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

[chrishooooo-netizen]

Summary

Addresses Issue #2 (Trust Score Algorithm) and Amey Parle's technical critique from 2026-04-13 (E-010..E-014 in the Challenge Register).

The current v1.2.0-draft Trust Score is count-weighted and Sybil-gameable: a coordinated cluster of 5 controlled issuers can drive D5 to 1.0 in weeks. This PR makes the score evidence-weighted along four axes.

Changes

Section	Change
§7.2.5 Probationary Tier (new)	New DIDs start with cap = 0.5 + 0.5 × min(1, interactions/100), min 30 days. No inheritance from issuer — explicitly retracts the "Partial Trust Inheritance" answer to E-002.
§7.3.2 Aggregate	S = m × Σ(wi × di) × (1 - p), capped by §7.2.5. Maturity, anomaly, and probationary effects compose.
§7.3.2 D5	Reformulated: Issuer Independence Score (cluster-discounted) + Diminishing Returns (1 - exp(-k × independence), k=0.45). The previous linear cap is deprecated.
§7.3.7 Anomaly Penalties (new)	5 rule-based graph detectors (A1-A5: circular endorsement, cluster collusion, velocity anomaly, interaction burst, co-registration cluster). ML capped at +0.20 contribution to preserve EU AI Act Art. 13/14 explainability.
§7.3.8 Maturity Multiplier (new)	m = min(1.0, age_days / 180).
§7.3.9 Signal Decay (new)	D1 step-function to 0.4 on re-verification lapse, D2 unchanged (12m rolling), D3/D4 exponential HW=30d, D5 linear over VC validity.

Sybil Gameability

• Before: 5 controlled issuers → D5 = 1.0 in weeks. Estimated gameability 8/10.
• After: Cluster collusion detected (A2: +0.40 penalty), independence-weighted D5 contributes ≈0.2 from same cluster, maturity multiplier holds score below 1.0 for 180 days, probationary cap holds early score ≤ 0.5. Estimated gameability 3/10.

Open for review

@AmeyParle — your memo from 2026-04-13 is the foundation for this PR. Specific points where I'd value your judgment:

1. §7.3.7 detector thresholds (A1-A5 penalty contributions) — calibration is a guess; I'd rather have your read.
2. §7.3.2 D5 k=0.45 — chosen so 5 fully-independent attestations yield ≈0.9. Reasonable?
3. §7.3.9 D3/D4 half-life of 30 days — too aggressive? 60 days?
4. §7.3.7 A2 cluster threshold (60%) — could be tuned.

Test plan

• Spec renders cleanly in markdown viewer
• Cross-references (§7.2.5 ↔ §7.3.2 ↔ §7.3.7-9) resolve
• No regressions to D1/D2 dimension definitions
• Resolution metadata example in §7.2.5 still validates against schema
• Reviewer (Amey) sign-off on detector calibration before merge

Related

• Closes part of RFC: Trust Score Algorithm - Seeking Expert Critique #2 (Trust Score Algorithm)
• Companion PR for Issue RFC: Federation Model Architecture Review #1 (Federation + Revocation Propagation §3.4 + §8.7) follows separately