Divergence
Divergence is what happens when the evidence does not agree. Two sources make contradictory claims about the same fact. A time series has a gap where there should be a record. A dissenting interpretation of an instruction exists alongside the primary interpretation.
AIEP does not resolve divergence by discarding the minority. It records it — deterministically, cryptographically, and in a form that can be retrieved and acted on when new evidence arrives.
Types of divergence
| Type | What it means |
|---|---|
| Evidential divergence | Two artefacts make contradictory claims about the same fact |
| Temporal divergence | A time-ordered evidence set has a gap — a period with no records where records are expected |
| Interpretive divergence | An instruction has alternative interpretations, none of which is definitively ruled out by current evidence |
| Branch divergence | A reasoning process has taken a fork — two branches are both active and neither has been resolved to dominance |
Each type is handled by a distinct mechanism, but all share the same underlying principle: detection and recording, not suppression and discard.
Evidential divergence (P37)
When two evidence artefacts are found to make contradictory claims about the same subject, the divergence detector:
- Confirms both artefacts are schema-valid and hash-verified
- Identifies the specific fields in which the contradiction exists
- Generates a DivergenceRecord — a typed record naming both artefacts, the contradicted fields, and the nature of the contradiction
- Appends the DivergenceRecord to the evidence ledger
- Routes both artefacts to the dissent archive — neither is discarded
The DivergenceRecord is hash-bound to both artefacts. Future retrieval agents can find all divergence records related to a given artefact by querying the ledger.
What the DivergenceRecord contains:
| Field | Contents |
|---|---|
artefactA | Hash reference to first artefact |
artefactB | Hash reference to contradicting artefact |
contradictedFields | The specific fields where contradiction exists |
divergenceType | Typed classification of the contradiction |
timestamp | When the divergence was detected |
detectorVersion | Schema version of the divergence detection rules |
Temporal divergence (P16)
Evidence is often time-ordered — logs, compliance records, sensor data, financial transactions. A gap in a time series is not just missing data. It is potentially significant: an unrecorded period during which something may have happened without a trace.
The temporal divergence detector constructs a TimelineIndex from a set of time-indexed evidence records:
- Each record is normalised and its timestamp reduced to a canonical
TimeIndexvalue - Records are sorted deterministically: primary by TimeIndex, secondary by CanonicalHash as a stable tie-break
- Adjacent records in the sorted index are evaluated against a GapRuleProfile
- If the interval between successive records exceeds the permitted threshold, a GapArtefact is generated
A GapArtefact is not an error. It is a first-class evidence record — cryptographically bound, timestamped, and appended to the ledger. It records that a gap exists, where it is, how large it is, and under which GapRuleProfile version the gap was detected.
If a downstream operation requires a complete timeline and a disallowed GapArtefact exists, execution is suppressed. The gap is not ignored — it gates further processing until resolved.
Dissent fork generation (P83)
When a reasoning process encounters a sustained divergence — a divergence that persists above a threshold magnitude for a period above a threshold duration — it does not arbitrarily resolve the divergence. It generates a dissent fork.
A dissent fork is a controlled split of the reasoning state into two branches:
| Element | Description |
|---|---|
| Primary branch | The current consensus interpretation, continuing |
| Dissent branch | The alternative interpretation, archived with its evidence chain |
| RecallScope | The deterministic specification of which archived entries are eligible for reactivation |
| RecallScopeHash | Cryptographic hash over the recall configuration |
| Termination record | Generated when the frontier constraint is reached — bounding the fork’s expansion |
The dissent fork is bounded. It cannot expand indefinitely — the frontier constraint limits the computational depth of the fork. When the frontier is reached, a cryptographically verifiable termination record is appended and the fork is closed. The fork is archived, not deleted. Its RecallScope is preserved so it can be reopened if new evidence meets the sustained divergence threshold again.
Sustained divergence threshold:
A fork is only generated when divergence magnitude and divergence duration both exceed their declared thresholds. This prevents spurious forks from transient noise — only persistent, significant divergence triggers a fork.
Divergence and the dissent loop
Divergence detection is the mechanical engine behind the dissent loop. The loop describes the philosophy — preserve outliers so they can be recalled when new evidence arrives. Divergence is the implementation:
- Evidential divergence → DivergenceRecord → both artefacts archived
- Temporal divergence → GapArtefact → timeline completeness check before execution
- Dissent fork → bounded branch archived with RecallScope → reopenable when new evidence meets threshold
Together, these mechanisms mean that AIEP never silently discards a contradiction. Every disagreement in the evidence is a record that can be found, retrieved, and re-evaluated.
Divergence is not failure
A system with no recorded divergences is not necessarily a trustworthy system — it may simply be one that discards contradictions rather than recording them. A system with recorded divergences is operating correctly: it is acknowledging that the evidence is not complete or not unanimous, and preserving that acknowledgement for future retrieval.
The divergence ledger is therefore not a list of errors. It is a structured record of what the system does not yet know, complete with enough information to revisit those questions when more is known.
Architecture of Knowing · Evidence normalisation · Constitutional substrate · Recall · Plausibility Matrix · Patents — P16, P22, P37, P42, P83, P94