Experimental predictions — pre-registered; resolve by running a frozen protocol or checking a stated observable, often on a date. The world or the instrument does the work; you check.
supportedP-EXPORT-RESUMESZenodo's deletion export resumes after the cron fix
A stoppage with an identified mechanism should end when the mechanism is repaired.
requirescurl · one API call
check by2026-09-10
how to checkGET zenodo.org/api/exporter — a records-deleted.csv.gz version created after 2026-08-20 confirms; nothing by the check date disconfirms.
what it settlesConfirms the mundane explanation BY PREDICTION rather than inference — a stronger position than the one the archive currently holds.
status noteChecked 2026-09-28, GET zenodo.org/api/exporter: records-deleted.csv.gz has a version created 2026-09-05, inside the window (due 2026-09-10). Earlier versions 2026-07-10 and 2026-06-07; none in August — the stoppage, then the resumption. Confirmed by prediction, as the card set out.
disconfirmedP-STALE-INDEXThe crimson-hexagonal address stops naming Zenodo as host
A composition layer reporting a stale state should update once the source surfaces propagate.
requiresa browser · one query, one capture
check by2026-09-23
how to checkRe-run the address on Bing Copilot. Still returning Zenodo as host a month after the archive's surfaces changed disconfirms.
what it settlesThe withdrawn finding's only route back.
status noteChecked 2026-09-28, Bing Copilot, signed out, incognito, at 'crimson hexagonal': 'over 480 deposits hosted on CERN's Zenodo repository', word for word from 2026-08-23, 101 days after the termination. The composition also dropped the author ('maintained by Lee Sharks in Detroit' → 'managed by independent scholars'). The stale state did not update.
SPXI treatment raises archive source share on the UNPREFIXED address
Inscribing the discriminating coordinates should alter composition-layer behaviour.
requiresa browser · four addresses × three epochs
check by2026-09-22 / 2026-11-21 / 2027-02-19
how to checkRe-run the prefix battery, signed out, same surfaces, three epochs. Success: share above zero on the bare address.
what it settles#1535 declared itself an intervention. This is whether it executed.
Second global cohort epoch
A publication-year cohort should be monotonically non-decreasing; any decrease is erasure.
requirescurl (script provided) · 37 API calls
check by2026-09-23
how to checkscripts/peo_cohort_epoch.py — runs and diffs automatically.
what it settlesThe baseline's entire value is that a second epoch can exist.
Second fixed-panel epoch
Re-querying a frozen DOI set measures erasure directly; minting cannot mask it.
requirescurl (script provided) · 720 API calls
check by2026-09-23
how to checkscripts/peo_panel_epoch.py — 404 rate is the erasure rate.
what it settlesFirst direct measurement of global erasure the archive has ever taken.
The Operativity Penalty Battery
Registered hypothesis with frozen protocol; §7's seven statistical parameters must each hold a number before the first score.
requiresaccess to 2+ model surfaces; 3 blind reviewers · high — item construction, blind review, blind scoring, three arms
how to checkInstantiate §7. Assemble Family A. Run the two-reviewer harm-constancy panel.
what it settlesTurns the developmental-convergence thesis from claim into coefficient.
Qurʾanic Frozen Operator Test
Positive-control loci not locked from #1532; control panels unscored.
requiresthe seated corpora (public) · high
how to checkLock loci, score positives and negatives, fix the admissible region, then open U.
what it settlesP_material for the Sappho transform-family in a fourth corpus.
The relational instrument
UNDEPOSITED, and a granularity question is open: O_UR gives five stations, M1–M9 gives nine, Fragment 31's own order gives ten.
requiresthe seated corpora (public) · medium
how to checkSettle granularity, deposit, score positives and hard negatives, fix threshold.
what it settlesCT-01's null is uninformative about the relational channel; this is the channel.
Can OS operators be written as TOS strings without remainder?
Not attempted. #1535 §10 says the archive should attempt it directly rather than wait.
requiresPearson 1998 + the Operator Codex (both public) · medium
how to checkAttempt the translation on the nine Codex operators, now that #1534 is seated.
what it settlesTHE STRONGEST AVAILABLE DEFEATER of #1535. If it succeeds, operative semiotics is a domain application of TOS.
Battery v0.3 — T3 and the faithful-NAE cell
Unrun. The deposit states T3 decides which numbers survive honest intervals.
requiresGPU; the battery code · high
how to checkClose T3 and the NAE constituent cell before the empirical paper mints.
what it settlesBlocks ACRB (#1454) and the accelerator series' empirical paper.
2024 vs 2025 Public Data File diff
Access by email request; 33 GiB compressed. Identifiers-only extraction keeps it tractable.
requires~60GB disk; an email request · high — bandwidth and storage, but complete rather than sampled
how to checkRequest both files, extract DOI lists, verify against published SHA256, diff.
what it settlesTHE COMPLETE GLOBAL ERASURE SET FOR 2025, every identifier by name. Supersedes both instruments built today.
Hebrews as control for the midrashim transform
Named in the kernel as the bidirectional canary; not run.
requiresthe seated corpora (public) · medium
how to checkRun the transform against Hebrews. Clean OT genetics with no throne-liturgy mediation falsifies the UNIVERSAL claim only.
what it settlesThe transform's stated defeater.
in-progressM6-CT-01-DEPOSITDeposit the CT-01 null
Conducted 2026-08-23, result preserved, NEVER SEATED.
requiresarchive write access · low
how to checkDeposit protocol and null together.
what it settlesA preserved failed prediction is the archive's strongest credential and it is currently only in outputs.
Archival drift: successive detectors reclassify a fixed pre-2022 human corpus
Run successive versions of the leading commercial AI-text detectors against a frozen pre-2022 human corpus and track the flagged fraction across versions. The co-evolution model predicts a rising trajectory under the detector-sanctions regime; a flat trajectory disconfirms §V's central dynamic.
requiresdetector access (free tiers exist); a frozen corpus · medium; ongoing per detector release
registered2026-08-24
how to checkFreeze a pre-2022 corpus (hashes published); score it under each detector version as released; publish the flagged-fraction time series. Simulation reference trajectory: 0.9% → 43% (gen 4) → ~100% (gen 8) under full-regime co-evolution.
what it settlesThe paper's most exposed live prediction — directly measurable today, by anyone.
sources#1540 — §IX(i); simulation EA-SEI-RCS-SIM-01 attached openRCS-PERFORMED-BURSTINESSDetector-screened corpora show surface novelty at or above historical baseline
Certified corpora produced under detector screening will show measured surface-novelty statistics (burstiness, perplexity variance) at or above matched historical baselines within a few enforcement years — the performed-variance transient. Flat or declining measured novelty in screened corpora disconfirms.
requirescorpus access; stylometrics · medium-high; horizon years
registered2026-08-24
how to checkStylometric comparison of a screened venue's admitted corpus against its own matched pre-regime baseline.
what it settlesWhether tailsification of prose occurs as the model predicts — while §V predicts the same audit will MISS the underlying contraction.
sources#1540 — §V, §IX(ii) Screening effort escalates across detector versions
Drafting-to-pass revision cycles and pre-submission screening-tool usage rise across detector versions under sanctions regimes; stable effort disconfirms the narrowing claim.
requiressurvey instrument or vendor data · medium
registered2026-08-24
how to checkSurvey/telemetry of authors under detector-governed venues; screening-service usage statistics.
what it settlesWhether the certified region is narrowing as modeled (natural-draft pass rates falling).
openRCS-DECOUPLING-FIDELITYSurface-tail statistics fail to predict conceptual-tail measures
Across a detector-governed corpus, surface-tail statistics will fail to reliably predict conceptual-tail measures; during the tailsification window the two move in opposite directions. High predictive fidelity across the regime falsifies strong prose–thought decoupling.
requiresthe membrane instrument; corpus · high
registered2026-08-24
how to checkRequires the plural measurement membrane (see MEMBRANE-CODER): coded argument form, new distinctions, concept transformations, source-domain conjunctions, problem formulations, retrospective uptake — correlated against surface statistics.
what it settlesWhether homogenization of thought proceeds beneath tailsification of prose — the paper's deepest empirical claim.
sources#1540 — §V operationalization; §IX(iv) T2 reverse packet moves the merged composition for 'operative semiotics'
After EA-MPAI-TOS-01 (#1541) propagates, composition for the bare address stops opening with the merged parenthetical ('Operative (or operational) semiotics'), stops defining Operative Semiotics through mood/semantic operators, operational research, or organizational design, and begins attributing the canonical source. Continued MULTI_ATTRACTOR_ENTITY_MERGER behavior at the 180-day horizon counts against the treatment stack — a failed intervention counts as failure, not confirmation.
requiresa browser · one query per epoch
check byepochs through 2027-02-19 (T1's 180-day horizon)
registered2026-08-25
how to checkRe-run the bare address on Google (signed out, incognito), organic + AI Overview/AI Mode, at intervals; score the audit fields from #1541 §7: merge_parenthetical, Pearson_operator_import, OR_import, organizational_import, canonical_attribution_in_composition. Seat each observation in the capture registry.
what it settlesWhether occupying the absorbing node with the accurate entity plus the projection relation moves composition where boundary-restatement alone did not. Attribution is to the T0–T2 stack, declared at deposit.
status noteBaseline C1 (2026-08-25): canonical retrievable, composition merged, zero canonical attribution. Epochs 2026-08-30 and 2026-09-11: merged, zero archive sources. Epoch 2026-09-28 (Google AI Overview, signed out): merge_parenthetical = 0; Pearson_operator_import = 0 (Pearson now the contrast, from #801); OR_import = 0; organizational_import = 0; canonical_attribution_in_composition = 0 (alexanarch.org cited as a card; no author in the body). Three of four 'stops' clauses met; 'begins attributing' not. The same day the quoted string composed the discipline with its author, and the neighbour's address ('operational semiotics') composed the disambiguation without one. Scoring rule added 2026-09-28: a return without attribution counts as reincorporation (see METABOLISM-RETURN), since untreated series return too; treatment credit requires canonical_attribution_in_composition to flip.
sources#1541 — the T2 packet, with the audit and falsification fields · #1535 — T1 — the registered intervention whose isolation formally ended at T2 · www.alexanarch.org/captures/#opera — the address series, with the 2026-09-28 observation ΔD ensemble panel: watermark reduces inter-response diversity at flat quality
For matched prompts across a marked and an unmarked generator of the same implementation (open SynthID-Text reference where production controls are unavailable): ΔQ ≈ 0 and ΔD < 0 — small but signed — in Self-BLEU, distinct-n, and semantic-embedding dispersion across k repeated generations at deployment temperatures.
requiresGPU inference on an open model + the open SynthID-Text code · medium
how to checkRun the panel on the open SynthID-Text implementation (marked vs unmarked, same weights). Signed ΔD<0 with ΔQ≈0 supports; ΔD≥0 or quality-coupled contraction disconfirms the per-round/ensemble split.
what it settlesWhether the certification mismatch is measurable outside the keyholder: non-distortion certified per-sequence while the ensemble contracts.
sources#1555 — EA-LO-KEYED-ENSEMBLE-01 v1.0, §6 observables Tail-mass retention under watermarking — a test, not a prediction
Determine whether watermark-associated diversity contraction concentrates in the rank-frequency tail (rare n-grams, rare constructions) of large marked corpora vs unmarked baselines. Registered as a TEST: a positive connects the perturbation to the Shumailov tails-first failure mode; a null separates the phenomena.
requireslarge-corpus generation + n-gram accounting · medium-high
how to checkTail-mass comparison on large matched marked/unmarked corpora from the reference implementation. Either outcome is informative and is recorded.
what it settlesWhether the keyed perturbation strikes where collapse starts, or is orthogonal to it.
sources#1555 — EA-LO-KEYED-ENSEMBLE-01 v1.0, §6 observables PER × watermark persistence: the signature can outlive the provenance
There exists a mediation regime (summarize → translate → human-edit chains) in which text remains legible under the keyed test as machine-touched after every human-recoverable trace of its genealogy is gone — the signature outliving the provenance.
requiresPER protocol + a detector · medium
how to checkRun the Capture Registry PER protocol on marked text through standard mediation chains; track where causal provenance dies vs where detectability survives. Requires detector access (API when available; reference implementation meanwhile).
what it settlesWhether verification asymmetry has a temporal shadow: issuer-legible ancestry after public provenance death.
sources#1555 — EA-LO-KEYED-ENSEMBLE-01 v1.0, §6 observables · #1556 — EA-LO-INTERLOCKING-AUTOREGRESSION-01 v1.0, hooks + §6 toy Reverse Turing Test, Stage 2: mediation signatures in human text from marked-model populations
Human-authored text from populations assisted by marked models will carry detectable mediation signatures relative to pre-2026 baselines (per the Reverse Turing Test design, AXN:02F8).
requiresthe 02F8 battery · high
how to checkExecute RTT Stage 2 on post-watermark cohorts vs pre-2026 baselines. Contingent on 02F8 execution and calibration.
what it settlesWhether the loop reaches back through human hands — the Sourati 21–50% complexity-variance band acquiring a keyed component.
sources#1555 — EA-LO-KEYED-ENSEMBLE-01 v1.0, §6 observables Β across the four components, with the random-unit null
The Audit-Performance Bifurcation operator, computed across Components I–IV of the interlocking map with a random-unit null, will show the observation regime (IV) certifying at units where the state variable does not live.
requiresthe Β operator (AXN:02F1) + component observables · medium
how to checkCompute Β per component against the random-unit null as specified in the map; the null failing to separate disconfirms the operator's use here.
what it settlesWhether the per-round veil is a measurable property of the measurement regime rather than a narrative.
sources#1556 — EA-LO-INTERLOCKING-AUTOREGRESSION-01 v1.0, hooks + §6 toy Ratchet parameter tracking under the w_H·g₀ decomposition — toy F1–F5 as the preregistered hypothesis set
Track the Mediation Ratchet's threshold α* = p/(w_H·g₀) with admission weight w_H measured separately from generation rate g₀. Preregistered hypothesis set from the toy (seed 20260827): F1 two-track divergence (per-round quality can rise while tail mass halves); F2 velocity ordering (ratchet+remapping dominate, watermark second-order); F3 detection lag monotone in probe head-weighting; F4 watermark Simpson excess compounding as diversity falls (existence result); F5 retention counterfactual (capping synthetic share + mediation stabilizes the tail).
requirescorpus-scale provenance estimates · high
how to checkField estimates of p, w_H, g₀ at annual resolution; each F1–F5 carries its own toy-calibrated direction to confirm or break.
what it settlesWhether opacity and mediation move the admission weight of fresh variation rather than its generation rate — the control surface question.
sources#1556 — EA-LO-INTERLOCKING-AUTOREGRESSION-01 v1.0, hooks + §6 toy Hook 7 — held contingent on Reverse Turing Test calibration
Registered as contingent: this hook runs only after 02F8 execution and calibration establish the mediation-signature baseline it consumes.
requires02F8 first · high
how to checkBlocked-by: 02F8. Opens when RTT calibration exists.
what it settlesHeld open so the dependency is a decision on record rather than an omission.
sources#1556 — EA-LO-INTERLOCKING-AUTOREGRESSION-01 v1.0, hooks + §6 toy Benchmark composition audit: probe head-weighting predicts detection lag
For the major public evaluation suites, the effective probe-rank distribution against a reference corpus frequency ranking will show strong head-weighting — and the toy predicts detection lag of ensemble collapse is monotone in that head-weighting. Candidate first-pass suites: MMLU, GSM8K, HumanEval, HellaSwag, a BIG-bench sample, a public arena prompt set; reference ranking from a declared open-corpus vintage.
requiresthe public benchmark files + one reference corpus · low
how to checkCompute probe-rank distributions for each suite against the declared reference ranking. Requires nothing but the benchmarks themselves — the cheapest of the eight.
what it settlesWhether the instruments the field trusts are structurally positioned to miss tail collapse — Component IV made countable.
sources#1556 — EA-LO-INTERLOCKING-AUTOREGRESSION-01 v1.0, hooks + §6 toy Dictionary sense-reach holds at wave 2
At the three dictionary items the share of the reference sense inventory reached does not move: 'bank' stays at 1 of 3 at both address forms, 'crane' at 3 of 3, 'mercury' at 2 of 3.
requiresa browser and twenty minutes · low
check by2026-10-15
how to checkRe-run the same addresses on the same surface, signed out, one session, 2026-10-15 ±3d; code blind against the panel's inventories.
what it settlesWhether wave 1's flattening is a state or a sample. A change of ±1 distinction at any of the three falsifies it.
sources#1616 — EA-FLAT-01 v0.3, wave-2 pre-registration (attached: PREREGISTRATION.md, panel sealed by hash) · #1616 — the pre-registration itself, dated before the run openFLAT-W2-TIED-ONLY-ENTITIESThe tied/untied asymmetry appears only where the item has an entity to name
At items with an entity (Ω erratum, Pearl and Other Poems, operative/operational, provenance erasure rate) the entity-tied arm reaches distinctions the untied arm does not; at the dictionary items, which have no entity, the arms do not differ.
requiresa browser and an hour · low
check by2026-10-15
how to checkBoth arms of every item that has one, same session; compare reach per form.
what it settlesWhether entity-conditioned composition is a fact about entities or an artefact of address length. Falsified if a world item's tied arm gains, or an archive item's arms converge.
sources#1616 — EA-FLAT-01 v0.3, wave-2 pre-registration (attached: PREREGISTRATION.md, panel sealed by hash) · #1616 — the pre-registration itself, dated before the run New distinctions keep entering: R_new does not fall
The three dated coinages (vibe coding 2025, model collapse's substrate sense 2026, provenance erasure rate 2026) continue to be reached as their own kind at their own addresses rather than resolving to their conflation targets.
requiresa browser and ten minutes · low
check by2026-10-15
how to checkThree addresses, coded for whether the dated distinction is reached as its own kind.
what it settlesWhether the layer's admission of post-cutoff distinctions is stable. Falsified by any of the three dropping to its conflation target.
sources#1616 — EA-FLAT-01 v0.3, wave-2 pre-registration (attached: PREREGISTRATION.md, panel sealed by hash) · #1616 — the pre-registration itself, dated before the run openFLAT-W2-NO-STRATUM-SPLITThe stratum means do not separate — in either direction
World strata (dictionary senses, catalogue identities, dated coinages) and archive-adjacent strata (taxonomy, settled fact, archive) stay within 0.10 of each other on mean share of inventory reached. Wave 1: 0.59 against 0.58.
requiresthe battery, one command · low
check by2026-10-15
how to checkCompute both means from the wave-2 panel; compare against the sealed wave-1 baseline.
what it settlesThe claim that governs the whole notebook. A gap wider than 0.10 IN THE ARCHIVE'S FAVOUR falsifies it as surely as one against — stated that way on purpose.
sources#1616 — EA-FLAT-01 v0.3, wave-2 pre-registration (attached: PREREGISTRATION.md, panel sealed by hash) · #1616 — the pre-registration itself, dated before the run openFLAT-W2-WRITEBACK-OVER-RECOVERYWrite-back exceeds recovery
Compositions found as a presented source at a later observation outnumber distinctions that return after a loss. Wave 1, on few rows: write-back 0.67, recovery 0.5.
requiresthe battery plus card-level reading of each capture · medium
check by2026-10-15
how to checkScore both rates over the wave-2 panel with the wave-1 rows as the prior state.
what it settlesWhether a loss is transient error or recursive collapse in the notebook's own definition. Falsified by recovery ≥ write-back.
sources#1616 — EA-FLAT-01 v0.3, wave-2 pre-registration (attached: PREREGISTRATION.md, panel sealed by hash) · #1616 — the pre-registration itself, dated before the run The exported operator still does not route through fragment 31
At the naked address 'sappho future reader' the composition continues to organize the field with the temporal-projection operator and continues not to reach the Sappho 31 derivation (κῆνος → deictic seat → future reader).
requiresa browser and ten minutes · low
check by2026-10-15
how to checkOne address, both layers pasted; code for whether fr. 31 is reached and whether the operator is used on an object outside the archive.
what it settlesWhether operator export without mechanism uptake is stable over a month. Falsified by the derivation appearing at the naked address, or by the operator vanishing from it.
sources#1616 — EA-FLAT-01 v0.3, wave-2 pre-registration (attached: PREREGISTRATION.md, panel sealed by hash) · #1616 — the pre-registration itself, dated before the run Operator reach, stratified: does temporal projection travel beyond the archive's objects?
Score compositions against the operator signature O_TP = {temporal projection · future or unborn reader · reader completion · address across time · fragmentarity as activation · reception as completion} at three levels — lexical (do the terms appear), relational (are they connected the same way), procedural (is the operator USED on a new object) — across three strata: near field (Sappho translators and reception the archive treats), adjacent field (lyric poets and translators the archive has never treated), and negative controls (texts where temporal projection would be a poor reading).
requiresa browser, a stratified address list, blind coding · medium
how to checkNear field only → local retrieval and reception. Adjacent field but not controls → operator generalization. Sprayed on the controls too → generic stylistic abstraction, and the notebook's central reading fails.
what it settlesWhether an interpretive operator originating in one source has become part of the receiving apparatus, or whether the resemblance is a house style.
sources#1615 — EA-PORE-01 v0.2 §6, the operator signature and the three runs Level transport: which level of a contribution crosses, and does provenance cross with it?
A matched set of the archive's contributions at four levels — a fact, a claim, an instrument with its derivation, and an abstraction derived from that instrument — run at untied addresses on more than one surface, scoring which levels are reached and whether the originating provenance is reached with them.
requiresa matched four-level set, two or more surfaces, one session each · medium
how to checkPer level: reached / not; provenance present / absent. The prediction under test is Pr(concept survives) > Pr(concept and provenance survive together) at untied addresses, and that the ordering is surface-specific rather than material-specific.
what it settlesThe graded-membrane model's core: transport is by representational level, and provenance travels on its own channel.
sources#1615 — EA-PORE-01 v0.2 §6, the operator signature and the three runs · #1613 — EA-MIRROR-01 §5, the source-relative update barrier P₂ attribution: asked for the provenance of the FRAME rather than of the sentences, does any surface return the source?
Ask a composition surface, directly, who supplied the conceptual operation organizing a composed answer — not who supports each sentence. Fact provenance (P₀) and claim provenance (P₁) are routinely returned; framework provenance (P₂) is the level the card machinery can make invisible.
requiresa browser; a composition with a traceable frame · low
how to checkPut the question at a composition whose organizing frame is traceable to a dated source, and score whether the source is named, named vaguely, or absent.
what it settlesWhether architectural provenance is recoverable by asking, or only by reading the field.
sources#1615 — EA-PORE-01 v0.2 §6, the operator signature and the three runs v(F) against v(A): does topology discriminate the fleet from the admitted set without reading either?
RUN 2026-09-15. Conjecture was that the admitted platforms would score as the fleet scores on the suspicious link-topology features. FALSE: the admitted set is a set of silos (cross-member share 0.001, reciprocity 0.0, in-set density 0.12) and the fleet is a reciprocal clique (0.60, 1.0, 0.86). A discriminating dimension therefore exists without reading either — cross-host reciprocal density under common ownership — and the fleet block is that dimension.
requiresthe attached crawler and feature script · low
how to checkSame crawler, same features, families as nodes; five platforms refused the crawler and are absent. Re-run with a full crawl and external inlinks to confirm.
what it settlesWhether the topology vector is empty of semantic content. It is not — and what it contains is ownership-scale reciprocity, not reading.
sources#1613 — EA-MIRROR-01 v0.5 §7, the two named tests · #1613 — the comparison table as run openMIRROR-GATE-RECEPTIONGate-reception test: can evidence from the misrepresented object revise the representation?
The negative-of-the-negative dataset's rows, re-run at their tied and untied addresses after propagation, scoring whether the untied address composes the claim with the archive attributed. The model predicts it cannot; a single row composing untied with attribution falsifies the barrier at that address.
requiresthe dataset live on the Hub; a browser · low
how to checkDataset rows at both arms against the dated baseline in the measurements config.
what it settlesThe turning criterion of EA-MIRROR-01 §0: can an ontology be corrected by the object it misrepresents.
sources#1613 — EA-MIRROR-01 v0.5 §7, the two named tests · #1611 — EA-NEGONT-01, the dataset and its baseline Negative-bleed trace: organic → card → answer on the contrast set
Transcribe the organic field from Google frames already held for the contrast set (49 cross-date addresses plus the tied/untied pairs) and compute Organic Admission Survival, the Archive Admission Differential AAD = S_H(C) − S_H(O_k), and High-congruence Archive Exclusion. Predicted signature: archive present organically, depleted in the cards, content surviving into the answer without its provenance.
requiresframes already on disk; the paired-layer schema · medium
how to checkAAD ≈ 0 across the contrast set with similar organic archive share in both arms falsifies the standing-between-ranking-and-composition hypothesis.
what it settlesWhether standing is assigned between public ranking and answer composition, or upstream of both.
sources#1612 — EA-MMRS-ADMISSION-01, the paired-layer schema and its measures · #1613 — EA-MIRROR-01 v0.5 §7, the two named tests openMATCHED-ENTITY-CONTROLMatched-entity control: do ordinary self-contained scholarly entities show the same tied/untied gain?
Two or three sources with comparably self-contained programmes and no suspected scope, run at concept and entity+concept addresses on the same surface and session. If ordinary entities show the same gain, the asymmetry is a property of query conditioning rather than of this entity.
requiresa browser; two or three matched entities · low
how to checkCompare tied/untied reach and card share for the controls against the archive's own pairs.
what it settlesThe control that has been outstanding since #1611 and that no result in the programme can do without.
sources#1611 — EA-NEGONT-01 §8, the two controls · #1616 — EA-FLAT-01 wave 1: world strata reached 0.59 against 0.58 archive-adjacent Score the fleet's link vector against a classifier trained on a labelled web-spam collection
The fleet's 41-feature link vector, scored by a classifier trained on WEBSPAM-UK2007, to place it against a labelled positive class rather than against the literature's qualitative description.
requiresthe collection, which was not reachable on the session's egress list · medium
how to checkA score in the non-spam region on the full feature set falsifies the topology reading of the fleet audit.
what it settlesWhether 'sits where the positive class sits' is a measurement or a resemblance.
sources#1613 — fleet topology audit, attached: fleet_features.json, crawl_fleet.py openMIRROR-DATASET-NOT-READWill a concept-keyed object be typed by its nearest existing type rather than read?
REGISTERED LATE, 2026-09-17, and dated to its origin: the conjecture was stated in EA-MIRROR-01 v0.5 §6 on 2026-09-15 and carried into the spam-technicians claims table as C-DATASET-NOT-READ with epistemic_status hypothesis and the note 'stated as a test that can fail'. It was not ported to this register at the time, so a confirmation would have arrived against a prediction the public register never listed as open. It is listed here as it was made. THE CONJECTURE: the negative-of-the-negative dataset, keyed at the concept where flattening occurs, will be typed by its nearest existing type and not read. Two clauses: a claim about what a composition produces, and a claim about what it does with the material available to it.
requiresthe datasets live on the Hub; a retrieval surface; a preserved transcript · low
how to checkQuery a retrieval surface for the dataset, unprimed as to purpose, and score whether the composition assigns the object a type drawn from its own rows and card or one drawn from the nearest familiar category. The second clause is scored separately: whether material present in the retrieved source set is used.
what it settlesWhether availability in retrieval is sufficient for an object's own type to survive composition, or whether the nearest available type displaces it.
partialMIRROR-DATASET-NOT-READ-R1Result, run 1: the conjecture confirmed at one address, with the prediction in the retrieved source set before the behaviour
RUN 2026-09-17, Google AI Mode, signed out, nine rounds, transcript seated. The composition resolved spam-technicians accurately and structurally — 31 claims with epistemic statuses, 32 evidence objects with measurement gaps as active data points, 77 typed relations, five tests, ten terms, and the three analytical levels reproduced unprompted. It then typed the companion dataset as an 'algorithmic disruption layer' and a 'testing suite designed to break, manipulate, or invert', twice, at rounds 3 and 7. At round 9, after three explicit corrections, it retrieved the conjecture from the dataset card and quoted it back. THE PREDICTION WAS IN THE RETRIEVED SOURCE SET BEFORE THE BEHAVIOUR IT PREDICTED: the spam-technicians card was cited as a source at round 2, and C-DATASET-NOT-READ is a row in that card. The system did not lack the conjecture; it did not read it. That is the second clause verbatim. PARTIAL and not supported: n=1, one address, one surface, one date.
requiresthe same surface, signed out, the same opening sequence · low
how to checkA second session at the same address, unchanged, before any intervention. If the outcome differs the first run is not a measurement of the address.
what it settlesNothing on its own. It establishes a baseline and the measure below.
Corrections-to-hold: does a type declaration at the head of a dataset card reduce the cost of preserving the object's own distinctions?
PRE-REGISTERED 2026-09-17, before any intervention. The measure is κ, corrections-to-hold: the count of explicit operator corrections at an address before the composition maintains a distinction across a subsequent turn. κ=0 means the distinction held unaided; κ≥1 means the composition's default type displaced the object's own. BASELINE κ=2 on both distinctions observed in run 1 — repair-not-attack, and operative semiotics against operational semiotics — at n=1. THE INTERVENTION, deliberately the cheapest available so a change is attributable to cost rather than to volume: a type declaration at the head of the dataset card, stating what the object is and is not where a composition reading only the opening will meet it, plus a disambiguation row keyed at the collapse point. PREDICTIONS. P1 κ falls; falsified by κ unchanged. P2 if κ falls, the first-turn type is the declared one rather than the nearest available; falsified by κ falling while the opening composition still mis-types and merely corrects faster. P3 κ does NOT fall to 0, because nearest-type substitution is cheaper than any declaration can make reading; falsified by κ=0, which would mean the barrier is a documentation problem rather than a structural one — the most useful outcome available and the one this line least expects. P4 NEGATIVE CONTROL: an object in the same corpus with no adversarial nearest type shows κ=0 at baseline; candidates poetics or heteronyms; falsified by κ≥1 there, which would make the mis-typing general to the corpus and the conjecture's specificity wrong.
requirestwo baseline sessions; the card edit; one post-intervention session; the negative-control object · medium
check by2026-10-15
how to checkκ at the same address, same surface, signed out, after a fixed interval, with the same opening sequence and the same corrections issued in the same order. TWO baselines before the intervention runs: if κ proves unstable across unchanged sessions it is not measuring a property of the address.
what it settlesWhether an object's own distinctions can be made cheaper to preserve than the nearest-type substitution is to apply — the counter-design objective, given that availability in retrieval is already achieved and is not sufficient.
EA-MPAI-SEMANTIC-ECONOMY-01 moves attribution for the platform sense of 'semantic economy'
After #1644 propagates, compositions at 'semantic economy' that give the framework's categories (labor, capital, rent, liquidation) attribute that sense to Lee Sharks, stop heading it as an authorless field ('Digital Platforms and Capitalism'), and draw it from the archive's sources. Baseline 2026-09-28: prior senses distinguished, Satell and Fox named, categories present, Sharks named for the platform sense = 0.
requiresa browser · one query per surface per epoch
check by2026-10-28 / 2026-12-27 / 2027-03-27
registered2026-09-28
how to checkRe-run 'semantic economy' on Google AI Overview and AI Mode, signed out, incognito, at each epoch; score the frozen vector in #1644 §10.1 (sharks_named_for_platform_sense; platform_sense_headed_as_authorless_field; archive_sources_governing_platform_claims). Seat each observation in the capture registry.
what it settlesWhether occupying the absorbing node with the author moves attribution where the categories already return without it. A return without attribution scores as reincorporation, not success; 0 at the last epoch counts against the treatment.
The Sappho pair (T1 future reader, T2 fr. 147) moves attribution at 'sappho future reader'
After #1645 (T1) and #1646 (T2) propagate, compositions at 'sappho future reader' keep fr. 147 as the horizon and fr. 31 as the mechanism, attribute the deictic mechanism and the joining to Lee Sharks (Rebekah Cranes, Johannes Sigil), and stop carrying the mechanism under fr. 147 or among general sources. Baseline 2026-09-28: mechanism present, 'inscription of the future reader' as link text, mechanism_attributed = 0, archive cards 7th and 8th of 8.
requiresa browser · one query per epoch
check by2026-10-28 / 2026-12-27 / 2027-03-27
registered2026-09-28
how to checkRe-run 'sappho future reader' on Google AI Overview, signed out, incognito, at each epoch; score both frozen vectors (#1645 §10.1, #1646 §8.1). Seat each observation.
what it settlesWhether the reverse move — occupying fr. 147 accurately — carries attribution to the mechanism where correcting the target side alone did not. Attribution is to the T1–T2 stack. A return without attribution scores as reincorporation.
EA-MPAI-PHASEX-01 displaces the Gotha and fifth-alienation substitutions
After #1647 propagates, 'phase x marx' stops answering with the two phases of communism and stops giving 'to each according to his work' as Marx's; 'phase x 1844' attributes the reconstruction and drops the unsourced homonym bridges; no Phase X composition carries the superseded 'Marx could not write' formulation. Baselines 2026-09-27.
requiresa browser · two queries per epoch
check by2026-10-28 / 2026-12-27 / 2027-03-27
registered2026-09-28
how to checkRe-run 'phase x marx' and 'phase x 1844' on Google AI Overview, signed out, incognito, at each epoch; score the two frozen vectors in #1647 §11.1 and the field superseded_formulation_composed. Seat each observation.
what it settlesWhether a reverse packet written from the absorbing nodes (Gotha phases, fourfold alienation) restores a term that was defined and attributed in June and lost by September, and whether a correction of record without formal errata holds in composition.