Skip to content

Contradiction Detection

A trust system that can only accumulate positive evidence is not measuring credibility — it is measuring evidence volume. Bukti's posterior is non-monotonic in evidence: contradictions reduce it.

Contradiction detection operates at two layers: a categorical flag layer that surfaces contradictions in the UI, and a score layer that flows penalties through the Beta-Binomial posterior.

Flag layer — four contradiction categories

A contradiction scan returns a list of flags, one per contradicted VOI, classified into the following categories.

Direct factual contradiction

A VOI from a higher-credibility source asserts an inconsistent fact about the same entity and time period. Example: the entity claims employment at Company X across one date range, but an employer-issued signed credential shows departure inside that range.

Capability ceiling violation

A claimed high competence level is directly contradicted by a near-contemporaneous outcome VOI. Example: the entity's profile claims expert-level proficiency in a framework, but a near-contemporaneous outcome VOI from a code review shows fundamentally broken patterns in the same framework.

Temporal impossibility

Two VOIs place the entity in logically impossible simultaneous states. The system detects narrow same-day platform-collision proxies today; full multi-week overlap analysis (detecting impossible full-time employment overlaps, for example) is on the roadmap.

Revocation

The credential issuer has set the relevant bit in their Bitstring Status List (W3C Bitstring Status List v1.0). A revoked credential is the strongest possible negative signal — the issuer themselves is asserting the claim is invalid.

Score layer

Contradicted VOIs contribute negative pseudo-counts to the Beta posterior's β accumulator rather than positive pseudo-counts to α.

The penalty multiplier for revocations is multiplicatively larger than the multiplier for ordinary contradictions, reflecting that revocation is an explicit issuer assertion of invalidity rather than a contextual inconsistency. Both multipliers are tunable parameters held in private config.

Contradictions flow through the same Beta-Binomial machinery as positive evidence. They are not a binary multiplier on the final score. A single strong contradiction against otherwise thin evidence will move the posterior significantly; the same contradiction against a deep body of independent cross-platform evidence will move it less. This is the correct Bayesian behavior.

Roadmap

LLM-based factual contradiction scan and structured capability-ceiling outcome metadata are on the roadmap. Multi-period temporal impossibility (overlapping employment claims spanning weeks) is also a future capability. Today's implementation focuses on the rule-based revocation and same-day platform-collision proxies, which cover the strongest signals.

See limitations.md for what is and is not detected today.