D0 — governed observation input
Validates the closed-world SRS plus verification input contract. Contract validity is not itself a verifier verdict, trust decision, claim-support decision, or admission act.
/citation-integrity
Citation Integrity
GARP treats citation support as an inspectable workflow. Analytical records may cite exhibits, filings, transcripts, correspondence, reports, datasets, articles, or internal records, but ordinary prose does not show whether those citations resolve, whether the cited fragment exists in source text, or which references remain unmapped.
Citation Integrity makes that support layer visible. It is a methodology with a reference implementation, not a general legal analysis system, a truth engine, or a substitute for human review.
Every citation should have a chain of custody
Counterpedia is designed to make it difficult for information to quietly change meaning as it moves through a system. Seeing a page is not the same event as acquiring bytes. Acquiring bytes is not the same as emitting a receipt. A receipt is not verification. Verification is not a claim decision. And evidence supporting a proposal is not the same as granting that proposal standing.
WHAT WAS OBSERVED
≠ WHAT WAS INDEPENDENTLY ACQUIRED
≠ WHAT INGEST DERIVED
≠ WHAT A RECEIPT DECLARES
≠ WHAT VERIFICATION ESTABLISHED
≠ WHAT AUTHORING PROPOSES
≠ WHAT IS ADMITTED
≠ WHAT IS PUBLISHED
The goal is not bureaucracy for its own sake. The goal is to make consequential transformations difficult to hide.
- 01Preserve — Keep the exact observation and material references distinct from later interpretation.
- 02Receipt — Record what the evidence layer declares without treating that declaration as a verdict.
- 03Verify — Check bounded mechanical claims such as subject binding, derivation inputs, or content digests.
- 04Materialize — Resolve the exact bytes bound by the governed observation input.
- 05Project — Expose only explicit byte spans, hash the raw bytes before strict decoding, and perform no summarization.
- 06Make citable — Map each exact projected unit 1:1 to a bundle-local evidence handle as research-only evidence.
- 07Stop — End at the EvidenceBundle. Claim construction, model composition, admission, standing, and publication remain separate acts.
Current landed exact-lineage boundary: citable evidence.
The implemented chain reaches a bounded EvidenceBundle and stops. No claim is approved merely because the evidence path succeeded. No model conclusion receives standing merely because it cites the evidence correctly.
What the landed exact-lineage path does
E0 — exact materialization
Resolves the material references and binds exact bytes to an authoring observation binding. The serialized binding is audit evidence, not a capability token.
EP0 — exact-byte projection
Re-runs the upstream materialization gate, enforces the observation anchors, and slices caller-declared half-open byte spans. Each span is hashed before strict UTF-8 decoding. There is no normalization, summarization, heuristic chunking, or model selection here.
EB0 — citable research evidence
Re-runs EP0, independently rechecks the runtime bytes, and maps each projection unit 1:1 into a bundle-local evidence handle as a research artifact. The result is an EvidenceBundle whose handles are content-addressed to the exact projected source bytes.
Three identities remain separate by design: SRS identities, projection-unit identities such as projection:U001, and bundle-local citation handles such as evidence:E001. A matching number never makes them the same identity; the mapping is explicit provenance.
Evidence is not a claim
A projected evidence item is research-only. Its summary is a deterministic display label, not the source payload. The exact source material is content-addressed by its digest and bound to projection provenance.
EvidenceBundle: these are the evidence units you may cite.
Projection: these exact bytes are what the evidence unit means.
Future composer: these are the exact bytes actually shown to the model.
The third statement is intentionally future work. The current landed chain stops before composer or model execution.
A separate URL-authoring path remains separate
Counterpedia Authoring also has a URL-based route that performs a fresh authoring-side acquisition and can produce proposal-only output. That route is useful, but it is not the same custody claim as reusing the exact SRS/Verify evidence bytes carried through D0, E0, EP0, and EB0.
URL
↓
fresh authoring acquisition
↓
proposal_only
SRS / Verify
↓
D0 → E0 → EP0 → EB0
↓
EvidenceBundle STOP
What it does
Source registration
Source files and extracted text are bridged into a registry so the system can distinguish a citation that resolves to a known source from one that merely appears in prose. Registered sources may carry stable identifiers, labels, paths, content hashes, provenance level, and project context.
Citation extraction
Analytical drafts are parsed deterministically for source references. The output is a structured inventory of what the draft claims to cite, with context preserved for review.
Coverage analysis
Extracted references are compared against the registered source set. Resolved references are reported with source identifiers, while unresolved references remain visible for operator review. The governing question is simple: Which references remain unmapped?
Verbatim fragment verification
Governed fragments are checked against extracted source text under exact or documented normalization rules. SR-07 verifies that fragment text appears in the declared source text; it does not decide truth, legal significance, admissibility, or completeness.
How it works
- 01Source files
- 02Source registry
- 03Citation extractor
- 04Coverage analyzer
- 05Verbatim fragment lint
Each stage produces inspectable output. A coverage report can be generated without creating fragments, fragments can be verified without rerunning coverage, and each layer can carry its own audit trail.
Source mapping table
Real source folders are messy. File names vary, extracted text may live separately from originals, and citation labels may come from court filings, attachments, internal naming, or operator review. GARP uses a source mapping table as the human-reviewable record of which source file corresponds to which citation label.
| Review status | Meaning |
|---|---|
| accepted | Mapping confirmed by operator review. |
| provisional | Mapping useful for mechanical testing but not evidentiary reliance. |
| needs_review | Mapping flagged for human disposition. |
| rejected | Mapping considered and declined. |
| no_source_found | No registered source corresponds to the label. |
The distinction matters because raw coverage and clean reviewed coverage are different measurements. Provisional or unresolved rows may help an operator debug a corpus, but they should not be treated as reviewed source support.
Reference implementation
The reference implementation includes source registration, citation extraction, coverage reporting, mapping validation, unresolved reference ranking, scoped fragment linting, clean coverage reporting, SR-07 verbatim checks, and the landed D0 → E0 → EP0 → EB0 exact-lineage chain in Counterpedia Authoring.
The exact-lineage engineering record is maintained separately from the public knowledge corpus. As of August 14, 2026, canonical EP0 landed through Counterpedia Authoring PR #51 and EB0 through PR #55. EB0 stops at the EvidenceBundle; downstream composer, ClaimMap, and admission work are not implied by that landing.
The implementation is not distributed from this page. This page documents the public methodology and its boundaries.
What “verified” should always answer
Whenever Counterpedia uses the word verified, the next question should be: verified what?
- a digest matches specific bytes;
- a receipt binds to a declared subject;
- an extraction reference resolves to content with the declared hash;
- a projected span hashes to the declared material reference;
- a bundle item maps 1:1 to that projected unit.
None of those findings, by themselves, establish the truth, completeness, significance, admissibility, or institutional standing of a broader proposition.
What it does not claim
SR-07 confirms that fragment text appears in extracted source text under exact or documented normalization checks. It does not confirm source authenticity, admissibility, completeness, or the truth of the proposition that cites the fragment.
Coverage analysis confirms that a citation reference resolves to a registered source. It does not confirm that the source supports the proposition, that the mapping is substantively correct, or that the citation meets a legal or editorial standard.
Source registration confirms that a source file has been bridged with an extracted text path. It does not prove extraction completeness, document recency, or external authentication.
Exact evidence lineage confirms bounded custody relationships between governed observation material, projected byte spans, and citable research-evidence handles. It does not turn evidence into a claim, a truth judgment, an admission decision, or a published Counterpedia record.
Current boundaries
Each registered source needs extracted text before verbatim fragment checks can run. Citation extraction operates on the draft as written, so references outside a documented label format are not visible to coverage analysis until the profile teaches the extractor that form.
The exact-lineage chain currently stops at citable research evidence. Evidence does not become knowledge merely because it has been captured, materialized, projected, and made citable.
The system makes a small number of mechanical claims and makes them precisely. It does not make broader claims that depend on judgment, doctrine, or external authority.