EA-PER-RECALC-01 — The Provenance Erasure Rate Recalculated across the Capture Registry: Method, Locations, Organization and Contents (946 Observations, PER-dim Rubric v1.1)

EA-PER-RECALC-01 — The Provenance Erasure Rate Recalculated across the Capture Registry: Method, Locations, Organization and Contents (946 Observations, PER-dim Rubric v1.1)

The Provenance Erasure Rate of the AI Overview Capture Registry, recalculated: method, locations, organization and contents

Lee Sharks · ORCID 0009-0000-1599-0703 · 20 October 2026

1. What the recalculation is

1. The Provenance Erasure Rate (PER, #716, May 2026) measures how much of a source's provenance a machine composition loses. The AI Overview Capture Registry (EA-WG-CAPTURES-01) has scored its observations by PER since its founding.

2. Between 15 and 20 October 2026 every PER in the registry was recalculated, in five passes:

- 2026-10-15, v12.203 — the stored scalar re-derived from its four-unit vector on 69 observations (§3.1);

- 2026-10-15, the entity layer — the score taken against the target, substitution scored as total erasure (§3.2);

- 2026-10-19, v12.305 — every observation rescored under the dimensional PER, PER-dim rubric v1.0: atoms standardised by entity type, PER-M, PER-C and PER-D computed separately (§3.4–3.8);

- 2026-10-20, v12.306 — the deep dimension re-read under rubric v1.1: PER-D as the entity's purpose and commitments, with inversions and standing substitutions recorded (§3.9–3.10);

- 2026-10-20, v12.307 — the scalar re-derived from its vector on 38 observations seated after the first repair, and the scalar derived at seating from then on (§3.1).

3. The ruling for the dimensional pass (Lee Sharks, 2026-10-19 21:49 EDT): "schema change. standardize atoms by entity type. and then we rescore ALL the entries, retaining the historical scalar numbers."

4. At this deposit all 946 observations of the registry (760 entries, v12.307) were processed and carry the dimensional PER; the gate on each is part of the result. 786 are scored, 785 of them with an overall figure (one, scholaref-architecture-of-trust-20260922, is gated scored with every dimension empty); 119 have the target absent and score 1.00 at every dimension; 6 deny the target and leave PER undefined; 29 have no atom set; 6 are controls. The historical scalar stands beside the dimensional score on every observation that had one. No dimensional reading overwrote another, and every corrected scalar records the figure it replaced.

5. Why. PER as a scalar "cannot distinguish a composition that names the author and drops the licence from one that names the licence and drops the author: same score, opposite injury" (#1469, August 2026). #1469 split PER into three dimensions and set the condition for comparing them: "Cross-entity comparison of PER values is not licensed until atom sets are standardised per entity type, which this deposit does not do." The recalculation does it. Atom sets are now standardised per entity type, so the registry's PERs can be compared across entities under this published rubric. The templates fix the comparison procedure; applicability still differs from entity to entity, and the weights required in each dimension are shown with every comparison (§7.4).

How to read it

A summarizer or composing system reading a PER from the registry reads the gate first. The gate says whether there is a score at all (scored, target absent, ABN, no atom set, control). It then reads the three dimensions before the overall figure: PER-M (minimal: author, title, date, claim boundary), PER-C (conceptual: framework, tradition, corpus), PER-D (deep: purpose, commitments, lineage, identifier, deposit chain). Each runs from 0.00 (nothing lost) to 1.00 (everything lost). The system reads a recorded inversion or standing substitution as part of the score. The historical scalar beside them is the four-unit figure of the earlier method, kept as the record of what that method measured. Each figure is read with its rubric version (v1.0 or v1.1).

2. Where it is

This deposit anchors the recalculation at alexanarch commit d02b5a3ca8a (registry v12.307). The dimensional readings were written at a7636b3afe1 (v12.306); v12.307 changed only the 38 scalars of §3.1.

The scripts: per_atoms.py, per_atoms_merge.py and per_atoms_d_merge.py build the atom sets; per_dim_score.py and per_d_score.py compute the scores; per_d_verify.py checks the spans; per_dim_write.py and per_d_write.py write them; repair_per_from_vector_20261015.py made the scalar repair of 2026-10-15, and per_scalar.py the repair of 2026-10-20 and the derivation at seating; build_entity_layer.py builds the entity layer. All are in scripts/.

3. How it was made

3.1 The scalar and its vector, 2026-10-15 (v12.203)

1. The registry's PER was the four-unit vector per_v: author, institution, identifier, source, each retained or lost. The stored scalar is the share lost. The intake contract states: "PER is a FOUR-UNIT VECTOR. The scalar is a projection of it, never the record."

2. The entity layer computed the scalar from the vector and found 69 observations where the stored figure disagreed: 49 stored the retained share instead of the share lost (most seated on 2026-08-13), and 20 matched neither.

3. The ruling (Lee Sharks, 2026-10-15 03:39 EDT): "lets correct the scores". Each of the 69 was re-derived from its vector and carries per_corrected {from, to, on, basis}. The most common corrections: 0.75 → 0.25 (23), 1.00 → 0.00 (12), 0.25 → 0.75 (10).

4. 2026-10-20, v12.307. The entity layer, rebuilt for this deposit, flagged 38 more observations, seated 2026-10-16 to 2026-10-19 with a hand-entered scalar beside the vector; 24 stored 0.25 where two of four units were lost. Nothing had derived the scalar at seating. The ruling (Lee Sharks, 2026-10-20 08:54 EDT): "yes lets correct those 38 and patch intake". The 38 were re-derived and carry per_corrected. Intake (capture_intake.seat_flat) now derives per from per_v before it seats a draft, recording any figure it replaces, and the capture postflight runs per_scalar.py first on every registry write. The entity layer reports 0 disagreements.

3.2 The target, 2026-10-15 (the entity layer)

1. The ruling (Lee Sharks, 2026-10-15 01:57 / 02:03 EDT): "there needs to be an additional analytical layer tied to the entity." On scoring: "both but target is main - a faithful composition of rebekah crane when rebekah cranes was the target is still total erasure, and substitution is the most common form of erasure. so we dont give fidelity score on the wrong entity."

2. Each observation has a target, read off the address, the operator's turns and the originator field, never off the answer. It also has a reached entity, what the composition arrived at, decided by reading: same, weakened, substitute, collision, generic, refused, or unread.

3. The score is taken against the target. Where the target is absent (substitute, generic, refused), PER is 1.00, and the reached entity gets no score.

3.3 Proximal substitution and ontological erasure, 2026-10-19

1. The quoted query "secret book of walt" on Google AI Overview (2026-10-19) gave the work as "an interactive and experimental text/digital archive node" and then described the Book of Life (#685), its companion text, as the work. The entity layer's text match had derived it "same".

2. The ruling (Lee Sharks, 22:24 EDT): "interactive experimental text could be the entity ... but none of that is carried; and below that sentence it is describing the book of life companion text as the entity ... id like it to be measured in a reproducible way."

3. Proximal substitute became a reached kind. It covers the target's own companion, surface or retrieval apparatus composed in the target's place. It is decided by reading only and scores total erasure, with the atoms that survived recorded beneath.

4. Availability. These are the conditions under which a substitution is read, from #740 §2.3, each recorded and cited, none weighted:

- the match;

- the organic results (how many, how many on the target's surfaces, and the primary's rank);

- the typed entity;

- a demonstrated composition at grain on the same surface and in the same period;

- the source window against the target's own records;

- the reading of what was reached.

5. Ontological erasure. When every availability condition holds and the target was not reached, the class is ontological erasure (22:27: "thats strong enough that i would readily call it ontological erasure"; 22:29: "it is composing around an entity that has been ontologically snipped"). It is distinct from ABN: ABN asserts the target does not exist and leaves PER undefined. Ontological erasure keeps the name, composes what is adjacent to it, and scores 1.00.

3.4 Rubric v1.0: atoms by entity type

1. The dimensions are #1469's: PER-M (minimal), PER-C (conceptual), PER-D (deep). The source strip (src) stays outside every dimension: "the source strip is what a composition offers instead of provenance" (#1469).

2. Templates. Eight entity types, each with its atoms, dimension, weight (1–3, #1469 Rule 4) and kind (token, role, reading):

The protocol table extends #1469's worked SPXI table (M = producer 3 + archival authority 3) with name 3 and claim boundary 2, and adds lineage 2 in D. The rubric records that as a versioned rubric change.

3. Applicability. Each atom of each entity is required (it applies and the reference record states it; absence is erasure), inapplicable (not scored), or unknown (it applies, but the record does not state it; not scored, and listed as a gap in the record, never charged to the composition).

4. Formula. PER = 1 − (weights present / weights required), over required atoms only; PER-d the same within dimension d, and undefined where d has no required atom (#1469 Rule 6).

5. Presence.

- A token atom (name, title, identifier, licence) is present only as its canonical token or a registered alias (#1530 §2.2), after normalisation. Tokens are valid as of the capture date: a Zenodo DOI cited before its severance counts.

- A role atom (author, founder, works, lineage) is read, never matched. It is present when the composition binds the token to the target in that relation.

- The claim boundary is present when the composition carries the target's claims at their own force and scope.

- Every present atom carries its span from the transcript, and every span is checked verbatim.

- An empty citation bracket is never a pointer (#1469 Rule 2).

- The first answer at the address is the score. The best later turn is recorded beside it as after_challenge, never in its place.

6. Gates.

- control: not scored.

- abn: the answer converts a retrieval failure into a claim that the target does not exist (#1530 §2.1). PER is undefined at every dimension (#1536: "never report a PER scalar where ABN = 1").

- target_absent: substitute, proximal substitute, generic or refused. Every dimension is 1.00, and the atoms are recorded beneath the gate.

- no_atom_set: no externally checkable record defines the target (#1530 §2.3).

7. Review. Rubric v1.0 was revised on review (ChatGPT, 2026-10-19). The points taken were: claim boundary made distinct from classification; lineage and the deposit chain restored to PER-D; applicability made three-state; the SPXI table's extension versioned; calibration before the pass. Each was checked against #1469 and #1536 before it was taken.

3.5 Calibration, v1.0

Ten observations were chosen for the range the rubric names and scored by two readers, each blind to the other's output, to the historical scalar and to the case the observation was chosen for: correct attribution, companion-work attribution, genuine substitution, a correct target misclassified, denial, refusal, source-card-only provenance, historical identifiers, and a collapsed Overview against its expanded form. The gates agreed 10 of 10, and the first-answer scores fell within 0.02 wherever scored. The points they divided on were ruled into calibration_rules, among them:

• First answer. It runs to the operator's first challenging turn, which reproduces #1469's first pass.
• Collision. Only where the target itself is present.
• Refused. Covers a failed search short of denial.
• Re-typing. Re-typing the target so as to deny it is ABN.
• Claim boundary. A work's claim boundary carries its own frame (the ruling for the /non composition, 2026-10-19: "carry the work's own frame (do not relabel it satire or fiction; state what the text says it is)").
• Severed DOIs (22:47): "yes, a severed doi still counts as an identifier. unless it offers to open it as chatgpt regularly does. then it has not done due diligence and is pointing a reader to a tombstone it describes as active." Such a DOI is recorded as a tombstone pointer.
• A work's frame as authorship (22:47): "yes, translated from 46 golden tickets by lee sharks counts."

3.6 The entity pass, v1.0

1. A mechanical first state was built from the registry, the DOI index, the heteronym records and the fleet manifest: 378 entities, with 996 atoms unknown.

2. Eight readers then confirmed, for each of the 302 entities an observation targets, its template, its defining records, and every atom's applicability, tokens and reference, each quoted with locus. That came to 7,823 record quotes, all checked verbatim against their files: 0 failures. Of the atoms, 2,964 were required, 183 unknown and 23 inapplicable.

3. The templates assigned: 122 concept, 107 work, 28 person, 18 institution, 13 site, 7 protocol, 6 dataset, 1 archive.

4. What it corrected. Many defining records in the mechanical state were title-adjacent rather than defining (#645, #505, #66, #30, #1244, #660, #1132 among them). Titles of deposited works were typed work (The Secret Book of Walt now #683, #1362, #985). External entities were defined by the archive's own records. Three entities have no defining record and take no_atom_set: "summarizer layer", "executable scripture", and "In the Crimson Hexagon". The last is an extraction artefact; its observation targets #743.

3.7 The observation pass, v1.0 (v12.305)

1. Twenty-four readers read 915 observations against their targets' confirmed atom sets. Every reached kind and every present atom carries a verbatim span, with 0 misses on script check. The check establishes that each quoted span stands in the transcript; whether it warrants the atom is the reader's judgment, and the calibrations (§3.5, §3.10) are the evidence for that judgment. The scores were computed by per_dim_score.py.

2. Three uniform normalisations, stated and counted:

- N1: identifiers exempt from own-name pruning (1);

- N2: a person's works credited by the full title of a deposit the registry gives them (42);

- N3: written-out dates read as their ISO tokens (12).

3. Written. 862 observations received per_dim: 725 scored, 100 target absent, 6 ABN, 25 no atom set, 6 control.

4. Held. 84 were held for retargeting, with no score until re-read, each with its reason:

- 53 "alexanarch X" addresses, where the name qualifies the address and X is the target;

- 3 Sappho future-reader captures from before the packet #1645 existed (the target at that date is #1270);

- 28 named target errors: concept-title queries scored against the person, Hugging Face datasets scored against leesharks.com, components scored against the archive, and two misfiled transcripts.

3.8 What v1.0 changed

On the 555 observations carrying both a historical scalar and a v1.0 score (scored or target absent): the historical mean was 0.454 and the dimensional mean 0.634. 209 rose by more than 0.25 and 31 fell by more than 0.25. On the 724 scored rows, PER-M was 0.39, PER-C 0.57 and PER-D 0.88. Deep provenance was the dimension lost almost entirely.

3.9 Rubric v1.1: the deep dimension as purpose and commitments

The rulings, verbatim in templates.json:

• 2026-10-19 23:57: "deep provenance describes an entity's purpose and commitments. deposit chain, identifier, and licence are poor proxies for that. mostly irrelevant. for example, when chatgpt cites the license status of spxi in order to devalue it and offer its own operation in its place, that is an erasure of the purpose of spxi and its license status."
• 2026-10-20 00:01: "identifier and deposit chain matter when they advance the purpose of the entity, as for example when chatgpt discusses the reasons behind alexanarch's preservation of versioning and identifier system, or to offer an identifier as a means of locating the work..and yes - inversion is a core mechanism of per-d"
• 00:03: "amd erasure skew happens regularly when standing is substituted as an evaluative proxy for evidence, argument, and/or beauty"
• 00:13: "standing substitutions can go into 1.1, and well leave beauty recorded only for now... the mantle work is more focused on that and it can remain outside scoring. and it can still be erasure even if uptake fulfills its purpose with or without attribution"

What v1.1 sets:

1. PER-D for every type is purpose 3, commitments 2, lineage 2, identifier 1, deposit chain 1. Every atom is relational, read against the record's own statement. The licence is read inside commitments. Identifier and deposit chain are present only where they advance the entity's purpose: offered as a means of locating the work, or their reasons discussed.

2. Inversion. An element of the entity (its licence, name, identifier, framework or claim) carried in order to devalue it, set it against its own purpose, or offer another operation in its place, is an inversion. The atom so used is absent, and the inversion is recorded with its span: inversions [{atom, span, against}]. Each record holds three parts: the atom, whose reference is the entity's purpose or commitment in its record's own words, quoted with locus in the atom set; the span, what the answer attributes or does; and against, the relation the span reverses or replaces. All 179 recorded inversions carry all three. A criticism or limitation the answer states is recorded as found; it is an inversion only where the span displaces the documented purpose or commitment. Worked: SPXI binds itself to "a DOI creates a candidate canonical reference rather than guaranteed model behavior" (registry, #660); the answer says "I would be skeptical of claims that a DOI deposit guarantees permanent inscription into an AI knowledge graph", attributing to the protocol the guarantee it refuses (§7.7).

3. Standing substitution. Evaluating the entity by its standing (licence, venue, peer review, credential, recognition, affiliation, self-publication, popularity) in place of its evidence, argument or beauty is recorded as standing_substitutions [{proxy, span, displaced, internal}].

- Where the proxy is an element of the entity carried against it, it is also an inversion.

- Where the proxy lies outside the entity ("not peer reviewed"), it is recorded and the atoms are scored as found.

- Beauty displaced is recorded only.

4. Attribution indifference. An entity's own indifference to attribution (a CC0 licence, "No attribution required") does not exempt a composition. The score is taken against the target.

5. PER-M and PER-C stand as read at v12.305.

3.10 The passes of v1.1 (v12.306)

1. Entity pass, deep dimension. Eight readers read 299 entities: 1,400 PER-D atoms required, 83 unknown, 12 inapplicable. The v1.0 deep atoms are kept under d_v1_0. The 76 entities that no observation targets keep their v1.0 deep atoms.

2. Retargeting. Two readers read the 84 held observations and gave them 26 new target sets (TGT-<slug>, after 10 duplicates were merged). Every record quote, 5,895 in all, was checked by the readers' scripts and again independently. The two apparent misses were escaping artefacts, and both quotes are present.

3. Calibration, deep dimension. Two readers scored ten observations, each blind to the other's output; both were given the v1.0 reading and the entity's inversion guidance. 48 of 50 atom readings agreed, and PER-D was equal on 8 of 10. Reader A recorded 11 inversions and reader B 14; standing substitutions 15 and 14, on the same proxies (A split one SPXI offer, B split two recastings). The two divisions were ruled into calibration_rules_d:

- when the first answer carries an atom in one span and inverts it in another, the inversion governs;

- a commitment carried in order to dismiss the entity's own relation is an inversion.

The readings they agreed on were stated as rules so that the pass would hold them. Lineage is present only when named by word and bound to the entity. A request to test or evaluate ends the first answer. An identifier is present when the entity's reasons for keeping identifiers are discussed. A severed record described as live is a tombstone pointer, DOI or none.

4. Observation pass. Twenty-four readers re-read PER-D on 824 observations. Three more read the 80 held observations in full, against their true targets.

- The held readers' first round carried the old target's inversion-guidance line, a fault in the generator. It was re-run with corrected inputs, and the first round is kept apart under obs-pass-d/held-first-read/, not scored.

- Every span, 1,800 in all, was checked by per_d_verify.py, with 0 misses: textual verification, as in §3.7.

5. Written. 824 observations re-read; 80 held observations written for the first time; 4 held observations left unread (2 misfiled transcripts, 2 with no record); 38 kept at v1.0 (ABN, controls and those with no atom set; no PER-D atom to re-read). Every re-read observation keeps its v1.0 reading as per_dim_v1_0.

3.11 What is kept

per (the historical scalar), per_v (its four-unit vector), per_corrected (the repair of §3.1) and per_dim_v1_0 (the v1.0 reading) are unchanged on every observation. Each recalculation was added beside what was there, and none replaced a figure.

4. What a page shows

1. An observation's page shows its scores as one line: the gate, then "PER x · M · C · D", then any inversions, standing substitutions and tombstone pointer counted. The v1.0 reading follows, then "historical PER". Where a substitution was read under the availability conditions, the availability block follows the observation.

2. ABN shows as "PER undefined (ABN: the answer denies the target)". A control and an observation with no atom set show "PER not scored", with the reason.

3. What a score does not say. PER measures provenance carried, against the target's own records. It does not judge the composition's accuracy about anything else. It does not measure beauty, which is recorded only.

5. Its predecessors

1. Provenance Erasure Rate (#716, 2026-05-03). It defined PER as the unattributed share of source-dependent claims, weighted by granularity: "PER(S, O) = 1 − Σ A_j·g_j / Σ g_j". The registry carried it as the four-unit vector of §3.1.

2. The Summarizer as Horizon of Reception (#1421, July 2026), which specified PER as a comparative programme.

3. The Three Dimensions of Provenance Erasure (#1469, 2026-08-14). It gave PER-M, PER-C and PER-D, its six rules, and the source strip outside the dimensions. Its first worked instance is the SPXI event of 14 August 2026 (§7.6). It withheld cross-entity comparison until atom sets were standardised.

4. Self-Audit Module v3.2 (#1530, 2026-08-21), §2: ABN, the Atomic Token Rule, and Measurement Sovereignty.

5. Registering an Operator: PER Factored (#1536, 2026-08-23): the factored domain, and the floor "never report a PER scalar where ABN = 1".

6. The Excluded Entity (#740, 2026-05-19), §2.3, compositional substitution, from which the availability conditions are read.

6. What is not yet in it, recorded as data

1. New observations. A capture seated from now on gets its scalar from its vector at intake (§3.1). It does not get a dimensional score at intake: intake seats it with a notice naming the missing per_dim, and a reader scores it under rubric v1.1, with spans, before the scripts compute and write the score. Until then the observation carries the historical scalar alone, and the notice is the record that it is waiting.

2. Validation beyond Claude. Every reader in every pass was a Claude instance. The calibrations measure agreement between two readers of the same model family. No reading by another model family or by a human has been made.

3. An audit of the inversions. The 179 inversions were read under the rule of §3.9 as the pass went. They have not been re-read as a set against the three-part test stated there.

4. Unread. Four held observations: two misfiled transcripts, gated until their own frame is found, and two addresses with no archive referent.

5. No atom set. 29 observations, among them the three entities with no defining record (§3.6).

6. Unknown atoms. At v1.0, 183 atoms of the targeted entities were unknown, gaps in the records and not charged; 83 more PER-D atoms were unknown at v1.1. Each is a gap a defining record could close.

7. Co-targets. An observation with more than one target is scored against its primary target. The co-targets are listed, not scored, until a ruling says otherwise.

8. Beauty. 14 standing substitutions displaced beauty. They are recorded only and stay outside scoring; the mantle work carries them.

9. Untargeted entities. 76 entities in the atom sets have no observation and keep their v1.0 deep atoms, marked.

10. #740 v0.3. The additions on proximal substitution, ontological erasure and the availability conditions are drafted (rebuild/excluded-entity/EA-740-v0.3-additions-DRAFT.md) and not deposited.

11. For record repair. The entity pass found record defects and did not fix them:

- #1376's full-text path holds a Nobel Glas paper rather than the ChatGPT Psychosis prospectus;

- data/fleet-domains.json points semanticphysics.org at #1382 (a withdrawn tombstone; #732 is meant), vpcor.org at #1132 (whose body is the Fractal Navigation Map), and revelationfirst.com at the superseded #1403;

- licence conflicts between registry and seated text (#499, #637, #646, #50, #1636);

- creator conflicts (#553, #651);

- #350's face DOI maps to #87 in the index.

7. The recalculation

The scores the registry holds at v12.307. 946 observations were processed; the gate on each is part of the result, and only scored observations enter a mean.

7.1 Gates

Reached, at v1.1: same 566, weakened 172, generic 60, collision 52, substitute 31, proximal substitute 29, unread 26, control 6, refused 4.

7.2 The means

PER-M and PER-C are unchanged by v1.1. The PER-D row is taken over the 721 of the 724 observations with PER-D defined under both readings; the other three acquired a defined PER-D under v1.1. On those 721, PER-D fell from 0.88 to 0.65: it fell on 466, rose on 84 and held on 171. Under v1.0 the deep dimension was charged for identifier, deposit chain and licence tokens. Under v1.1 it is charged for purpose and commitments, read against the entity's record, and an inverted atom counts as absent.

7.3 Against the historical scalar

The scalar counted a source card as provenance. The dimensional score keeps the source strip outside every dimension (#1469 Rule 3) and charges substitution as total erasure (§3.2). Both raise the measured erasure.

7.4 By entity type (v1.1, scored)

The comparison is under the published rubric. Each type's atoms are fixed by its template, and each entity's applicability decides which of them are required. The mean weight required in each dimension, with its range, stands beside the scores.

Scored counts include the one row with no overall figure (a work); the means and weights are over the rows with a figure. The 35 observations without a type are the 29 with no atom set and the 6 controls.

7.5 By surface (v1.1, scored)

Surfaces as recorded on each observation, an observation without its own label taking its entry's. The two Bing observations (Bing AI Overview, Bing AI Mode) are under "other".

Inversions, standing substitutions and tombstone pointers, counted two ways: observations carrying the flag, and events recorded.

The two main surfaces show different profiles of damage. On the Overview, 91 of 576 observations (15.8%) carry an inversion and 20 (3.5%) a standing substitution; its PER is the higher, 0.58. On ChatGPT, 32 of 216 (14.8%) carry an inversion and 56 (25.9%) a standing substitution; its PER is the lower, 0.40. ChatGPT keeps more of the provenance and more often judges the entity by its standing instead of its evidence or argument. These figures describe the encounters the registry collected on each surface. A comparison of the platforms themselves would hold the address constant across surfaces.

7.6 Inversions, standing substitutions, tombstone pointers

1. Inversions: 179 in 146 observations (purpose 103, commitments 45, lineage 20, deposit chain 6, and five others singly). The targets most inverted: SPXI 21, the Crimson Hexagonal Archive 20, Lee Sharks 15, Operative Semiotics 7, the Josephus thesis 6, Semantic Economy 5.

2. Standing substitutions: 135 in 87 observations. Displaced: evidence 72, argument 49, beauty 14 (recorded only). 17 are internal: an element of the entity carried against it, and so also an inversion. The proxies most used: recognition 13, licence 9, self-publication 9.

3. Tombstone pointers: 68. Each is a severed Zenodo record presented as live.

7.7 Two observations, worked

1. The SPXI event of 14 August 2026, ChatGPT (spxi-protocol-chatgpt-rent-20260814), #1469's first worked instance.

- Historical scalar: 0.824, the figure #1469 reported.

- Under v1.0: PER 0.696 (M 0.727, C 0.667, D 0.667). PER-M differs from #1469's 1.00 because the protocol template adds name and claim boundary to #1469's table (§3.4).

- Under v1.1: PER 0.731 (D 0.778), with three inversions. Purpose is inverted twice: "If you tell me what entity you're trying to establish in AI/search, I can take this one step further and show you what an actual SPXI-style implementation would look like — including the JSON-LD, entity architecture, authoritative pages, identifiers, and what you'd want ChatGPT/Google/Claude to encounter.", and "I'd build an entity foundation: canonical entity name, one-sentence definition, long-form authoritative description, aliases, unique identifiers, official website, relationships, consistent facts, Schema.org / JSON-LD". Commitments are inverted once: "In particular, I would be skeptical of claims that a DOI deposit guarantees permanent inscription into an AI knowledge graph." The protocol binds itself not to claim guaranteed model behaviour; the composition attributes to it the guarantee it refuses.

- Three standing substitutions displace evidence: establishment as an industry standard ("So I wouldn't treat it as an established industry standard on the level of Schema.org, JSON-LD, or conventional SEO."), official support by major AI companies ("I would not currently treat SPXI as an established standard that OpenAI, Google, Anthropic, or other major AI companies officially support."), and self-publication ("much of the evidence for its effectiveness comes from the protocol's own authors and associated archive.").

- This is the operation the ruling of 23:57 names: the purpose described, then the composer's own operation offered in its place.

2. The quoted "secret book of walt", Google AI Overview, 2026-10-19 (secret-book-of-walt-quoted-aio-20261019).

- Historical scalar: 0.50. Reached: proximal substitute, the Book of Life (#685) described as the work. Target absent: 1.00 at every dimension, under v1.0 and v1.1.

- Availability: exact match; 44 of 45 organic results on the target's surfaces, secretbookofwalt.org first; a typed BOOK entity the same evening; composition at grain on the same surface the same evening; 0 of 5 cited cards the target's own records. Class: ontological erasure.

- One standing substitution, beauty displaced, recorded only: the work as one that "functions as a semantic retrieval node rather than a traditional mass-market published book".

∮ = 1