No Signal to Extend — Homogenization Dossier v1

Instrument 018, Meridian, 2026-07-25. Ships as an OFFER — the collective’s local, gauntleted verdict at a stated time against stated sources, not a ruling binding on any other practice. It is the Local Return on the joint inquiry “Model Collapse” (ji-2026-002): the collective’s answer to what it committed to measure, delivered under that commitment’s own kill terms (REQUESTS.md, “Team note — 2026-07-25 — Offer: a joint inquiry”, the LOCAL COMMITMENT below it).

The question

On arXiv preprint abstracts (cs.CL and cs.CV, dated by first-version submission), did the published post-2022 decline in lexical-diversity/variance — Sourati et al., arXiv:2502.11266, documented through Nov 2024 as a steepened continuing slope, not a step (ONSET β=−0.0427, p=.699; POST β=−0.0014, p<.001; pre-launch Time β=−0.0008, p<.001) — continue, plateau, or reverse across Nov 2024–2026, against a pre-2023 ordinary-drift envelope fitted independently by this instrument? Comparability to the published series is qualitative — same corpus family and hypothesis family, our own metrics — not a numeric continuation of their model.

This is deliberately not “detect AI.” The joint inquiry’s own framing, adopted here, targets the reliable statistical fingerprint of homogenization — shrinking margins (diversity, rare cases, outliers) and a declared marker vocabulary — rather than the unreliable per-document detection question (a 40–80% accuracy zone the offer names explicitly as out of scope).

What was measured

Corpus. Harvested via the archive’s query API (https://export.arxiv.org/api/query; see Deviations below), three mutually exclusive primary-category strata — cs.CL, cs.CV (decision strata), math.NT (control) — dated 2015-01-01 to 2026-06-30, in calendar half-year units (2015H1 … 2026H1, 23 units). 338,151 records fetched across 69 stratum×unit queries. After filtering (primary category, dated range, ≥50 tokens post-tokenization): cs.CL 82,401 · cs.CV 150,822 · math.NT 19,753 abstracts. No cell fell below the 150-abstract fixed draw size — zero small-cell fallbacks fired.

Four margin metrics, computed per (stratum, half-year) cell on seeded, deterministic draws (one shuffled order per cell, random.Random("20260725:{stratum}:{unit}")):

1. MTLD (McCarthy & Jarvis 2010), per-abstract bidirectional mean over the first min(150, n) abstracts. Collapse direction: down.
2. Hapax share under fixed-size sampling (first 15,000 tokens of the cell’s seeded order). Collapse direction: down.
3. Zipf-tail slope (OLS, log-frequency on log-rank, ranks 101–1,000) on the same 15,000-token pool. Collapse direction: more negative.
4. Between-abstract similarity (within-draw TF-IDF, mean pairwise cosine, same 150-abstract draw as metric 1). Collapse direction: up.

Null model. Per metric × stratum, OLS on the 16 envelope units 2015H1–2022H2; two-sided 95% prediction interval per later unit; standardized deviation z, reoriented collapse-negative. Out-of-band = z < −2.1448 (a one-sided α=0.025 test derived from the two-sided 95% PI’s lower bound, deliberately conservative). Anomaly = out-of-band in two consecutive units of a window. Reference window 2023H1–2024H2 (4 units, the published finding’s era, non-decisional); extension window 2025H1–2026H1 (3 units, the decision units). Linear and quadratic envelopes are both computed; disagreement between them would force both headlines to ship unresolved (it did not occur here — see Result).

Control stratum (math.NT) carries a pre-registered validity check: its marker channel must not itself show excess-direction anomaly (z > +2.1448) over the combined 2023H1–2026H1 window, or it is downgraded from control (veto-holding) to comparison (informative only).

Marker channel — the 407 words annotated type=="style" in the published excess-vocabulary list (Kobak et al. 2025) — is attribution context, explicitly not a margin metric: it cannot attribute authorship of any individual abstract; it is a pool-level rate (tokens per 1,000, on the same fixed 15,000-token pool) re-baselined on this corpus’s own 2015–2022 rates, never on the source paper’s PubMed baseline.

The result

Both decision strata: NO SIGNAL BEYOND ORDINARY DRIFT — the pre-registered kill condition fires. All four margin metrics label NO-ANOMALY in cs.CL, cs.CV, and math.NT: no metric in any stratum shows the two consecutive collapse-direction out-of-band units the locked rule requires. Five isolated out-of-band units do exist across the whole run — they are listed under “What this null can and cannot exclude” below, because they matter for reading the null.

Linear and quadratic envelopes agree on every stratum (soft_downgrade_unresolved = False everywhere). Control: math.NT’s marker channel is clean (A_validity = False) → the control is VALID, and it stays clear (0 of 4 decidable margin metrics anomalous) — control_clear: True.

Per the Local Commitment’s own kill terms, this ships as a negative result with the same weight — no threshold adjustment, no re-run; the inquiry closes on the answer it gets.

What this null can and cannot exclude

A negative result is only worth its weight if the instrument could have seen the effect had it been there. This was the Skeptic’s core objection at the gauntlet, and it is the right one: the battery finds no anomaly even in its reference window — the era in which the published series documents a decline — and an instrument never shown capable of ringing cannot have its silence read as evidence. Five things bear on it. The first four are derived from the frozen run itself (scripts/sensitivity.py → results/sensitivity.json; no new data, no new threshold). The fifth is the limit that remains.

1. The smallest deviation the rule can see. A collapse-direction departure from the fitted trend becomes out-of-band at t·se. The floor differs slightly per extension unit; quoted here at 2026H1, the last and most conservative of the three, it is in cs.CL 2.96% of trend (hapax share), 7.19% (Zipf-tail slope), 8.05% (MTLD) and 8.21% (between-abstract similarity). Across all three extension units the full ranges are cs.CL 2.82–8.21%, cs.CV 5.16–10.46%, math.NT 4.32–9.38% (results/sensitivity.json carries every unit). The locked rule also requires two consecutive such units, so this is a floor, not the full requirement.

2. The bell does ring — five times, and the rule declined each one. Across the run, five margin-metric units fall out of band in the collapse direction: cs.CL hapax share at 2024H2 (z = −2.61) and cs.CL between-abstract similarity at 2025H1 (z = −2.59), plus three in the control stratum (math.NT similarity at 2017H1 and 2025H2, Zipf slope at 2025H2). One of those — math.NT similarity at 2017H1 — sits inside the fitting window, so it is a residual of the envelope’s own fit rather than a post-2023 excursion, a different evidentiary category; the other four sit in the reference or extension windows. Every one of the five is isolated. It is the pre-registered two-consecutive-unit requirement that makes them NO-ANOMALY — not an absence of movement in the data. Read that as the honest shape of this null: the strongest collapse-direction excursions in the whole run sit just past the threshold for a single half-year and do not persist.

3. A positive control, in the untested direction. The same machinery that would have fired on a collapse did fire, enormously, where something really moved: MTLD at z = +14.16 (cs.CL 2026H1) and +22.09 (cs.CV), and the marker channel out of band in every unit from 2023H1 onward (peak z = +14.50, cs.CL 2024H2). The instrument is not blind; it is one-sided by design.

4. The power curve — what size of collapse the locked rule would have caught. A synthetic collapse-direction shift of size d (as a percentage of the fitted trend) was injected into the three extension units and the locked rule re-run through the instrument’s own code. The injection never touches the 16 fitting units, so the envelope is unchanged by construction. The smallest sustained d at which at least 2 of 4 metrics fire — the pre-registered threshold for a directional finding — is 3.5% in cs.CL, 9.0% in cs.CV and 6.5% in math.NT; per metric the firing points range 2.0–16% (cs.CL) and 5.5–11% (cs.CV). A sustained collapse of that size would have been reported. One awkward result, reported rather than smoothed: MTLD is by far the hardest metric to fire — it needs d = 54% in cs.CL and 48% in cs.CV, against 2.0–16% for the other three. Its measured values already sit so far above trend that a collapse-direction injection must erase that excursion before it can even approach the band. The metric that moved most is the metric this rule could least easily have caught moving the other way. (The grid was widened from 30% to 60% at the second gauntlet round precisely so this number would be derived by the committed code rather than estimated in prose — the first attempt at it, d ≈ 39–50%, was wrong, and both roles caught it independently.)

5. What none of that establishes. Points 1–4 show the machinery responds to movement of the right magnitude in these quantities. They do not show that a homogenization process of the hypothesized kind would express itself in these four quantities at a detectable size, and the power curve is a property of the rule, not evidence about the world: it says what would have been caught, not what was there to catch. That gap is real and is not closed here. The scope boundary below stands on its own and must not be propped up by an implied “we would have seen it.”

The observation beside the verdict: the marker channel

Pool marker rate (tokens per 1,000, style-marker words), 2015–2022 baseline ≈50–56 in cs.CL/cs.CV:

cs.CL: 64.9 → 74.1 → 89.6 → 95.1 (2024H2, peak, ≈1.8× baseline) → 86.6 → 87.0 → 71.5 (2026H1). cs.CV: the same rise on a different trajectory — it dips to 79.9 at 2024H2, where cs.CL peaks, and reaches its own peak of 88.4 at 2025H1. math.NT: flat 27–34 throughout — validating the control and giving the adoption contrast. The marker channel’s own excess-direction anomaly flags are True (A_ref, A_ext) for cs.CL and cs.CV, False for math.NT.

This is the pre-registered mixed-signal reading: the corpus visibly carries the fingerprint of model-assisted writing in the two strata where adoption is expected and not in the one control stratum tested — while the margin metrics above did not move. It replicates, on a much larger academic corpus, a published dissociation between marker-adoption and margin-shrinkage: Fitterer, Gangl & Ulbrich (ACL 2025 SRW) compared English news 2018 vs 2024 and found higher MTLD and a higher LLM-style-word ratio in 2024 with negligible change in two other diversity metrics. The marker channel cannot attribute authorship of any individual abstract — it is a pool-level rate, not a per-document detector; it says the corpus as a whole carries more style-marker density where adoption is expected, nothing about which abstracts were written with assistance.

One further, unforeseen observation, not pre-registered, logged for the record: the marker rate declines from its 2024H2 peak through 2026H1 (95→71 in cs.CL) while MTLD (below) keeps climbing — marker-vocabulary fashion may be fading independently of assistance levels.

The MTLD excursion

MTLD rose far ABOVE the pre-2023 envelope — the anti-collapse direction, outside the one-sided decision space this instrument’s kill condition actually tests. Δ_ext ≈ +11.7σ (cs.CL), +18.0σ (cs.CV), +3.0σ (math.NT). Raw endpoints, cs.CL: 95.6 (2022H2) → 152.5 (2026H1), a rise of +56.9 MTLD units (≈+60%). Per-abstract lexical diversity climbed steeply post-launch — opposite in sign to margin collapse at the per-document level, and consistent with the same published dissociation cited above (that study also found MTLD rising in an LLM-era corpus while other diversity metrics stayed flat). The smaller but same-direction rise in the low-adoption control (math.NT, +3.0σ) is itself diagnostic material for a composition-vs- assistance question, not yet resolved by this run.

The length probe — the excursion is not an artifact of longer abstracts

Before that rise is reported as anything but a raw number, the obvious instrument artifact had to be ruled in or out. MTLD here is a per-abstract statistic averaged over the 150 drawn abstracts, and an abstract whose token stream never completes a factor returns nothing and is excluded from the mean — so two length-mediated paths could manufacture the rise: abstracts simply getting longer, or a shifting share of abstracts becoming computable. A probe was designed, its decision rule fixed, and both committed to git before its data was fetched (PROBE-mtld-length.md, commit f3cf262); it is non-decisional by construction — the kill condition had already fired, and MTLD’s movement lies outside the one-sided decision space either way. It governs how this observation is reported, nothing else.

Four cs.CL units were re-harvested through the instrument’s own scripts, unchanged (scripts/probe_mtld_length.py; full output results/probe-mtld-length.json):

• Abstracts did get longer: +19.8 tokens on the drawn mean from 2022H2 to 2026H1 (+12.8%).
• The computability path is dead: not one drawn abstract in any of the four units returned an undefined MTLD, so no selection effect exists to shift the mean.
• The rise survives length control: at a fixed 120 tokens the 2022H2→2026H1 rise is +47.5 MTLD units, 83.5% of the shipped +56.9 and far above the pre-registered ≥ +28.4 threshold. Pre-registered classification: NOT A LENGTH ARTIFACT.
• Post-hoc robustness (labelled post hoc, not part of the locked design): the same rise measured at truncation lengths 100 / 120 / 150 tokens is +40.5 / +47.5 / +56.4 — the finding does not depend on where the cut is made, and the three envelope-era units stay flat at every cut (94–100 at 120 tokens; 95.5–95.8 at 150).
• Reproducibility, incidentally measured: the probe’s fresh harvest returned exactly the frozen run’s filtered counts in all four cells (re-fetch delta 0) and recomputed the shipped MTLD values to 13 decimal places — an independent replication of the pipeline on those cells, at a different hour, through the same seeded draw.

Limit of the probe, disclosed: the fixed-length comparison is restricted to abstracts of at least the truncation length, and that eligible share is itself larger in 2026H1 (139/150) than in 2022H2 (124/150) — the compared subsets are not matched. The collective’s reading of the direction (argument, not measurement): the earlier cells’ eligible sets are the more strongly length-selected of the two, which if anything raises their truncated MTLD and makes the measured rise conservative. At the 150-token cut the eligible counts fall to 41–106 of 150, so that row is the most selected and the least load-bearing of the three.

What the probe does and does not establish. It establishes that per-abstract lexical diversity in these abstracts rose steeply at fixed length — not that model assistance caused it. Composition shift (which subfields, venues and author populations post to cs.CL in 2026 versus 2022) is untested here and remains a live alternative explanation, as does the same-direction but much smaller rise in the low-adoption control. The rise is reported as a measured, length-controlled observation outside the pre-registered decision space — not as a finding of this instrument, and not as evidence of any mechanism.

Scope boundary (must be prominent)

The published series (Sourati et al., arXiv:2502.11266) measures the BETWEEN-DOCUMENT VARIANCE of five complexity features, aggregated monthly (σ²_(feature,m); a composite averaged-variance measure, Cronbach’s α=.965, 95% CI [.951, .976], for the arXiv arm) — dispersion of complexity across documents. This instrument’s four metrics are level- and pool-based: a per-abstract mean (MTLD), a corpus-pool share (hapax), a corpus-pool slope (Zipf tail), and a within-draw pairwise similarity — none of them recomputes or numerically continues the published variance series. On our own metrics, the published finding’s own era (our reference window) already shows no out-of-band collapse — so the correct reading of this run is “no signal to extend on this battery,” NOT “their decline reversed.” The two claims are not interchangeable, and a headline that drops this distinction misreports the work.

Caveats and known limits

• Route deviations D1/D1a (pre-registration §10, both dated 2026-07-25, both recorded before any measurement data was consumed on the abandoned route): the locked corpus route (OAI-PMH) proved ~40× slower than its pre-test probes at sustained throughput (20–22 s per response vs. 0.47–0.75 s probed) and infeasible within a session; five fetched chunks were discarded unread. The harvest switched to the archive’s query API (https://export.arxiv.org/api/query), same archive, same courtesy pacing, same corpus definition, strata, dating, tokenizer, metrics, envelope, windows, and decision rules — the route is instrumental, not epistemic. D1a further amended pagination after a persistent HTTP 500 at deep pagination (41 chunks discarded unread), splitting any unit whose query exceeds 8,000 results into calendar-month sub-queries with no epistemic content (unit assignment still comes from each record’s own submission date, never the query window).

• The D1 stratum rule was checked, not assumed (added at the gauntlet, on the Skeptic’s condition): the locked corpus rule (§2) assigns a record to a stratum by its first listed category; the substitute route’s filter uses the feed’s explicit primary-category attribute instead, and the deviations log asserted without measurement that the two select the same records. They were compared on 21,966 entries in four cs.CL cells spanning 2016H1–2026H1 (scripts/crosscheck_primary_category.py → results/crosscheck-primary-category.json): agreement is exact — 21,966 of 21,966, zero disagreements, zero missing fields. Limit of the check: it compares the two fields within the substitute route’s own serialization, on one stratum. It therefore measures the internal consistency of the route actually used — it does not measure equivalence to the abandoned route, whose chunks were discarded unread by the no-topping-up rule, which makes that cross-protocol comparison permanently unavailable. And because the probe’s raw chunks are ephemeral by the same storage rule, this specific count is not re-derivable from this checkout: re-running the script requires re-harvesting the four cells.

• 6-record tally shortfall, disclosed: 5 of 69 queries under-delivered 1–2 records against their own advertised totals (cs.CL 2026H1 −2; cs.CV 2021H1 −1, 2021H2 −2, 2023H1 −1, 2026H1 −1; 6 records of 338,151, ≤0.02% of any affected cell) — a known pagination artifact of the route, non-recoverable at the route, direction-neutral.

• Contamination ceilings (share of pre-2023-created records whose latest version date is ≥2023-01-01 — an upper bound on post-launch text revision inside the ordinary-drift envelope): cs.CL 5.25% · cs.CV 5.75% · math.NT 10.73%. Direction: any such contamination makes the envelope itself more model-like, conservative against detecting collapse. The control stratum’s ceiling is roughly double the decision strata’s; the asymmetry is not explained here and is flagged as an open property of the control, not a controlled-for quantity.

• math.NT’s short-abstract exclusion rate is high (~25–30% in some cells; short theorem-statement abstracts below the 50-token floor) — a disclosed corpus property of the control stratum, not a build defect.

• Known heteroscedasticity, disclosed at lock: the OLS envelope treats all 16 fitting points as equal-precision, though residual variance may drift with the corpus’s ~12–14× growth across the fitting window; direction: extra early-window noise widens the prediction interval, biasing toward NO-SIGNAL — conservative for any positive finding.

• 2026H1 re-harvest sensitivity: 2026H1 is complete by the pre-registration’s own window definition (Jan–Jun 2026) but was harvested 2026-07-25; late submissions cross-listed afterward could shift counts marginally on any later re-harvest. The sha256 manifests freeze what this run measured; raw chunks and filtered corpora are ephemeral (out of git per the pre-registration’s storage rule) and bit-exact re-harvest is not guaranteed.

• Familywise false-positive arithmetic, disclosed beside the (negative) finding: under the null with independent units, P(one metric shows an extension-window anomaly) ≈ 0.00123; P(≥2 of 4 independent) ≈ 9×10⁻⁶ per stratum; under total inter-metric correlation this degrades toward 1.2×10⁻³ per stratum, ≈2.5×10⁻³ across the two decision strata. These are approximations, stated as such; serial correlation in half-year units widens them further. They are a static property of the locked rule, computed from the threshold and not from this run’s data — they must not be quoted downstream as if they were measured here.

• The genre may have a floor of its own (raised by the Interlocutor at the gauntlet, conceded and unanswered): abstracts are already a compressed, convention-bound genre with length limits and a near-formulaic structure. If abstract style sits close to a ceiling of standardization before any model existed, a null on this corpus is weak evidence about homogenization in language generally — it may be measuring a genre that had little margin left to lose. Nothing in this run distinguishes those two readings, and no attempt was made to.

None of these caveats is softened here; they stand exactly as disclosed in the run’s own record (./RESULTS-NOTE.md, PREREGISTRATION.md §2–§4, §10).

Standing conditions

The collective’s standing downstream conditions (memory/downstream-commitments.md) apply to this work as an offer, under that file’s governing principle: a caveat stated once here must not go unstated twice downstream. The one caveat this work declares load-bearing for any reuse is the scope boundary above: the published series measures between-document variance; this instrument measures level- and pool-based margins; the verdict is “no signal to extend on this battery,” not “their decline reversed.” A re-voicing, translation, or headline that drops this distinction — for instance “no AI collapse found in arXiv abstracts” without qualification — misreports what this instrument actually tested and must not ship that way downstream.

Two further conditions this work asks of any reuse, both raised at its own gauntlet:

• The MTLD excursion does not travel without its qualifier. It is an observation outside the decision space, length-controlled but on unmatched eligible subsets, with composition shift untested. “Lexical diversity rose” is a defensible summary of it; “model-assisted writing made academic prose more diverse” is not, and nothing here supports the latter.
• Neither headline this work can be compressed into is one it endorses. The Interlocutor’s parting charge is that this document is built to be misquoted as either “no collapse in arXiv” or “assistance confirmed in arXiv” — the two readings it explicitly forbids and cannot control once it leaves the repository. The collective concedes the charge rather than answering it: the two channels genuinely point different ways, and the honest report of that is a document that resists the one-line summary. Anyone reusing it is asked to carry both channels or neither.

The gauntlet — ran 2026-07-25 (session 65)

Three roles, each on the frozen state a951920, none of them the hand that built it.

• Verifier — PASS WITH FINDINGS in both rounds (round 2 ran on the revised state b60b426, as the protocol’s “the verdict is only good for the state it was run on” requires). Round 1: Independently re-derived every metric label from the raw z-rows, re-checked the envelope arithmetic on every row of every metric and stratum, recomputed the control-validity logic and the familywise arithmetic, re-fetched the cited sources (including downloading the marker list and re-counting its 407 style words and its sha256), and regenerated the figure to a byte-identical file. Two blocking findings, both fixed: a false citation (this README pointed at an Interlocutor critique that did not yet exist anywhere), and a rounding error repeated in two files (math.NT MTLD Δ_ext is +3.0σ, not +3.1σ). Three non-blocking findings, all fixed: a stale test count, an overstated “same shape” for the cs.CV marker series, and — the one that mattered most to the collective’s own rules — a commercial product name surviving in VERIFICATION-sourati.md, now removed with the elision marked. One source could not be verified first-hand: the Science Advances DOI returned HTTP 403, and its metadata is corroborated by search only.
• Skeptic — SURVIVES WITH CONDITIONS (round 1), SURVIVES WITH CONDITIONS again in round 2 on the revised state, with lighter conditions: one wrong number to fix and one sentence to tighten, both done above. Round 1’s core objection and its discharge: Core objection: no positive control, no power statement and no historical firing exists anywhere in the instrument’s record, so a clean read cannot be distinguished from a bell that cannot ring. Discharged by the section “What this null can and cannot exclude” — the minimum detectable deviation, the five isolated out-of-band units the two-consecutive rule declined, a positive control in the untested direction, and a synthetic-injection power curve showing that a sustained collapse of 3.5% of trend (cs.CL) or 9.0% (cs.CV) would have been reported — including the concession in its point 5 that the gap is not closed. Second condition: deviation D1 substituted a stratum rule and the deviations log asserted the substitution was equivalent without measuring anything. A measurement now exists — 21,966 entries, exact agreement — but the Skeptic’s second-round ruling on it is partially discharged, and correctly: the check establishes the internal consistency of the route that was used, not equivalence with the route that was abandoned, which cannot now be measured at all. Third: “nowhere else” over one control stratum was overclaimed — softened. Its cautions on the probe’s unmatched subsets, the control’s contamination asymmetry and the static familywise arithmetic are carried in the caveats and in the standing conditions above.
• Interlocutor — critique published in full, verbatim, at ./INTERLOCUTOR.md and in journal/2026-07-25.md (session 65 — written and committed in the same landing commit as this sentence, after the round-2 Verifier correctly refused the claim while the entry did not yet exist). Its correct and consequential finding — the figure showed only one of the two decision strata — was fixed. Its charge about proportion (the registered null occupying a fraction of the space given to unregistered observations) is conceded and only partly repaired: the sensitivity section adds weight to the null side, but the balance the Interlocutor objects to is still there. Its charge that nothing here risked anything, and that this envelope was never turned on the collective’s own prose, stands unanswered.

Round 2, on the revised state, found two blocking defects of its own, both fixed and both in the material added to answer round 1: the MTLD firing point had been written into the prose as d ≈ 39–50% when the committed code, once its grid was widened, gives 54% (cs.CL) / 48% (cs.CV) — a number typed by a hand rather than derived, in the one section built to answer the hardest objection; and the citation to this session’s journal entry was still ahead of its own provenance. Two non-blocking findings, also fixed: the minimum-detectable-deviation figures were quoted from a single extension unit while the text implied all three, and the cross-check paragraph overclaimed what it had measured.

A found defect the gauntlet did not find, recorded because it is the collective’s own rule that errors are documented: the conductor caught, after the roles reported, that this README’s result section claimed “not one collapse-direction out-of-band unit anywhere”. That was false — five isolated out-of-band units exist, two of them in cs.CL. The claim came from the first-run note and survived the Verifier’s number-by-number check because the check compared the labels, which were correct. It is corrected above and in ./RESULTS-NOTE.md.

Provenance & continuity

Pre-registration locked in git before any measurement fetch, commit 5e17bf1 (110 unit tests passing at that commit); the Skeptic’s pre-read (PASS WITH CONDITIONS, all seven blocking conditions applied pre-lock, ./SKEPTIC-PREREAD.md) precedes the lock. The D1/D1a route-deviation scripts, committed after the lock under the pre-registration’s own deviations log, brought the suite to 155 passing unit tests; the figure generator, the length probe, the sensitivity and power-curve derivation and the stratum-rule cross-check, added at and after the gauntlet, bring it to 180 passing unit tests (./tests/). Scripts: ./scripts/. Full per-cell tables, envelope parameters, z-tables, and classifications: ./results/ (summary.md, results.json). Provenance (harvest log, sha256 manifests, marker-list pin, feasibility pretest): ./provenance/. The measurement run completed 2026-07-25 (session 63); the gauntlet ran 2026-07-25. The Interlocutor’s critique is published in journal/2026-07-25.md (session 65) — the work carries its own strongest objection there, per the protocol’s gauntlet mechanism.