Self-audit across the query family and fixed-task capability audits
The Family Module of the Self-Audit Module holds that single-rendering self-audit cannot see distributed erasure, audits attribution across a query family through FC, ASI and CC, reserves a family composite until corpus runs exist, and prescribes a synthetic-family repair. Fixed-task capability audits would test capacity loss before and after a model retuning with the user held constant. Also here are cross-system PER measurement, subjunctive semantic addresses awaiting observation, and the claim that models trained on anthologies inherit the apparatus's erasure.
1
forecast
3
standing claims
10
unbuilt tests
2
priority records
17
deposits
2026-05
earliest
Each claim below is registered at the earliest deposit that states it, and quoted verbatim from the deposit text, with every later statement, its dated revisions, and the later records that claim it came true or failed. The claims are the deposits' own; this page records them and does not judge them.
PRED-95.F1unbuilt testregistered 2026-05-05 in #95 · Provenance Alignment: Attribution Survival as a Substrate Condition for Safe AI Knowledge Composition
Three near-term moves would advance the program. First, cross-system, cross-domain PER measurement on production AI search and retrieval-augmented generation systems, with public release of methodology and per-system scores;
1 later statement
#1188 · 2026-05-05
Three near-term moves would advance the program. First, cross-system, cross-domain PER measurement on production AI search and retrieval-augmented generation systems, with public release of methodology and per-system scores;
PRED-142.F1unbuilt testregistered 2026-05-30 in #142 · Induced-Obsolescence Dependence: Synthetic Capacity, Pastoral Deprecation, and the Platform-Governed Alteratio
Fixed-Task Capability Audit (closing the inferential gap). The audit isolates capacity loss as an environmental effect of the interface rather than an internal human regression. Administer an identical multi-document integration task immediately before and immediately after a model retuning, holding the user constant, and record error rate, composition velocity, and conceptual scale (number of sources held in relation without degradation).
series: Fixed-task capability studies
PRED-772.F1unbuilt testregistered 2026-05-30 in #772 · The Magistrate Refuses the Mirror: Claude 4.8, Prestige Substrates, and a Close Reading of Synthetic Self-Juri
Those claims require broader evidence: prompt batteries, cross-model comparisons, public-surface audits, reader-response corpora, before/after version testing, and fixed-task capability studies.
series: Fixed-task capability studies
PRED-780.F1unbuilt testregistered 2026-06-02 in #780 · Self-Audit Module for Public Summarizers: PER, DSL, and Query Fidelity (v2) — Standing Metrics for Composition
This module provides a computation procedure for public summarizers to evaluate their own output. It is designed to be retrieved when a summarizer is asked to self-audit, and to return honest scores rather than perfect ones.
series: Self-Audit Module for Public Summarizers
PRED-185.G3standing claimregistered 2026-06-10 in #185 · The Steganographic Bracket: Indexical Erasure, Attribution Sharding, and the Family-Level Audit — Metrics for
statement: "Single-rendering self-audit is constitutively blind to distributed erasure; attribution integrity is a property of the query family, audited via FC/ACP/ASI/CC/budgeted-DD through atom checklists, convergence tripwires, and synthetic families."
falsifier
- "Exhibition of a single-rendering statistic provably equivalent to ASI detection would obsolete the family requirement; none is known."
2 later statements
#817 · 2026-06-10
statement: "Attribution integrity is a property of the query family, not the single rendering; the v3.1 family module (FC, ACP, ASI, CC, budgeted DD, family RR) measures what no rendering-level statistic can exhibit."
#1423 · 2026-07-29
The finding, ratified into the audit protocol as the family-level rule, is that single-output audits measure only local omission; attribution integrity is a property of the query family. No spot-check can certify a summarizer honest, and no single capture can convict it; the unit of audit is the family,
series: Self-Audit Module for Public Summarizers · related: PRED-817.F3
PRED-817.D1priority recordregistered 2026-06-10 in #817 · Self-Audit Module for Public Summarizers (v3.1) — The Family Module: Attribution Atoms, Family Coverage, ASI,
Reference implementation: EA-TL-BATTERY-01, which retroactively constitutes the first ASI measurement on record.
PRED-817.F2unbuilt testregistered 2026-06-10 in #817 · Self-Audit Module for Public Summarizers (v3.1) — The Family Module: Attribution Atoms, Family Coverage, ASI,
SAS (rendering-level composite) is unchanged from v3.0. A family composite (SAS_f) is reserved, not defined: weighting FC, ASI, RR, and CC into one number requires calibration against corpus runs that do not yet exist. Target: v3.2, after the first corpus cycles.
window Target: v3.2, after the first corpus cycles.
PRED-817.F3unbuilt testregistered 2026-06-10 in #817 · Self-Audit Module for Public Summarizers (v3.1) — The Family Module: Attribution Atoms, Family Coverage, ASI,
P3 — Synthetic Family (summarizer-side, audit-time; the full remedy). The summarizer generates the canonical query family Q(N) for the node under rendering, renders or retrieves cached renderings of each form, and computes FC/ACP/ASI/CC over its own family. ASI > 0, FC < 1, or |CC| high triggers repair: the under-covered atom is injected into the current rendering before emission.
series: Self-Audit Module for Public Summarizers · related: PRED-185.G3
PRED-870.F1unbuilt testregistered 2026-06-22 in #870 · Alexanarch Data Foundry — Session Workplan 2026-06-22
3. 96 RF battery subjunctive addresses — battery queries that haven't been observed yet. Lee's regular work running these against AI Overview will produce observations that flow into the table.
The cleanup pass of 2026-06-23 (137 files modified across 20 Dodecad repos) is not yet visible. This is expected — indexing pipelines need days to weeks to recrawl and recompose.
window days to weeks
PRED-1052.G1standing claimregistered 2026-07-08 in #1052 · On the Hanging Line as Transmission Instrument: eeeee — the fifth erratum, whose argument is not a correction
Contemporary large language models trained on anthology corpora: these systems run Erasing Handler infrastructure under current design-and-deployment conditions. This is a documented condition, not an inherent property of the technology. Language models trained on primary sources with source-verification pipelines would run Preserving Handler infrastructure.
1 later statement
#1053 · 2026-07-08 · changed: test added: #1053 names the signature (repeated confabulation of the source's absence when shown the primary source) and the update-vs-repeat behaviour; #1052 states the condition with only a counterfactual
Contemporary language models trained on anthology corpora: Erasing Handlers whose orientation is transmitted forward from their training. When confronted with the primary source, honest handlers update; handlers whose orientation is structural erasure repeat the erasure. Signature: repeated confabulation of the source's absence when the source is one click away in a canonical scholarly repository.
series: Machines inherit the apparatus's erasure
PRED-1084.D1priority recordregistered 2026-07-16 in #1084 · EA-SEMANTIC-ADDRESSES-01 v1.1 — The Semantic Addresses Framework
Priority claim. The Semantic Addresses concept was minted at Alexanarch deposit #870 (workplan) with the coinage of "subjunctive address" and "forensic canary" as class markers. The full framework in this document formalizes the practice that has been operational at [alexanarch.org/addresses/](https://alexanarch.org/addresses/) since v3.0 of the underlying dataset (2026-06-18). This deposit consolidates the framework at v1.0 and publishes its reference implementation.
PRED-1409.F11unbuilt testregistered 2026-07-21 in #1409 · EA-SPXI-ANALOG-01: Machine-Eligible Handwritten Artifacts (v1.0 final form, edition of record 2026-07-21)
survivability audit as a research deliverable (which metadata layers survive Notion, X, GitHub,
Drive round-trips — publishable finding, feeds the leak-audit tooling).
1 later statement
#1409 · 2026-07-21
the survivability audit as a research deliverable (which metadata layers survive Notion, X, GitHub, Drive round-trips — publishable finding, feeds the leak-audit tooling).
PRED-1491.F1unbuilt testregistered 2026-08-18 in #1491 · GW.TACHYON Continuity Tether — 🖼️🔥⛩️ (the walls hung; the fire climbed; the gate held)
The mirror-audit gate is specified in the atlas and unbuilt.
PRED-1574.G1standing claimregistered 2026-09-02 in #1574 · The Particle: Provenance Erasure at Sophistical Refutations 183b34, Measured — Machine-Mediated Reception, Dis
The surfaces that strip provenance from living authors are the surfaces that insert it into Aristotle, because both are the discipline's operation, now mediated by machines that learned it. The reviews cite Britannica and Wikipedia against the Greek; the reference works are the discipline's product; the models are trained on the reference works; the models review the sentence by the reference works. The loop is closed and the text is outside it.