feat(spec): evidence-weighted Trust Score (Issue #2)

spec/did-method-trail-v1.md

@@ -862,6 +862,49 @@ Self-signed credentials MUST include `"trailTrustTier": 0` in the credential sub
 | EU AI Act support | Minimal | Standard | Full |
 | Trust Score | N/A | Computed | Computed + audited |
 
+#### 7.2.5 Probationary Tier
+
+New DIDs registered in the TRAIL ecosystem MUST enter a **probationary state** before becoming eligible for their full computed Trust Score. This prevents trust laundering, where a reputable issuer could grant a newly registered agent an unearned high score through inheritance.
+
+**Probationary State Rules:**
+
+- A DID enters the probationary state at the moment of registration.
+- While in probationary state, the effective Trust Score is **capped** independent of the raw computed value:
+
+  ```
+  effective_score = min(computed_score, probationary_cap)
+  probationary_cap = 0.5 + 0.5 × min(1, verified_interactions / 100)
+  ```
+
+- The minimum probationary duration is **30 days** from registration, regardless of interaction count.
+- A DID exits the probationary state when BOTH conditions are met:
+  1. `verified_interactions >= 100`
+  2. `age_days >= 30`
+
+- **Issuer reputation does NOT bypass the cap.** A high-reputation issuer contributes only to dimension D1 (Identity Verification), not to the probationary cap. There is no "partial trust inheritance" from issuer to subject.
+
+**Resolution Metadata:**
+
+Registries MUST expose the probationary state in resolution responses:
+
+```json
+{
+  "trailTrustScore": {
+    "overall": 0.50,
+    "computedOverall": 0.82,
+    "probationary": {
+      "state": "probationary",
+      "cap": 0.50,
+      "ageDays": 12,
+      "verifiedInteractions": 23,
+      "exitsEarliest": "2026-05-14T00:00:00Z"
+    }
+  }
+}
+```
+
+Verifiers MUST use `overall` (the effective capped score), not `computedOverall`, when making trust decisions. `computedOverall` is exposed for transparency only.
+
 ### 7.3 Trust Score
 
 The TRAIL Trust Score quantifies the trustworthiness of a registered identity across five independently verifiable dimensions.
@@ -881,10 +924,13 @@ The TRAIL Trust Score quantifies the trustworthiness of a registered identity ac
 The overall Trust Score `S` is computed as:
 
 ```
-S = Σ(wi × di) for i = 1..5
+S_raw      = Σ(wi × di) for i = 1..5
+S_mature   = m × S_raw                        // see §7.3.8 Maturity Multiplier
+S          = S_mature × (1 - p)               // see §7.3.7 Anomaly Penalties
+effective  = min(S, probationary_cap)          // see §7.2.5 Probationary Tier
 ```
 
-Where `wi` is the weight and `di` is the dimension score (0.0–1.0) for each dimension.
+Where `wi` is the weight and `di` is the dimension score (0.0–1.0) for each dimension, `m` is the maturity multiplier (§7.3.8), `p` is the anomaly penalty factor (§7.3.7), and `effective` is the score verifiers MUST use for trust decisions. Per-dimension decay is defined in §7.3.9.
 
 **Dimension Formulas:**
 
@@ -916,10 +962,22 @@ Where `wi` is the weight and `di` is the dimension score (0.0–1.0) for each di
   Where `compliant_checks` is the number of automated policy compliance checks that passed. Measured over trailing 90 days.
 
 - **D5 (Third-Party Attestations):**
-  ```
-  d5 = min(1.0, valid_attestations / 5)
-  ```
-  Where `valid_attestations` is the count of non-expired, non-revoked Verifiable Credentials from distinct accredited attestors. Capped at 5.
+
+  D5 is **evidence-weighted, not count-weighted**. Two effects compose:
+
+  1. **Issuer Independence Score** — attestations from issuers in the same governance/control cluster are discounted:
+     ```
+     independence = Σ (1 / cluster_size_i) for each attesting issuer i
+     ```
+     Where `cluster_size_i` is the number of issuers known to share a control cluster with issuer `i` (including itself). Two issuers operated by the same legal entity, or sharing >50% governance overlap as recorded in their `did:trail:org` documents, count as one cluster. An attestation from a singleton issuer contributes `1.0`; one from a 5-issuer cluster contributes `0.2`.
+
+  2. **Diminishing Returns** — additional independent attestations contribute progressively less:
+     ```
+     d5 = 1 - exp(-k × independence)
+     ```
+     Where `k = 0.45` (calibrated so ~5 fully independent attestations yield d5 ≈ 0.9). This replaces the previous linear cap.
+
+  Where `valid_attestations` are non-expired, non-revoked Verifiable Credentials from distinct accredited attestors. Cluster membership is computed from the issuers' published `did:trail:org` governance metadata; absent metadata, registries SHOULD treat issuers sharing a `controlledBy` field as one cluster. The previous formula `d5 = min(1.0, valid_attestations / 5)` is deprecated as of v1.2.0 — it permitted Sybil attacks via controlled issuer clusters (see §7.3.7).
 
 #### 7.3.3 Score Transparency
 
@@ -1013,6 +1071,50 @@ The trust score model has inherent limitations that verifiers MUST be aware of:
 - **Gaming resistance:** The formula-based approach creates potential for strategic behavior optimization. The TRAIL Registry SHOULD implement anomaly detection for sudden score changes and MAY require minimum observation periods before scores are considered stable.
 - **D5 bootstrapping:** Early in the ecosystem, few accredited auditors may exist, limiting the practical value of D5. The protocol anticipates this through the three-phase governance model (§11.1).
 
+#### 7.3.7 Anomaly Penalties
+
+The Trust Score MUST be discounted multiplicatively by an anomaly penalty factor `p ∈ [0, 1]` derived from rule-based graph analysis of the attestation and interaction graph. Rule-based detection forms the normative core; machine-learning classifiers MAY be used as a secondary signal but MUST NOT solely determine `p`.
+
+**Rule-based detectors (normative):**
+
+| Code | Detector | Penalty Contribution |
+|------|----------|----------------------|
+| A1 | **Circular endorsement** — DID A attests B, B attests A within 30 days; or any closed cycle of length ≤ 4 in the attestation graph | +0.30 per detected cycle (max +0.50) |
+| A2 | **Cluster collusion** — ≥ 60% of a DID's attestations originate from issuers in a single control cluster (see §7.3.2 D5) | +0.40 |
+| A3 | **Velocity anomaly** — D5 grows by > 5 independent attestations within any 7-day window during the first 90 days of the DID's lifetime | +0.20 |
+| A4 | **Interaction burst** — D2 `interactions_12m` increases by > 10× the trailing 30-day average within 24 hours, with > 50% of new counterparties sharing a control cluster | +0.30 |
+| A5 | **Co-registration cluster** — ≥ 5 DIDs registered within a 60-minute window from a single registrar account, all attesting one another within their first 30 days | +0.50 |
+
+`p = min(1.0, Σ contributions)`. Detected anomalies MUST be exposed in the raw inputs endpoint (§7.3.4) under an `anomalies` field with the detector code, evidence, and detection timestamp. A DID with `p ≥ 0.5` MUST be flagged in resolution metadata (`"anomalyFlag": true`).
+
+**Machine-learning role:** Registries MAY run ML-based anomaly classifiers (graph neural networks, embedding-based clustering) and MAY surface their findings as additional `anomalies` entries with detector code `M*`, but ML findings MUST NOT contribute more than 0.20 to `p`. The rule-based core ensures explainability and auditability for EU AI Act Art. 13/14 compliance.
+
+#### 7.3.8 Maturity Multiplier
+
+The Trust Score MUST be scaled by a time-based maturity multiplier `m`:
+
+```
+m = min(1.0, age_days / 180)
+```
+
+Where `age_days` is the number of days since the DID was first registered in any TRAIL Registry. A DID that has existed for less than 180 days cannot reach a Trust Score of 1.0, even with perfect dimension scores. This complements §7.2.5 (Probationary Tier) by extending the trust-building period beyond the probationary cap exit.
+
+The maturity multiplier MUST be exposed in resolution metadata as `"maturity": { "ageDays": N, "multiplier": M }`.
+
+#### 7.3.9 Signal Decay
+
+Each Trust Score dimension decays differently based on signal type. Decay is applied to each `di` *before* aggregation in §7.3.2.
+
+| Dimension | Decay Type | Specification |
+|-----------|------------|---------------|
+| **D1** Identity Verification | Step function | Full value until annual re-verification deadline; on the day after the deadline, `d1` drops to `0.4` (the self-declared baseline) until re-verification is recorded. |
+| **D2** Track Record | Rolling window | 12-month rolling window (unchanged from §7.3.2). No additional decay. |
+| **D3** Information Provenance | Exponential | Half-life of **30 days**: `d3_decayed = d3_observed × 0.5^(days_since_observation / 30)`. Behavioral signals must reflect current practice. |
+| **D4** Behavioral Consistency | Exponential | Half-life of **30 days** (same formula as D3). |
+| **D5** Third-Party Attestations | Linear over VC validity | Each attestation contributes its full independence-weighted value at issuance, decaying linearly to zero across `validFrom`–`validUntil`. Expired or revoked attestations contribute zero. |
+
+Registries MUST recompute decayed dimension scores at least once per 24 hours and on every resolution request that occurs more than 1 hour after the last computation.
+
 ### 7.4 EU AI Act Alignment
 
 TRAIL provides technical infrastructure that organizations can use to support their compliance efforts with the EU AI Act. This section maps TRAIL capabilities to specific EU AI Act requirements.