Evidence Weight

Evidence weight is the formal measure of how strongly an evidence artefact supports a claim — and how that strength changes over time as the evidence ages, is challenged, or is corroborated by new independent sources.

This is defined in P20 and P21.

Why weight matters

Not all evidence is equal. A single underpowered study supports a claim less strongly than a systematic review of ten. A compliance certificate issued yesterday carries more weight than one issued three years ago. A claim backed by five independent organisations has more weight than one backed by one.

Without a formal weight model, a retrieval agent cannot distinguish between weak and strong support. It retrieves what is available and acts as if all of it is equally operative.

With evidence weight, the retrieval agent can:

Report confidence proportional to the evidence base
Flag claims that rest on ageing or single-source evidence
Decline to execute a GoalVector instruction when the supporting evidence falls below the plausibility threshold

The weight model

Every evidence artefact carries an EvidenceWeightVector with four dimensions:

Dimension	What it measures	Range
`source_authority`	How authoritative the issuing organisation is for this evidence type, as declared in the registry	0.0 – 1.0
`corroboration_score`	How many independent sources support the same claim (independence verified via registry)	0.0 – 1.0
`temporal_decay`	Weight reduction due to age, computed against the domain’s decay curve	0.0 – 1.0
`challenge_penalty`	Weight reduction from unresolved `DivergenceRecord` events referencing this artefact	0.0 – 1.0

The composite weight:

Field	Value
`composite_weight`	`source_authority × corroboration_score × temporal_decay × (1 − challenge_penalty)`
`weight_computed_at`	ISO 8601 timestamp of most recent computation
`weight_version`	Version of the weight algorithm applied

Temporal decay (P20)

Evidence ages. The temporal decay component models this formally:

Artefact age	Default decay factor	Notes
0–6 months	1.00	No decay — full weight
6–12 months	0.95	Negligible decay
1–2 years	0.85	Moderate decay — still highly operative
2–5 years	0.70	Significant decay — corroboration increasingly important
5+ years	0.50	High decay — requires strong corroboration to remain in execution band
Domain override	As declared in `normalisation_profile`	Medical: faster decay. Legal precedent: slower decay. Mathematical proof: near-zero decay

Decay does not invalidate an artefact. It reduces its weight. An artefact with high temporal decay but a high corroboration_score — many independent sources citing the same old evidence — may still maintain a composite_weight in the execution band.

The decay curve parameters are declared in the domain’s normalisation_profile. This allows each domain to set its own evidence half-life without changing the weight algorithm.

Corroboration (P21)

Corroboration increases weight. When multiple independent issuers publish artefacts supporting the same claim, the corroboration score rises:

Independent sources	Corroboration score
1	0.60
2	0.75
3	0.85
4	0.92
5+	0.98

Independence is verified through the issuer registry. Two artefacts from the same registered organisation — even different departments — do not count as independent corroboration. The registry maps legal entity relationships to prevent corroboration gaming.

Weight and the plausibility gate

The composite_weight feeds directly into the plausibility gate (P03). A GoalVector instruction cannot enter the execution band unless the evidence supporting it meets the minimum composite_weight for its instruction type:

Composite weight	Band	Effect
≥ 0.85	`EXECUTABLE`	Proceeds to probability certification (P04)
0.60 – 0.84	`REVIEW`	Flagged for human review before execution proceeds
< 0.60	`BLOCKED`	Routed to dissent archive with weight record preserved

This is the mechanical connection between evidence quality and AI execution safety. A model cannot act on weakly supported evidence without a governance override recorded in the audit trail.

Weight decay and recall

When a claim’s composite_weight drops below the EXECUTABLE threshold due to temporal decay, it transitions to OUTLIER state in the knowledge state machine. It is not deleted or discarded — it enters the dissent archive with its full EvidenceWeightVector history preserved.

If new corroborating evidence arrives later — a new independent study, a new trial, a new independent confirmation from a different organisation — the corroboration_score rises, the composite_weight recovers, and the claim may be recalled to CONSENSUS state.

This is the formal mechanism behind AIEP’s recall capability. A claim does not become permanently non-executable because its evidence aged. It becomes recall-eligible — waiting for new evidence to restore its weight above threshold.

Patents

P20 — Evidence Weight and Temporal Decay Modelling. Filed November 2025, UK IPO.
P21 — Corroboration Score and Independent Source Verification. Filed November 2025, UK IPO.

Plausibility Matrix — how weight feeds the execution gate
Recall — how weight recovery triggers recall from the dissent archive
Evidence Normalisation — normalisation profiles that set decay parameters
Evidence Stitching — how multi-source evidence is assembled before weighting
Divergence — how challenge penalties are applied
Patents — P20, P21