How Does Ethos Detect Wrong, Missing, and Stale AI Citations?

Direct answer

Ethos detects wrong, missing, and stale citations by separating target resolution, evidence matching, fingerprint freshness, and source capability into structured checks. A citation can be Grounded, Mismatch, NotFound, Stale, CapabilityBlocked, unsupported, or erroneous depending on what the verifier can establish.

This taxonomy is more useful than a generic hallucination score. It tells an application whether to regenerate a locator, refresh the source, fix malformed input, obtain stronger parsing, route to review, or block the claim.

The Ethos citation-verification walkthrough explains the complete flow. This article focuses on failure classification.

Verification is a sequence, not one comparison

validate request
  -> check source fingerprint
  -> resolve page / element / span / bbox / table cell
  -> verify source capability
  -> compare quote, value, presence, or structured evidence
  -> compute overall grounding state
  -> derive proof summary

Earlier failures constrain later claims. A stale report cannot become reusable merely because one quotation still matches.

Grounded: the positive state

A supported claim is grounded when the requested target resolves, the required source capability exists, the evidence matches according to the active verification rules, and no report-level invalidator such as a stale fingerprint prevents certification.

Grounded is still a citation-level result. It does not automatically mean:

• the evidence is relevant to the user’s question;
• the source is authoritative under business policy;
• the answer’s synthesis is valid;
• a calculation is correct;
• all claims in the answer are covered.

Applications should use the canonical check plus relevance and claim-type policy before release.

Mismatch: the target exists but evidence differs

Mismatch means Ethos found a relevant target but the requested evidence did not match. Common examples include:

• altered quotation;
• missing or inserted negation;
• correct element ID with different text;
• wrong value at an otherwise valid target;
• table cell resolves but contains another value;
• bounding region differs from the requested region under verification rules.

Do not let fuzzy search silently turn a mismatch into a pass. A correction workflow may suggest another target, but the original citation remains wrong.

NotFound: the target cannot be resolved

NotFound identifies missing evidence targets. Ethos reason codes can distinguish:

• missing locator;
• page not found;
• element not found;
• span not found;
• bounding box not found;
• table not found;
• table cell not found.

This distinction helps identify the failing subsystem:

Unknown evidence IDs must fail closed. They should never be treated as optional citations that the application quietly drops.

Stale: the citation belongs to another source version

When staleness policy requires fingerprint matching, Ethos compares the citation envelope’s document fingerprint with the trusted grounding source fingerprint.

A mismatch sets report-level staleness and prevents all_evidence_grounded from becoming true. Even an otherwise grounded individual check is not reusable while the report fingerprint is stale.

This is important because old and new source versions can contain identical text. Matching words do not prove that the citation was created from the approved version.

The stale RAG citation guide covers source authority and effective-date policy.

CapabilityBlocked: the source cannot prove the request

CapabilityBlocked means the verification request is meaningful, but the grounding source lacks necessary evidence. Examples include:

• span claim against a source without spans;
• table-cell claim against flattened text;
• fingerprint-required policy against a source without fingerprint support;
• bounding-box claim when coordinate origin is unknown;
• crop request against a source without crop support.

Capability-blocked is not the same as mismatch. The verifier did not prove that the evidence is wrong; it proved that the available source cannot establish it.

Do not coerce unknown into false certainty. Route capability-blocked claims to stronger parsing, manual review, or a policy-approved lower-confidence path.

Unsupported claim kinds and errors

A verifier should also distinguish unsupported evidence types from operational errors.

• Unsupported claim kind: the current contract does not know how to verify the requested proposition.
• Error: processing did not produce a valid evidence result.
• Invalid request: malformed schema, locator, or configuration prevented verification.

These results should not count as grounded and should not be averaged into a neutral score that appears safe.

Overall certification

The report-level all_evidence_grounded gate is intentionally strict. A request is not certified when:

• a required check is not grounded;
• the source fingerprint is stale;
• a claim kind is unsupported;
• a check is semantically unverified under the contract;
• there are no usable supported checks.

An application may derive:

The proof summary is a communication layer. The canonical verification report retains the exact failure evidence.

Failure precedence matters

When multiple issues exist, the verifier needs stable precedence. For example, a stale source is a report-level invalidator. A target that cannot be found should not become a text mismatch merely because expected text was supplied.

Stable precedence improves:

• deterministic CI results;
• support diagnostics;
• alert routing;
• audit explanations;
• compatibility across adapters;
• regression testing.

Do not build application logic by parsing human-readable messages. Use structured status and reason fields.

Application response matrix

Avoid showing internal IDs or sensitive diagnostics directly to end users. Preserve them in controlled audit artifacts.

Examples by evidence type

Quote claim

• element exists and text matches: grounded;
• element exists and text differs: mismatch;
• element missing: not found;
• span required but unavailable: capability blocked.

Table-cell claim

• table and cell resolve with expected value: grounded;
• cell resolves with different value: mismatch;
• table or cell missing: not found;
• tables unsupported: capability blocked.

Region claim

• page and bounded geometry resolve: grounded;
• requested geometry differs materially: mismatch or not found under contract;
• coordinate origin unknown: capability blocked;
• page missing: not found.

Fingerprint-pinned request

• fingerprint matches: continue checks;
• fingerprint differs: stale;
• fingerprint required but unavailable: capability limitation or policy failure.

CI fixtures for every state

Build fixtures for:

1. fully grounded request;
2. text mismatch;
3. missing page;
4. missing element;
5. missing span capability;
6. missing table capability;
7. table-cell mismatch;
8. stale fingerprint;
9. missing fingerprint capability;
10. unknown coordinate origin;
11. unsupported claim kind;
12. malformed citation request;
13. partial report with one reusable check;
14. stale report containing an otherwise grounded check.

Assert status, reason, report-level certification, proof summary, and release action.

How the taxonomy improves operations

• Product: safer final, partial, review, and abstention states.
• Engineering: failures route to parser, generator, adapter, source, or policy owners.
• QA: exact negative fixtures replace vague score thresholds.
• Security: unknown IDs and stale sources fail closed.
• Compliance: reports explain what was and was not proven.
• Users: evidence warnings are more precise than “AI may be wrong.”

Definitive verdict

Ethos detects wrong, missing, and stale citations by treating them as different evidence failures. Mismatch means evidence differs at a resolved target. NotFound means the locator cannot be resolved. Stale means source identity drifted. CapabilityBlocked means the source cannot prove the request.

Use DocuShell Parse PDF for source-aware evidence, Ethos for structured classification, the RAG citation CI/CD guide for regression coverage, and the DocuShell hallucination index for broader model-risk context.

Primary keyword: Ethos wrong missing stale citations
Optimized meta title: How Ethos Detects Wrong and Stale Citations
Optimized meta description: Learn how Ethos classifies mismatched, missing, stale, unsupported, and capability-blocked AI citations.
Proposed URL slug: how-does-ethos-detect-wrong-missing-and-stale-ai-citations