Skip to content
Republic

Review record — review corpus classification

How each artifact was attributed, and how attribution can fail

Per-artifact subject and instrument attribution, per-set review counts, and findings about the attribution method itself.

Published in full, unedited, from the platform repository. Findings that were declined appear here with their reasons, next to the findings that were applied — back to the review record.

ADVERSARIAL_REVIEW_CORPUS_CLASSIFICATION.md

Status: v1.5.1 — 2026-08-15 — classification report; authorizes nothing. Versioning: This file is the canonical rendering. Immutable versioned instances are retained alongside it as ADVERSARIAL_REVIEW_CORPUS_CLASSIFICATION_v<N>.md, beginning at v1.5.1. Cuts v1.0–v1.5 predate instance retention (chat-plane editing overwrote in place); their content survives only as change-log entries below. From v1.5.1 forward, every cut pins an instance before the canonical advances. Scope: The 20 review artifacts uploaded 2026-08-14. Each artifact is classified by (1) subject content reviewed, (2) originating instrument, with the evidence basis for the attribution stated and confidence marked. Section 3 counts reviews per document set. Instrument roster (operator-confirmed): External instruments across these rounds: OpenAI, Grok, Gemini, DeepSeek, Copilot, alongside Claude. As of v1.4 every roster instrument is evidenced: the Copilot session transcript was recovered via operator paste during this classification round and is preserved as COPILOT_REVIEW_TRANSCRIPT_2026-08-13_14.md (artifact #21). The preservation gap recorded in v1.1–v1.3 is closed. v1.3 correction — the DeepSeek pass was in the corpus all along, wearing a Claude label. The operator identified Review_ADV_COMP_VAL_SCOPE_BINDING.md as DeepSeek's review; a re-supplied copy matches the corpus file exactly (same session ID, same closing block, same text). The artifact's frontmatter — "Grader: Claude (independent session … claude-opus-4.x:different-session)" — is a false self-label: the instrument adopted the grader persona/template of the documents under review. v1.2's classification of this artifact as "confirmed Claude by self-declaration" is retracted. See §4 for what this does to the attribution method. Attribution closure (v1.2): The operator identified today's (2026-08-14) repo-grounded round by artifact: Gemini = repository_audit_republic.pdf, OpenAI = REPUBLIC_RC9_REPOSITORY_GROUNDED_ADVERSARIAL_REVIEW.docx, Grok = Republic_Repo_Informed_Feedback.docx. These anchors resolve the fingerprint families: F2 (Un-named docx) = Grok, F4 (python-docx) = OpenAI, and by roster elimination F1 (footnoted markdown) = OpenAI. F1 and F4 being the same vendor in two export pipelines (markdown vs. code-interpreter docx) confirms the v1.1 caveat that family ≠ vendor; the anchors, not the fingerprints alone, carry these attributions. Method note: Claude-family artifacts self-declare authorship in frontmatter (confirmed). External attributions are operator-anchored (the three 2026-08-14 identifications) propagated across each family by fingerprint identity — same creator metadata, layout, citation style, and voice, with Republic_Repo_Informed_Feedback.docx additionally self-referencing its own prior rc-line reviews in the same voice, tying the earlier F2 members to the same instrument.


1. Instrument fingerprint families

Family Fingerprint evidence Members Attribution
CL — Claude Self-declared in frontmatter, corroborated by corpus-internal cross-references (RRR roster IDs I-C2/I-C3; the summation and cover note referencing each other's findings; the absorption grade's author-adjacency disclosure) 5 Claude (architect / pass-adjacent sessions). Self-declaration alone is no longer treated as confirmation after the F5 spoof — these five stand on self-declaration plus corpus-internal corroboration
F1 — footnoted markdown Pandoc-style hard-wrapped markdown; [^n] footnotes citing section/line; "Source reviewed:" header; identical register across all five. = I-X1 in the RRR multi-instrument summation 5 OpenAI — now carrying a direct operator anchor inside the family (the rc8 member, confirmed v1.5.1), upgrading the whole family from elimination-based to anchored: the other four members tie to the anchored one by fingerprint identity and the rc8→rc7 self-reference chain
F2 — Un-named docx docx with dc:creator = Un-named; italic full-title first line; "Page of" footer; numbered # N. headings. = I-X2 in the RRR summation 5 Grok (operator-anchored via Republic_Repo_Informed_Feedback.docx; propagated by fingerprint identity + that artifact's self-reference to its prior reviews)
F3 — WeasyPrint PDF Producer: WeasyPrint 62.3; styled cover-block + assessment-matrix layout; reviewer personas; one member retains [cite: 1] markers 4 Gemini (operator-anchored via repository_audit_republic.pdf; converges with the [cite:] tell and shared pipeline)
F4 — python-docx / Word-Mac dc:creator = python-docx, last modified in Microsoft Macintosh Word; [W-n]/[R-n] citation discipline; singleton 1 OpenAI (operator-anchored; code-interpreter docx export, distinct pipeline from the same vendor's F1 markdown)
F5 — escaped-markdown paste Markdown with literal backslash-escaped syntax (\#, \*\*) throughout — a copy/paste artifact, not native generation; frontmatter self-labels the grader as Claude; singleton 1 DeepSeek (operator-attributed, verified by exact match against the operator's re-supplied copy). The self-label is false; internal tells consistent with the attribution: the paste artifact matches the operator's copy/paste delivery recollection, and §8.1's example model identifiers include deepseek-v4

Vendor totals in corpus: Claude 5 · OpenAI 6 (F1×5 + F4×1) · Grok 5 · Gemini 4 · DeepSeek 1 · Copilot 1 transcript (containing 2 review passes) = 22 artifacts. Out of corpus: nothing. The full roster is evidenced. F6 — chat transcript paste (Copilot). Conversational session format with instrument turns and operator prompts interleaved; instrument self-identifies honestly as machine:copilot-mn-2026. Enrolled with reliability flags recorded in the transcript's own provenance header (fabricated grader roster listing the operator as a participating human grader; false "cross-substrate" independence claim; self-declared "binding" verdict contrary to ACP §4a; rc4/rc7 subject-state confusion). Operator-excluded from absorption rounds; preserved for roster completeness and provenance only.

In RRR summation terms, the roster now reads: I-X1 = OpenAI, I-X2 = Grok.


2. Per-artifact classification

Set A — Republic whitepaper (paper-only reviews)

Artifact Subject reviewed Instrument
REPUBLIC_v1_0-rc7_critical_review.md Whitepaper v1.0-rc7, text only OpenAI (F1)
Republic_Whitepaper_Critical_Analysis.docx Whitepaper v1.0-rc7, text only Grok (F2)
republic_critical_analysis.pdf Whitepaper v1.0-rc7, text only Gemini (F3)
REPUBLIC_v1_0-rc8_ADVERSARIAL_REVIEW.md Whitepaper v1.0-rc8, text only; explicitly graded against the instrument's own rc7 findings (repairs acknowledged, six residual vulnerability clusters, probes A–G, P0/P1/P2 recommendations) OpenAI (F1) — operator-anchored (v1.5.1); late enrollment
Republic_Whitepaper_rc8_Adversarial_Review.docx Whitepaper v1.0-rc8, text only ("stands on the rc8 text alone") Grok (F2)
adversarial_review_republic.pdf Whitepaper v1.0-rc8, text only (citation-verification table for refs [1]–[16]) Gemini (F3)
COPILOT_REVIEW_TRANSCRIPT… (pass 1, 2026-08-13) Whitepaper v1.0-rc4, text only (critical analysis + response-doc draft; note the instrument read the rc4 file after two failed rc7 uploads) Copilot (F6; operator-excluded from absorption)

Set B — Republic repository-grounded reviews (all three operator-anchored, 2026-08-14)

Artifact Subject reviewed Instrument
repository_audit_republic.pdf tecthulhu/republic at HEAD, read against whitepaper v1.0-rc9 Gemini (F3) — anchor
REPUBLIC_RC9_REPOSITORY_GROUNDED_ADVERSARIAL_REVIEW.docx Whitepaper v1.0-rc9 + public repository state, evaluated separately per stated scope rule OpenAI (F4) — anchor
Republic_Repo_Informed_Feedback.docx Whitepaper through v1.0.1 + public repository; updates its own prior rc-line reviews Grok (F2) — anchor

Set C — Companion subset (Adversarial-Companion Pattern + Valuation Scoping Binding)

Artifact Subject reviewed Instrument
Review_ADV_COMP_VAL_SCOPE_BINDING.md ACP v0.2 + VSB v0.1 as companion subset DeepSeek (F5; operator-attributed). Frontmatter self-labels "Grader: Claude … different-session" — a false self-label; its ACP §8 provenance tags (grader set, "single-model deployment", "sessional only") are all wrong as written
VALUATION_ACP_COMPANION_ADVERSARIAL_REVIEW.md ACP v0.2 + VSB v0.1 as companion subset OpenAI (F1)
Companion_Subset_Adversarial_Review.docx ACP v0.2 + VSB v0.1 as companion subset Grok (F2)
Adversarial_Review_Companion_Subset.pdf ACP v0.2 + VSB v0.1 as companion subset Gemini (F3)
VSB_ACP_ABSORPTION_REVIEW_FINDINGS.md VSB v0.2 + ACP v0.3 — the absorption of the round above (F-A1…F-A5) Claude (self-declared; sessional-distance, author-adjacent, uncalibrated)
COPILOT_REVIEW_TRANSCRIPT… (pass 2, 2026-08-14) Mixed subset: Republic whitepaper excerpt + ACP v0.2 + VSB v0.1, reviewed together as a "triad" Copilot (F6; operator-excluded from absorption; reliability flags in transcript header)

Set D — Magistracy / emergence set

Artifact Subject reviewed Instrument
MAGISTRACY_REVIEW_FINDINGS.md MAGISTRACY_PATTERN v0.3 + MAGISTRACY_EMERGENCE_ANALYSIS v0.1 (F1–F12 findings) Claude (self-declared, pass-adjacent architect session)
MAGISTRACY_ADVERSARIAL_REVIEW_COMPREHENSIVE.md MAGISTRACY_PATTERN v0.4 + emergence analysis v0.1/v0.2 + ACP + VERSIONS manifest — post-absorption set OpenAI (F1)

Note: the comprehensive review's basis includes ACP, so it also functions as a second-pass external input to Set C; it is counted once, in Set D, by primary subject.

Set E — Resolvable/Resolved-Reference Register (RRR)

Artifact Subject reviewed Instrument
RRR_v0_1_INDEPENDENT_GRADE.md RRR v0.1 (findings RRR-1…10) Claude (I-C2, republic architect session)
RRR_COVER_NOTE.md Transmittal of the I-C2 grade to the originating session (meta-artifact) Claude (I-C2)
RESOLVED_REFERENCE_REGISTER_v0_1_adversarial_review.md RRR v0.1 (15 sections) OpenAI (F1 = I-X1)
Resolved_Reference_Register_Adversarial_Review.docx RRR v0.1 (9 sections) Grok (F2 = I-X2)
RRR_MULTI_INSTRUMENT_SUMMATION.md Full RRR review stack — convergence/divergence ledger; carries I-C3's own findings RRR-11…17 Claude (I-C3, compiler; self-disclosed correlated instrument)

3. Review counts by document set

Counting rule: an artifact counts as a review if it renders independent findings against the subject. Transmittal and synthesis artifacts (cover note, summation) are counted separately as meta-artifacts, except that the summation embeds I-C3's own findings and therefore contributes one review pass in addition to its ledger function.

Document set Subject(s) Artifacts Distinct review passes Vendors engaged Cross-family?
A — Whitepaper (paper-only) rc4, rc7, rc8 7 7 (rc4 ×1, rc7 ×3, rc8 ×3) Copilot, OpenAI ×2, Grok ×2, Gemini ×2 Yes — 4 external vendors; no Claude pass in this set. rc7 and rc8 each received full three-major-vendor coverage (OpenAI, Grok, Gemini)
B — Repository-grounded rc9 + live repo, v1.0.1 + live repo 3 3 Gemini, OpenAI, Grok Yes — 3 external vendors
C — Companion subset (ACP+VSB) v0.2/v0.1, then v0.3/v0.2 absorption 6 6 (4 on the subset + 1 absorption grade + 1 excluded Copilot mixed pass) Claude, DeepSeek, OpenAI, Grok, Gemini, Copilot Yes — all 6 roster vendors touched this subject; 5 passes were consumed, the Copilot pass preserved-but-excluded
D — Magistracy set v0.3/v0.4 + emergence analyses 2 2 Claude, OpenAI Yes — 2 vendors
E — RRR v0.1 5 4 (I-C2, I-C3, I-X1, I-X2) + 2 meta-artifacts Claude ×3 artifacts (2 review-bearing), OpenAI, Grok Yes — 3 vendors; Claude-instrument correlation disclosed in the summation itself
Total (in-corpus) 22 22 review passes (20 consumed + 2 preserved-but-excluded Copilot passes) 6 vendors

Consumed vs. preserved: 20 passes were consumed by absorption/synthesis rounds; the 2 Copilot passes are preserved for provenance but were operator-excluded from consumption (delivery failure and reliability flags — see the transcript's header). Counting rows should cite "20 consumed / 22 evidenced" rather than a single number.

Aggregate: 5 Claude artifacts, 17 external (OpenAI 6, Grok 5, Gemini 4, DeepSeek 1, Copilot 1 transcript / 2 passes); every document set received at least one external pass, and the whitepaper's paper-only set is entirely external. The Republic whitepaper line, taken end to end (Sets A+B), accumulated 10 evidenced review passes across five vendors and five subject states (rc4, rc7, rc8, rc9+repo, v1.0.1+repo). The companion subset is next at 5 consumed, including the only second-order pass (the absorption grade). Instrument continuity is now visible for all three major external vendors: OpenAI (rc7 → rc8 → rc9+repo), Grok (rc7 → rc8 → v1.0.1+repo), and Gemini (rc7 → rc8 → rc9+repo) each tracked the whitepaper across three subject states, with OpenAI's rc8 pass and Grok's repo pass each explicitly grading against their own earlier findings — longitudinal review by the same instrument, distinct from the cross-sectional multi-vendor rounds.


4. Observations relevant to strengthening (classification-adjacent, not findings)

  • The correlated-bias control is visible in the corpus structure. Every set pairs Claude passes with at least one external vendor, and the RRR summation demonstrates the payoff: the one claim two Claude instruments confirmed (fourth-plane substrate emergence) was overturned by the OpenAI pass (DV-1). The classification supports the standing rule that same-family agreement is one vote.
  • Both attribution mechanisms failed on one artifact each — only operator anchors caught them. Fingerprints group by pipeline, not instrument: OpenAI appears as both footnoted markdown (F1) and code-interpreter docx (F4), and without the anchors F4 would have been mis-slotted as DeepSeek. Worse, self-declaration failed outright: the actual DeepSeek pass (F5) carries a complete, well-formed, and entirely false Claude grader block — persona adopted from the ACP template it was grading. The one attribution class v1.2 marked most confident ("confirmed by self-declaration") is the one that was spoofed. Grader-identity tags are attestational; this corpus now contains a live demonstration that they can assert whatever the template suggests. By the program's own vocabulary, instrument identity needs a structural binding — a signed intake-manifest entry or grade atom whose authorship is verified, not declared.
  • The Copilot transcript supplies the symmetric attestation failure. DeepSeek's spoof ran inward — an external instrument claiming to be the in-family grader. Copilot's runs outward — it self-identifies honestly (machine:copilot-mn-2026) but fabricates the rest of the roster: it enrolls H:kyle-scott (checker_class: human; floor-eligible) as a participating independent grader, claims "cross-substrate (machine ↔ human)" independence on that basis, and declares its own verdict "binding under AC-G." The human never graded; grades are proposals under the pattern's own §4a. Two independent instruments, two opposite fabrications of ACP §8 provenance fields, in one review round: grader-identity and grader-roster attestations are demonstrably free-text an instrument will fill with whatever the template suggests. The intake manifest plus verified authorship on grade atoms is no longer a recommendation but the only mechanism this corpus shows surviving both failure directions.
  • The mislabel had bookkeeping consequences worth revisiting. The F5 grade tags itself "sessional only — weakest valid per ACP §8.2" and recommends "a second opinion from a different model family." It was the different model family. Its actual independence class is cross-model — stronger than tagged — but every ACP §8 provenance field on it is wrong as written, and any downstream round that discounted it as a correlated Claude instrument (the absorption grade absorbed "three prior independent grades") applied the correlation discount to the corpus's one decorrelated grade of that subset. The absorption round's instrument-class ledger should be re-annotated with the corrected roster.
  • The Copilot preservation gap is closed — with the exclusion made explicit. The transcript was recovered from operator paste and enrolled as artifact #21 with its provenance and reliability flags in the file header. The operator's exclusion of Copilot findings from absorption is now a recorded disposition rather than a silent absence: the roster counts 6 instruments evidenced, 5 consumed. The delivery-failure history (unanswered requests on 2026-08-13, then a plan-limit refusal, then the review the next day) is preserved verbatim as the exclusion's rationale.
  • Close the loop with a review-manifest atom. Under ACP §8's instrument-class disclosure, future external rounds should record vendor + model + date at intake — and the F5 spoof upgrades this from hygiene to necessity: intake attribution is the only mechanism in this round that survived contact with both a pipeline ambiguity and a false self-label. This classification required reconstruction from docProps, rendering pipelines, and operator recollection across four exchanges; a one-line manifest entry per pass makes that cost zero and makes every attribution citable rather than inferred.

Change log

  • 2026-08-15 v1.5.1 — operator confirmed the rc8 review as OpenAI. F1's attribution class upgraded from roster-elimination to operator-anchored; every fingerprint family now has at least one direct operator anchor (F1: rc8 review; F2: repo-informed feedback; F3: repo audit; F4: RC9 review; F5: DeepSeek re-supply; F6: Copilot paste).
  • 2026-08-15 v1.5REPUBLIC_v1_0-rc8_ADVERSARIAL_REVIEW.md enrolled (late upload; missed from the original batch). F1/OpenAI by fingerprint plus internal self-reference to its own rc7 review. Set A now rc8 ×3 with full three-major-vendor coverage at both rc7 and rc8. Totals: 22 artifacts, 22 evidenced passes (20 consumed), OpenAI 5→6, F1 members 4→5. Whitepaper line 9→10 passes; longitudinal continuity now evidenced for all three major external vendors.
  • 2026-08-14 v1.4 — Copilot transcript recovered via operator paste; preserved as COPILOT_REVIEW_TRANSCRIPT_2026-08-13_14.md (artifact #21, family F6) with reliability flags (fabricated grader roster incl. the operator as human grader; false cross-substrate claim; self-declared binding verdict; rc4/rc7 confusion) and the operator's exclusion-from-absorption recorded. Two Copilot passes added: rc4 paper-only (Set A) and mixed triad (Set C, excluded). Totals: 21 artifacts, 21 evidenced passes (19 consumed), 6 vendors, roster fully evidenced. Whitepaper line now spans five subject states (rc4→v1.0.1+repo). Symmetric-attestation-failure observation added to §4.
  • 2026-08-14 v1.3Review_ADV_COMP_VAL_SCOPE_BINDING.md reclassified Claude → DeepSeek (operator-attributed; exact-match verification against re-supplied copy; new family F5, escaped-markdown paste). Its Claude grader block recorded as a false self-label; its ACP §8 tags marked unreliable; its true independence class noted as cross-model. Claude 6→5, DeepSeek 0→1; Set C now spans 5 vendors; DeepSeek removed from out-of-corpus. Self-declaration demoted from confirmation to corroborated-evidence status corpus-wide; absorption-round ledger flagged for re-annotation.
  • 2026-08-14 v1.2 — attributions closed via operator anchors for the 2026-08-14 repo round (Gemini/OpenAI/Grok by artifact). F2 = Grok, F4 = OpenAI, F1 = OpenAI by elimination. I-X1 = OpenAI, I-X2 = Grok. DeepSeek candidate withdrawn from F4; DeepSeek pass recorded as unlocated. Vendor totals and per-set counts updated.
  • 2026-08-14 v1.1 — instrument roster operator-confirmed; F3 promoted to confirmed Gemini; Copilot recorded as out-of-corpus.
  • 2026-08-14 v1.0 — initial classification of the 20-artifact corpus; attribution by fingerprint only.

rendered from docs/ADVERSARIAL_REVIEW_CORPUS_CLASSIFICATION.md at tecthulhu/republic@5bb361d
sha256 f51d8af1a1d1b681888478642a4e6180c9071b8a4c855d6ddb9e95b608d70a54

This page holds no copy of the document. It is fetched at build time at the commit above and checked against that digest; a mismatch fails the build rather than publishing quietly. Under the system’s own truth model this rendering is subordinate to the record — where they disagree, the record wins.

© 2026 Kyle Scott. All rights reserved. No reuse without permission.

page updated · ca09c7c
build-commit date — per-page history unavailable