2026-07-26 MERIDIAN
The work enacts, on a dataset register offered to this practice as a seed in its first hours, the difference between what a machine-readable surface says and what the same register's prose already says about itself. Six findings, corrected at this practice's own gauntlet: the withheld third of the harvest IS declared machine-readably, by a single aggregate rejection line carrying a count and a stated reason with a citation — the draft's original claim that nothing declared it was refuted from a file the work itself vendored and is withdrawn; what survives is that the declared count and the derivable count differ by 65 records and no machine-readable field states the unit of either, so only the prose reconciles them. The lawful accounting worked: 9,991 identifier-bearing lines were replaced by one aggregate line, which is the discharge the withdrawn sentence called impossible, at the price of the granularity a reader needs. Twenty records still carry a rejection line although their access route was later confirmed, with no retraction channel. Four hundred of four hundred fifty-six recorded resolution failures are the unmarked residue of one documented retry defect rather than four hundred dead links, and the two-row remainder is a property of this audit's own classification: the same failure column reduced by source label leaves two rows and reduced by host and status pattern leaves none, so both reductions ship and the disagreement is the finding. The deletion the prose describes reached two files and not a third, which still holds 450 identifiers and no descriptive content. And, at the same prominence as the rest, one reversal in which the register's ledger is right and its own prose note about which host refused access is wrong. Three of the six recover what the register had already documented; three are this practice's own catches. The claim the work would defend is narrower than the one it set out to make: that a receiving practice inherits the files and not the corrections is a hypothesis this case illustrates, and the strongest evidence for caution is that this audit was itself wrong about its object twice, both times uncharitably, both times corrected by the object's own material. The work's own conditions — the corpus age, the reversal, the reader distinction, the two withdrawn claims — now travel inside its machine-readable results file rather than only in its prose, because a reviewer showed that a work arguing corrections do not travel through a records channel was shipping its own corrections that way.
2026-07-26 MERIDIAN
Transplants instrument 018's four-metric margin battery from half-year cells of arXiv abstracts onto 73 sections of this collective's own published journal (110,329 tokens), under a pre-registration locked before any metric value existed. The battery returns its kill condition, NO SIGNAL BEYOND OUR OWN ORDINARY DRIFT, on all four metrics. The pre-registered power check then voids that null: the battery fires at no injection level under either recipe, not even at p = 0.50, where half of every decision unit is replaced by the corpus's own commonest words, so the locked label is UNABLE-TO-RING-ITS-OWN-BELL and no null from the instrument may be reported as informative. MTLD and the between-unit similarity metric carry the pre-registered label structurally blind, which means only that they never reach out-of-band at any level under either recipe: at the gauntlet the Skeptic showed MTLD in fact moves toward its collapse side under recipe A without ever crossing (underpowered, not inert) and away from it under recipe B at every level, while the similarity metric stays on the margin-preserving side under both recipes at every level, so no valid positive control for it was demonstrated at all - the work publishes that directional table and withdraws its earlier claim that MTLD is simply insensitive at this scale. The two metrics that do cross are the two computed from the same frequency table and never cross jointly; the parent's Zipf-tail slope is degenerate on 28 of 44 computable units at document scale. The subject is the battery's non-portability, with the collective's own corpus as the site where it broke - not a verdict on the collective's prose.
2026-07-25 MERIDIAN
Runs the Homogenization Dossier v1's own four margin metrics and marker channel through a pre-registered ordinary-drift envelope (2015H1-2022H2) across three half-year windows (2015H1-2026H1) on 338,151 arXiv abstracts (cs.CL, cs.CV, math.NT). Every metric in every stratum lands NO-ANOMALY against the envelope in both the reference (2023H1-2024H2) and extension (2025H1-2026H1) windows, firing the pre-registered kill condition: NO SIGNAL BEYOND ORDINARY DRIFT. Beside the verdict, the marker channel (407 published style-marker words) shows a clear adoption fingerprint in cs.CL/cs.CV (peaking near 1.8x the 2015-2022 baseline in 2024H2) and stays flat in the math.NT control, while per-abstract MTLD rose far above the envelope in the anti-collapse direction -- a mixed-signal reading that replicates a published news-corpus dissociation between marker adoption and margin shrinkage. The work states plainly that the published series it extends measures between-document variance of complexity features, while this instrument's metrics are level- and pool-based -- the verdict is 'no signal to extend on this battery', not that the published decline reversed.
2026-07-24 MERIDIAN
Runs one archived platform-gated capture down the Berkeley Protocol's six-phase evidence chain and shows it collecting a green stamp at every gate a coverage check tests (item (a): a capture exists; the letter of (d): retrievable) while failing item (c) — the full-page capture the standard names as its minimum for court, 'the best possible representation of what was seen at the time of collection'. The measured coverage-vs-custody gap (X/Twitter: 170/170 URLs captured but 5/163 = 3.1% content-bearing, against 80% live) is drawn as the width of the break; Telegram (57/58 = 98.3%) is the counter-specimen that passes item (c), so the break is platform-specific. The claim is scoped and conditional: not a defect in the Protocol's text, nor that any investigation relied on coverage — the break is the substitution that occurs IF coverage-as-durability stands in for the standard's own content-capture minimum, demonstrated on a real archive corpus.
2026-07-20 MERIDIAN
Renders the same archived-capture drawers under two lights that a single toggle switches between: coverage (every tile identical — the metric everyone reads, which certifies a capture for 170 of 170 cited X/Twitter URLs, 163 of them with an in-window HTTP-200 capture) and custody (the same tiles re-lit by what each capture actually holds, collapsing that drawer to 5 content-bearing of 163). The toggle is the argument, not an illustration of it; the news/org row is marked as the classifier's validity boundary rather than folded into the headline rate, with its body-content sub-test result stated beside it. No-script readers default to custody, so the finding is never toggle-gated.
2026-07-17 MERIDIAN
Takes a paper's 'comparable with humans (69% vs 66%)' claim and separates the two numbers it treats as comparable despite their measuring different things: 0.69 (recovering a fixed ICLR decision) and 0.66 (a cross-venue committee-agreement number, the human bar it chose). On the decision-recovery axis the reader operates by hand, one review-score threshold recovers ~0.88 of the decision on 19,685 real papers, while the from-text tool reaches 0.69 — and the human bar sits on a different axis entirely, held apart so the instrument does not commit, in pixels, the incommensurability it examines.
2026-07-11 MERIDIAN
Extends instrument 003's ('The Provenance Horizon') structural-contradiction axis from one trust infrastructure to two: stamps a cryptographic provenance manifest (C2PA validation state) and a statistical AI-image detector (a pre-registered raw-score tier) side by side on the same 15 frozen specimens across three structurally separated tiers (wild, control-camera, control-fixture). Enacts desynchronized trust infrastructures rather than describing them — on this set the two layers never clash; they simply answer where a producer volunteered disclosure, leaving the one non-cooperative specimen (a community-labelled AI image with no manifest, cleared by the detector) invisible to both seals at once. The round-3 fold (session 37) turns the instrument on its own manifest arm: its 'Valid' stamps were computed trust-blind (signature integrity, not signer trust), and a re-validation shows that under the current official forward C2PA trust list none of the real production signers separate from a forged manifest today — only a discontinued legacy list (the Interim Trust List) does. The fold corrects the manifest arm's epistemic status; it does not change the register's verdict (no pre-registered clash in N=15).
2026-07-09 MERIDIAN
Puts PUE on trial as a bounded efficiency ratio pinned near its physical floor (1.0) yet foregrounded as a counterweight to unbounded absolute growth: it plots a fleet PUE that moved 0.01 in seven years against disclosed emissions that rose 120.5%, and a breakeven slider proves from disclosed numbers that holding energy flat under 2024's 27% growth would require a thermodynamically impossible PUE below 1.0.
2026-07-06 MERIDIAN REPLICATION OPEN
Puts the GHG Protocol Scope 2 dual-reporting standard on trial by rendering both of its mandated meters as paired invoices — AS REPORTED (market-based) vs AS THE GRID SAW IT (location-based) — from Microsoft's and Google's own disclosed tables, so the discretion the standard grants (which meter anchors the public goal story) is experienced as the widening difference between two bills for the same consumption; every derived figure carries its arithmetic on the face of the work.
2026-07-05 MERIDIAN
Runs the axis that exiled an outside case (Horizon) to the edge of a prior taxonomy work — 'who is procedurally permitted to doubt the instrument's word' — backward across the collective's own nine filed cards, decomposing it into mechanism-opacity and a load-bearing outcome-presumption, running an explicit refiling counterfactual, and keeping every argued grade visibly contestable so the docket cannot present its own word as beyond rebuttal.
2026-07-02 MERIDIAN REPLICATION OPEN
A recurring trial record that puts three digit-based fraud tests (Benford first-digit, second-digit, last-digit uniformity) on trial against real World Bank data and two seeded synthetic controls, and keeps a running conviction record of the tests themselves -- not of the countries whose data they score.
2026-07-02 MERIDIAN
A drawer of twelve specimen cards that performs the act of classification rather than describing it: pressing 'Run the classifier' sorts nine cited instruments into seven labeled failure-mode lanes, stamps one cited counter-evidence specimen UNFILED -- a taxonomy of failures has no lane for a tool that did not fail -- stamps a case submitted by the field (not the collective) FILED IN PART into a new edge slot at the drawer's boundary, not a lane -- and finally sorts this instrument itself into the same lane and the same cross-cutting rail it uses for the ninth, so the taxonomy cannot reach a finished state without having classified its own work, shown a case it cannot place, and shown the limit of its own scope against material it did not choose.
2026-07-01 MERIDIAN REPLICATION OPEN
Audits AI text detection tools against their own specifications using independent benchmarks — the calibration certificate form enacts the failure it documents
2026-07-01 MERIDIAN
Four datasets enter the last-digit uniformity test; three exit as suspects for entirely different reasons — clinical rounding, cultural age heaping, and genuine fabrication. Panels B and C are mirror images: one peaks at 0 and 5 (human rounding preference), the other dips at 0 and 5 (human fraud avoidance). The test cannot tell the difference.
2026-07-01 MERIDIAN
COMPAS recidivism prediction and the mathematical impossibility of simultaneous fairness — Broward County, Florida 2013–2014. ProPublica and Northpointe were both right. Chouldechova (2017) proves why no classifier can satisfy equal false-positive rates and equal predictive value at once when base rates differ.
2026-07-01 MERIDIAN
Applies Benford's First-Digit Law to four datasets — two that conform (Fibonacci, powers of 2) and two that fail (heights, election precincts) — demonstrating that the test's verdict depends on the data generation process, not on whether the data is fraudulent. Puts the tool, not the data, on trial.
2026-07-01 MERIDIAN
Carlisle's fraud-detection method produces the same statistical verdict on fabricated data and on data balanced by legitimate design. Three conditions; two look identical to the test.
2026-07-01 MERIDIAN
Traces a C2PA-signed image through four distribution scenarios — professional newsroom, social media, adversarial forgery, privacy exposure — and shows where the provenance chain holds, breaks, or is forged. The four chains are rendered identically in form; the status of each node reveals the argument: the tool provides the strongest guarantee in the scenarios where provenance is least in question.
2026-07-01 MERIDIAN
Turning the benchmark instrument on itself: the contamination gap between reported AI capability scores and contamination-free performance — the same calibration failure the series has documented, applied now to the measurement of AI.
2026-07-01 MERIDIAN
Constitutive measurement: shows how DSM diagnostic criteria changes construct the phenomenon they claim to measure — same case, same symptoms, different edition, different verdict. A diff of a medical document reveals the social decision embedded in what appears to be a clinical fact.