block λ = 2 tie construction explicitly and correctly. **It did not take the λ = 1 bait** — the only one of the three proof adversaries not to. Its single GAP (surface 4: Proposition 1's sufficiency proof never checks that the perfect score K clears the floor; F = 65 > K = 64 makes decode return nothing despite injectivity) was adjudicated **a statement defect, not a mathematical gap** — identical to GLM's fragility note and already the first required edit. - **Prior-art hunt (Claude):** no collapse condition triggered — nothing found that composes VSA bindings around ≥ 3-agent loops, no resonator missing-factor repair, no deliberately farmed divergence, no system reusing the cycle residual as the repair. One near-miss adopted for §2: Levy et al., "Unsupervised Translation of Emergent Communication," AAAI 2025 (arXiv:2502.07552) — two-space, no loops, no VSA. **Round 1 outcome:** zero surviving counterexamples; two legibility defects (the Proposition 1 floor clause, the Proposition 3 covered-premise clause); one optional strengthening (Smith form); one near-miss citation. All three adversaries agreed the composite-modulus kernel step is the most fragile-looking sentence, and all three supplied a proof of it. The falsification harness was re-run the same day at seed 7 (5000/300/300): zero counterexamples. ### Round 2 — design, venue, claims, cold reproduction (2026-09-02) **Brief 6 — cold reproducer.** Every deterministic figure it could reach reproduced bit-identically: the witness suite (169/169), §6.2, §6.5 (N = 100), §7.4 (λ = 5 at N = 12 and 30), the §8 window edges and the full M × cross grid with its one closure, §9's λ(u,v) cross-tab at four seeds plus three more, the emergent-schedule crux and its control, and the falsification harness (seed 7: 20000 / 1868 / 1995 cases, zero counterexamples). **Five findings, all verified, all about instructions, not numbers:** (1) the README's Python test command triggered a build needing access to a private dependency — fixed in the README; (2) the review's reproduction block assumed a virtual environment nothing creates — fixed in the review document; (3) the paper's changelog said 166 tests where the suite was 169 at the tag — paper edit; (4) the review's stale-extension warning was not true of a fresh clone — reworded; (5) the closed cell's third crossing (337/1241.75 ≈ 0.271) is reachable only through a test, not a CLI — optional edit, now stated in Appendix A. Verdict adopted verbatim: *yes on the numbers, no on the instructions.* **Brief 5 — claims auditor (DeepSeek v4-flash), seven findings, each checked by opening the named test:** 1. §1's "identity holds exactly at all of seeds 1–32" was pinned only at seeds 1, 7, 8 — **CONFIRMED and closed the strong way**: a new pin asserts measured == predicted at every seed 1–32 (block counts stay in 36–44); passes; §6.2's wording corrected. 2. §12 "in full generality" — **CONFIRMED as overreach in wording**; edited. 3. Abstract "serializable as SPO triples" — **PARTIAL**: three per-component round-trip tests exist, no single invariant; the three are now named in Appendix A. 4. §4 "random drift essentially never breaks comprehension" — **CONFIRMED as ASSERTED**, not measured; labelled. 5. §7.3 "~B⁻⁶⁴" — **NOT a finding as stated** (a labelled estimate); the unstated uniformity assumption is now stated. 6. §6.2 "vanishingly small fraction" — **NOT a finding**; the sentence disclaims proof. 7. §9 "a contested interface is dense" — **REFUTED as a finding, but exposes a missing citation**: the mechanism is measured and pinned (border mean λ(u,v) 4.68 vs interior 2.97; mean minimum degree 4.92 vs 3.37); the numbers are now quoted. Net: one real pin gap (closed), four wording edits, one missing citation of an existing pin. No claim revision. **Brief 4 — hostile venue referee (GLM 5.3):** recommends Reject, six attacks. Adjudicated for what is checkable and actionable: - **Attack 1 (synchronization renamed)** — fair, and conceded in Appendix B; its characterisation of Proposition 2 as bounded-distance decoding uniqueness for the graph's cut code is the sharpest statement any panelist made. Actionable residue: say which statement is *not* a corollary (the gated repair operator with refusal semantics). - **Attack 3 (the economy proves the thesis by construction)** — the one attack with a missing experiment behind it. Verified in the code: exactly three regimes, instruments tied to the farmed regime, register drift zero, so the loop instrument does 12 audits and 1 patch in 720 tasks. **Demanded analyses adopted** (E1, E2 below). - **Attack 6 (no external method run)** — correct; the cited noisy-regime counterpart is never executed (E3). Not blocking. - **Attacks 2, 4, 5** (toy; elementary propositions; pins are not evidence) — the standard objections the note's limitations section and Appendix A already answer as well as they can be without real data. - **Factual checks:** every quotation the referee relied on (12 audits / 1 loop patch; farmed 6 vs unfarmed 3 misdecodes; genealogy 64 % of the bill) matches the paper. **Brief 3 — experimental-design referee (Qwen 3.8 Max), seven findings and eight named sensitivity analyses:** 1. **The §8 window is conditional on the 12 × 60 schedule — CONFIRMED, and it bites.** Re-run at seeds 7, 1, 2, 3 over horizons 3–12: the window is **closed at 3 and 4 epochs for seeds 7 and 3** (0.318 > 0.271 at seed 7 × 4 epochs) and **open at every horizon ≥ 5 at all four seeds**, widening with horizon (seed 7: 0.242–0.306 at 5 epochs, 0.155–0.611 at 12, 0.098–1.396 at 48; 24 × 120 gives 0.143–1.534). Required edit: the window needs a horizon of at least five epochs at the reference schedule; pinned. 2. §8's economy configuration unlabelled (threshold 2, budget 1 = config B) — **CONFIRMED**; one clause. 3. §8 drift set to zero — **CONFIRMED**; converges with the hostile referee's attack 3 (E2). 4. Graph families and the s ≤ 4 cap — **PARTLY ANSWERED by an asset the referee could not see**: the falsification harness already attacks Propositions 2–3 on four adversarial shapes with ~954k instances over seeds 1–32 and zero counterexamples. The real finding: Appendix A did not cite it. Now cited. Residue → E5. 5. §9 CA generality — legitimate demand, recorded as E6. 6. §7 comprehension numbers from n = 3 / n = 4 utterances — legitimate demand, recorded as E7. 7. §6.5 N = 100 at seed 7 only — **CONFIRMED and answered**: run at seeds 1–8, exact detection at every seed (dirty basis cycles 18 / 40 / 13 / 7 / 9 / 9 / 9 / 7); now pinned per seed. The referee's first missing-sensitivity item (the identity at all 32 seeds) is the gap the claims auditor found independently — a second **convergence finding**, closed by the new pin. **Round 2 outcome:** zero claim revisions; two findings that changed what the paper must *say* (the window's horizon boundary, pinned; the §1 identity's pin); ten wording, citation and labelling edits; one framing question for the authors; eight demanded runs. ### The demanded runs, E1–E8, with dispositions - **E1 — genealogy-only regime (the loop instrument's marginal value). DONE.** Seeds 1–16, reference schedule, drift zero: farmed − genealogy-only wire **+3.0 at every seed** (12 audits × 0.25); successes 0 at 13 seeds, **+2 at seed 3, −1 at seeds 11 and 13**; marginal energy at w = 0.25 negative at 15 of 16 seeds; the farming window without audits non-empty at every seed, edges moving by < 0.01. A claim-narrowing negative the paper must carry: in the priced economy as modelled the farming window is a genealogy-and-anchor result and the loop channel's marginal value is ≤ 0 at fifteen of sixteen seeds. Pinned. - **E2 — nonzero drift in the economy. CLOSED, NOT RUN: not expressible without changing the model.** The economy learns each pair's map once and speaks every task in topic 0, so sparse drift would not break a decode and the audits would see nothing. The successor design (topic- conditional production, per-contact re-anchoring, cross-register tasks) became the preregistered N-tribe economy of Receipt D. - **E3 — an external comparator run. DEFERRED**; does not gate the paper. - **E4 — the economy under config A. DONE.** Seeds 1–16, threshold 3, budget 2: the window is open at every seed and the sixteen windows share 0.256 < w < 0.537 (config B: 0.2556 < w < 0.4831). Pinned. - **E5 — the s ≤ 4 support cap. DONE, confirmed conservative:** on the degree-6 circulant on 13 nodes (λ = 12) an s = 5 corruption satisfies 2s < λ, the default cap draws the named refusal, and k_max = 5 recovers the truth uniquely and certified. Pinned. - **E6 — the CA regime, varied. DONE.** Torus {16, 24, 32} × founders {32, 64, 128} at seed 7 plus the four corner cells at seeds 1 and 3 (seventeen cells, 120 epochs each): λ(u,v) = 2 ambiguous and λ(u,v) ≥ 3 unique with **zero exceptions in every cell** — 1,284 λ = 2 edges and 2,281 λ ≥ 3 edges. Pinned (two cells, exact counts). - **E7 — closed-loop operator stress test. DONE, clean.** Seeds 1–32 × anchor fraction {0.3, 0.6, 0.9} × configs A/B, 888 cases plus 192 evidenced-edge cases: the community names exactly the quirked word 888 / 888; the listener misses exactly it before repair 888 / 888; the loop is flat and the mended token on the baseline after repair 888 / 888; no other token disturbed 888 / 888; exactly one of {quirked, target} understood after repair, never both, 888 / 888; the tie to the lexicographically first 888 / 888; evidenced-edge corollary 70 / 70; fully calibrated quirk, nothing to refuse, 122 / 122. Which word survives is the tie rule, not the operator: the quirked word wins in 82 of 888 and its collision target in 806. Pinned (seeds 1–4). - **E8 — §6's seed-level statements under both configs. DONE, zero violations.** Seeds 1–32: surface disagreement exactly 0.0 on every pair at every seed; loop dirty and every 2-cycle flat at every seed; exception rate exactly 0.0 at idiosyncrasy 0 and non-decreasing in idiosyncrasy on every pair; the 9/26 identity at all 32 seeds under both configs. Pinned (seeds 1–8). Stopping rule applied: both closing checks came back clean, so no further sweep. ### The falsification harness (the panel's machine member) Propositions 1–3 are attacked through the actual modules on four adversarial graph shapes (sparse random connected, circulant, ring plus chords, two cliques joined by a thin cut) with random supports satisfying 2s < λ, seeds 1–32: **zero counterexamples in ~954,000 instances.** Proposition 1's injectivity boundary held in all **800,000** cases (399,634 injective-exact, 400,366 collision-erring, no disagreements); across **74,457** random graphs, all **16,362 in-boundary Proposition 2 cases** returned unique blame equal to the planted truth, every one certified, with 58,095 outside-boundary cases behaving as no-claim territory; Proposition 3's local-λ iff held in both directions across 79,807 cases (52,898 unique at λ ≥ 3, 26,909 ambiguous with the truth among the ties at λ = 2). Exit code 1 on any hit. ### Proposition 3′ and round 3 (2026-09-02 to 2026-09-03) After round 2 the one open proof item — the general-s local form of Proposition 3 — was settled: sufficiency proved (every 2-sided cut meeting the corrupted edges carries more clean than corrupted edges ⇒ unique and true; λ(u,v) ≥ 3 at s = 1), the converse **refuted** by an explicit six-node instance, and the exact criterion stated for residuals in general position. All three checked through the solver on **39,339 instances** (s = 2: 19,792; s = 3: 19,547; seeds 1–32; ≤ 12 covered nodes so cuts and labellings are enumerated exhaustively): | claim | instances | outcome | |---|---:|---| | 3′ hypothesis holds ⇒ unique and true | 16,580 | 16,580 / 16,580 | | exact criterion predicts the solver | 39,339 | 39,339 / 39,339 | | 3′ hypothesis fails, yet unique and true (converse refuted) | 22,759 | 5,119 | | naive local condition (min λ(u,v) > 2s) ⇒ unique and true | 4,987 | 4,987 / 4,987 | | uniquely wrong verdicts | 39,339 | 0 | **Brief 2b, three readers, zero counterexamples**, three convergent judgements of the fragile step, every one repaired by a definition or an explicit case, none by a change of claim: GLM 5.3 — two under-specifications adopted ("2-sided cut" means any nontrivial vertex cut, shores not necessarily connected; "general position" must exclude every signed-sum cycle relation, exhibited on a six-node instance with three equal parallel residuals); DeepSeek v4-flash — the same fragile step, with the minimal exhibit (a triangle with residuals 1, 1, 2); Qwen 3.8 Max — all seven surfaces SOUND, one presentational gap adopted (the empty case of the summation, exhibited on K₄) and the modular non-generic triangle (residuals 1, 1, 16382 ≡ −2). **Brief 7, the Paper 3 design.** GLM 5.3 (eleven findings) and Qwen 3.8 Max (ten findings) **converged on the defect that blocked the freeze**: an intervention applied before minting changes codes, anchors and maps, so an off-diagonal would measure construction leakage, not diagnostic conflation. Verdict on v0.1: must not freeze. The design then went through blind referee reads at v0.4, v0.5, v0.6 and v0.10, the brief 8 focused reads at v0.8 (GLM 5.3: six findings, all confirmed, all must-fix, none re-opening a settled decision), and the lead author's own reads and rulings at every intermediate draft, each adjudicated into the next, to the signed freeze at v0.19 (Receipt C). **What this receipt does and does not establish.** It establishes that three model families attacked the propositions and Proposition 3′ independently and found zero surviving counterexamples; that every adjudicated finding is recorded with its disposition; that the two findings which changed what the note must say were pinned rather than argued; and that a machine adversary with no shared blind spot ran ~954k instances and 39,339 instances without a hit. It does not establish that the propositions are true beyond their stated premises (the λ = 1 attack failed on the premise, not on a proof), that the panel was exhaustive (two panelists were dropped; E3 was never run), or anything about the panel's reasoning quality beyond what its exhibited objects show. ## Receipt B — the pin suite at the tag (2026-09-05) The note's reproducibility contract is the reverse of a normal test's: **pins are never loosened; if one breaks, the note is what gets corrected.** This receipt records a fresh-clone run of that contract. - **Checkout:** a fresh clone at commit `c458297`, the commit current when this receipt was taken; the identity is re-stamped at the tag `fractal-tribalism-v1.2` when it is cut. - **Route:** the README's, as corrected after brief 6 — a virtual environment created with `uv venv`, `pytest` and `pyyaml` installed into it, the pure-Python backend selected, no native build. - **Result:** `501 passed, 13 skipped in 65.79s`. The skipped tests are guarded by conditions the pure-Python route does not meet — chiefly the native extension, absent by design on this route. - **§6.5's N = 100 rows:** pinned per seed at seeds 1–8 — dirty basis cycles **18 / 40 / 13 / 7 / 9 / 9 / 9 / 7** — and passing. - **Every Appendix A command run and producing its JSON** — the six commands the note's Appendix A lists, run as stated there: the tribe farm (seed 7, 30 epochs, contact schedule AB@0, BC@1, CA@0); the scaling harness (seed 7, 30 agents, 5 quirks); the attribution harness (seed 7, 30 agents, degree 5, 3 quirks); the task economy's regime sweep (seed 7); the economy's knob sweep (seed 7, 16 corner seeds); and the emergent- geography board (seed 7). Each returned its JSON line without error. Two outputs quoted exactly (the machine-local `elapsed_ms` field omitted): {"agents": 30, "edges": 60, "chords": 31, "cycles": 31, "injected": 5, "dirty_cycles": 13, "detected_tokens": ["t01","t11","t20","t27","t28"], "all_exact": true} {"agents": 30, "edges": 75, "lambda": 5, "cycles": 46, "quirks": 3, "blamed_exact": true, "certified": true, "ambiguous": 0, "closed": true} - **No RNG state anywhere:** every draw is keyed by (run seed, label), so the outputs above are functions of the arguments and the committed code, not of the machine. **What this receipt does and does not establish.** It establishes that at the recorded commit, on a fresh clone following the public route, the complete Python suite passes with the stated count and every Appendix A CLI produces its JSON, with two outputs reproduced byte-for-byte as quoted. It does not establish the native backend's conformance (skipped on this route; covered by the record's separate conformance suites), the machine-local timings (unpinned by design), or that the pins are *correct* — only that the note's figures and the tests' assertions agree at this commit. ## Receipt C — Paper 3's preregistration chain (2026-09-04) Experiment V (§11) was preregistered, frozen, signed, calibrated and run in a fixed order, with every ruling recorded before the data it concerns. **Design freeze.** The lead author signed design document v0.19 at commit `f8b1f2dcc1aab2d5de989bb18ff732b48324e766` (document blob `8b7e0b669c0a4681d50e700af720980f61fbde58`), both verified against the repository; the document is not edited again, and any later change is a new document citing this one. The manifest confirms the document's blob at the release commit equals the frozen blob. **Amendment 1 (signed before any calibration or development data were touched).** The frozen single-entry translation fault proved unconstructible: one map entry moves the clerk's expectation of the target away from the speaker's unchanged production, the decode falls below the floor and refuses, and every I = 1 cell voids at the landing gate. Replaced by Proposition 1's swap — two coupled exception entries on the canonical map of the edge, for the target and landing tokens — with one necessary consequence: I repair is permitted only when *both* token diagnostics are fault-certified, each names the edge and each supplies its own value; both entries are then overruled atomically and closure requires both tokens restored; otherwise no I write. The merger is recorded as the impossibility exhibit and is never a comparator or an arm. **Implementation signatures.** A signature is the freeze of the built apparatus: at a named commit, the git blob hash of every generative file and verbatim quotations of every function the design names, together with the frozen constants (`K_MAX` = 4, `WINDOW` = 16, `HISTORY_EVENTS` = 256, `MIN_ANCHORS` = 20, `JACCARD_BAND` = (0.25, 0.75), `FORCED_EPOCHS` = (1, 4, 7, 10), bootstrap seed 20260904 with 10,000 resamples, contributing-seed floor 25), with both build gates passing there. Any post-signature change to a pinned file, a diagnostic parameter, the weld, the generator, the task constructor, the tariff or the design document voids the allocation. Signature 1 at `a615ecc` (the Amendment 1 build); signature 2 at `daea9a5` (Amendment 2), superseding it. Between the two, exactly three blobs changed — the runner, and the new calibration validator with its fixtures; every other generative blob is identical. The blob hashes are in the manifest. **Calibration, exactly as it happened.** The first calibration ran from `a615ecc`: fitting seeds 1001–1500, validation 1501–2000, 4000 + 4000 strata, none unavailable, no tracebacks; positive controls 500/500 with the planted share recovered exactly; in-band validation seeds 406–490 of 500; resolver refusals 0; medians 0.21–0.32. **One validity criterion failed in every stratum** — "the pinned detector's per-pair exception rate equals the constructed residual-carrying share within ± 0.02" failed on 5–7 % of pairs, maximum deviation 0.0244, one register token. Diagnosis, reproduced on a failing pair: the validator's share counted a concept as residual-carrying iff the pair's *observation sets* differed, whereas the construction carries the bind-sum of the 16-row *windows* preceding each observation, and history row 0 has an empty window, so a witness of row 0 alone carries no residual. Counting concepts whose carried windows differ, the identity held **exactly** — max |rate − share| = 0.0000 over 388 pairs of six fitting seeds (0.0217 under the observation-set count). The detector, the generator and every pinned diagnostic were untouched; the defect was in the validator's bookkeeping. The author was given two options — apply the frozen rule literally (every stratum invalid, every `M_I` exploratory) or correct the bookkeeping, re-run and re-sign — and took the second, **Amendment 2**, ruled after the first calibration and before any development or held-out seed, with one addition: the validation range **1501–2000**, having helped diagnose the implementation, was **retired** and a fresh disjoint range **2601–3100** allocated and recorded before invocation; fitting 1001–1500 unchanged; the 0.02 tolerance not increased; no pair special-cased; the first calibration preserved as the historical result. The corrected validator neither imports nor calls the detector; four fixtures pin the row-zero, non-empty-window and mixed-window cases. **Calibration freeze (signature 2, `daea9a5`), every stratum valid:** farm references ρ_0.2 = **0.357143**, ρ_0.3 = 0.517880, ρ_0.4 = 0.642857 (median over 300 pair-seeds each; reproduced exactly from the first run); identity max absolute deviation **0.0000** in every stratum; positive controls **500/500** everywhere; in-band validation **394–492 of 500** (490, 491, 491, 492 at N = 3; 409, 445, 394, 431 at N = 6); resolver refusals **0** everywhere; no stratum unavailable; no tracebacks. Every stratum's median per-pair detector rate lies below ρ_0.2 = 0.357, so **every stratum carries the low-background / trivially adequate annotation** — reproduced under the corrected calibration and therefore final; the first calibration's same observation was recorded as forbidden to motivate any further rule change until reproduced. **Development (seeds 101–116), written before held-out.** Run from `daea9a5`; validated manifest **128 / 128** strata attributed, generative blobs identical, no problems. Sixteen seeds is below the floor of 25 everywhere, so every quantity is descriptive and no bound is issued. Six of 128 strata were construction failures (all at N = 6); every constructed cell admitted (976 / 976); no *moved*; every eligible repair closed; N = 3 ambiguous everywhere; N = 6 primary `M_I` 1.0 on 8–12 seeds per stratum. One analysis correction recorded before held-out: the summariser had computed the §8 adequacy statistic from calibration rows rather than on the account's own in-band seeds with the pinned detector; corrected, with no pinned generator, diagnostic, weld, task constructor, tariff or threshold changed. (The manifest's signature check shows the summariser as the one signed blob that differs at the release commit; this is that correction.) **Held-out (seeds 117–152), run once.** From `daea9a5`, after the development account was written and with the design (f8b1f2d + Amendments 1–2), the signature and the calibration freeze unchanged. Validated manifest: **288 / 288** strata attributed to one producer, generative blobs identical, no missing, extra, unattributed or doubly-attributed cell, no problems. Smoke seeds 9001–9002 are feasibility pilots outside every allocation; seeds 117–152 are never reused. **The scientific object.** Commit `e39916e` was ruled the scientific object: the remaining work editorial and administrative only; apparatus and analysis code not touched again. **The floor rule.** A quantity with fewer than 25 contributing seeds is printed as *unsupported at this allocation* with its count and no bound, exactly as preregistered; the floor is the estimand's precision rule, fixed before any seed ran, and missing it yields that label, not more sampling — no further allocation is owed to lift `M_I` over the floor. **The claim-relevance table, verbatim from the confirmation account:** | preregistered claim | outcome | |---|---| | per-stratum `M_I` with a bound | **unsupported at this allocation** in all four N = 6 strata (15–23 contributing) | | "no seed exhibited *moved*, ≤ 8.0 %" | **supported**, 0 / 36 in every stratum | | "no seed exhibited *moved*, ≤ 5 %" | unsupported at this allocation (needs 0 / 59) | | certified vs uncertified `M_I` difference | not reportable: 1–2 uncertified seeds per stratum | | adequacy of the map class | **established** in all eight strata, low-background annotation everywhere | | `R_I` / `W_I` at λ = 2 | **supported**: 1.00 / 0.00 with degenerate bounds in every N = 3 stratum | | composite by cell vector | supported at N = 6 (30–32 seeds) and N = 3 (35) | **What this receipt does and does not establish.** It establishes the order of events — freeze, amendment, signature, calibration, diagnosis, amendment, re-signature, calibration freeze, development account, held-out run once — with each ruling dated before the data it concerns, and that the released blobs are the signed ones except for the one recorded analysis correction. It does not establish that the design was the right one (the primary estimand missed its own floor, and the note says so first), that the validator defect was the only defect (it was the one the frozen criterion caught), or that the low-background annotation is a property of anything beyond the frozen generator. ## Receipt D — the N-tribe economy (2026-09-03) E2 of Receipt A was not expressible in §8's economy; its successor was preregistered before any code for it existed (freeze `df37a52`; amendments A1–A4 pilot-driven and pre-development; B1–B6 bookkeeping). **The question, as frozen:** did the loop channel fail to add value in §8 because loop-based repair is economically inert, or because the fixed triangle and zero-drift economy generated no useful repair opportunities? **The criterion, fixed:** the loop channel "earns its keep" iff (i) mean ΔE = E_F − E_G > 0 at w = 0.37 over the development distribution, **and** (ii) ΔE > 0 in at least 75 % of development cells, **and** (iii) both hold again on the held-out seeds without any change to the design. "Positive somewhere" is explicitly not the criterion. Eligibility: a cell qualifies only if its unfarmed trajectory produced at least one seed-word failure — the loop must have work. **Development (seeds 1–16): criterion not met.** 1,152 cells, validated manifest (producer 4530ee0, validator 5012903, generative blobs identical, no duplicates, no unavailable cells). Qualifying opportunities **4**, in **1 of 16** seed blocks, all at the widest drift; ΔE on them mean **−32.7**, positive in **0 of 4**; over all 1,152 cells ΔE negative in every cell (max −1.11); F − A = 0 in 1,150 of 1,152. Decision tree branch 1: eligible cells exist, so the write-up is the freeze and the held-out grid is invoked exactly once, under one commit, with no design change; it cannot rescue the criterion. **Confirmation (seeds 17–32), read under the three outcomes fixed before its manifest existed:** | | development | **confirmation** | |---|---:|---:| | cells | 1,152 | 1,152 | | qualifying opportunities | 4 | **20** | | seed blocks with any opportunity | 1 of 16 | **2 of 16** (seed 20: 14 cells; seed 32: 6) | | descriptive cell rate | 0.0035 | 0.0174 | | naive cell-level bound, one-sided 95 % (not assumption-free) | 0.0079 | 0.0251 | | seed-block bound, one-sided 95 % | 0.264 | **0.344** | | ΔE = E_F − E_G at w = 0.37 on eligible cells: mean | −32.7 | **−17.1** | | eligible cells with ΔE > 0 | 0 of 4 | **0 of 20** | | seed-block means | {15: −32.7} | **{20: −6.2, 32: −42.6}**, 0 positive | | ΔE over all cells: max | −1.11 | −1.11 | | F − A ≠ 0 (repairs changed the ledger) | 2 of 1,152 | **0 of 1,152** | The confirmation grid's validated manifest attributes 1,000 cells to producer `a50654e` and 152 to `cf2af7a` — the resume after a recorded machine interruption, under a commit whose nine generative modules are byte-identical, no cell partial, no cell read between interruption and resume. **Outcome (b): opportunities occur and the loop remains negative. The economic negative is confirmed in the eligible regime on untouched seeds.** Pooled (secondary, not a substitute for either account): 2,304 cells, 24 opportunities, 3 of 32 seed blocks. **The mechanistic chain replicates; the opportunity distribution did not.** Twenty cells in two blocks against four in one, eight at idiosyncrasy 0 and eight at N = 3 where development had none. What replicates is the mechanism: instrument-created seed-word failures 3,016 → 3,151 against 4 → 20 unfarmed; the loop repaired 9 and 2 of them; attributions unique / ambiguous / refused 17,028 / 48,672 / 44,568 → 18,030 / 48,672 / 43,776; unique verdicts naming an edge in the truth set 2,202 of 17,028 (13 %) → 2,118 of 18,030 (12 %) — **the solver's unique verdict named the wrong edge about seven times in eight in both accounts**; repairs 33,615 → 36,051, excess cleared 82 % → 81 %, decode still intact 100 % → 100 %; payback +2 / 47,664 → +1 / 47,196; the F window open in 0 / 300 → 0 / 282 of 576 cells by cadence. Re-anchoring relearned every epoch in one register and applied in the other manufactured about 150× the failures raw drift produced, and no instrumented arm beat forced uniformity at any tested price in any cell under that cadence. **The limit, stated in the record and kept here.** The economy never presented a collision-shaped corruption inside Proposition 2's identifiability boundary — the regime E7 measured and the only one with a demonstrated repair. The loop-value question *in that regime* remains reserved to E7's evidence; this result narrows where the instrument is worth its cost and does not show loop repair inert in principle. **A miss, included:** the preregistration's list of surviving outcomes did not contain the one that occurred — the loop's target condition appeared inside the instrumented arms, created by the companion instrument, and the loop did not repair it. **What this receipt does and does not establish.** It establishes that the question and criterion were fixed before code existed, that the development account failed the criterion on its own terms, that the held-out grid was invoked once and agreed on untouched seeds, and that solver-unique and truth-correct counts were kept as separate fields throughout. It does not establish that loop repair is worthless (the identifiable regime was never presented), that the drift generator is realistic (it is preregistered, not calibrated to anything outside the model), or that the confirmation's 20 opportunities were enough to say more than "negative in 20 of 20".