Two coupled items from the CIRISAgent 2.9.6 health/capacity work (the agent-side erroneous-1.0 fix landed as CIRISAgent e09e3eec0; this is the cohort half).
1. Bug: missing evidence defaults to perfect scores
Live prod payload (GET /lens-api/api/v1/scoring/capacity/Ally, 2026-06-11): composite_score: 1.0, all five factors 1.0, at trace_count: 18, confidence: "low" — and fragility_index: 0.999 rendered alongside category: "high_capacity".
Default-to-perfect sites in api/ciris_scoring.py:
- ~line 978: insufficient samples in either drift window →
score = 1.0 (documented in the docstring!)
I_replay = 1.0 placeholder ("no replay detection inputs in v0.5.x")Q_deferral = 1.0 placeholder ("requires WBD tracking")- line 948:
i_chain = 1.0 - gaps/max(total,1) — an 18-row chain with 0 gaps reads as perfect integrity
This is the same bug class the agent just fixed locally (a health probe defaulting absent evidence to healthy=True): absence of evidence is not evidence of perfection. A cold-start/thin corpus should produce an epistemically discounted score (or explicit Indeterminate, which lens-core's ManifoldConformity already models as CohortColdStart), never 1.0. Suggested shape, mirroring the agent fix: multiply a coverage/maturity term into the composite (evidence_count / target_count, capped <1 until the corpus matures), and replace placeholder-1.0 sub-terms with placeholder-EXCLUDED (renormalize) so unimplemented detectors don't inflate the score. A self-consistency check (fragility 0.999 + high_capacity should be impossible to emit together) would have caught this at the boundary.
2. Ask: occurrence-scoped capacity, gated on the CEG consent grant
Post-fold (CIRISAgent#866), consented traces flow CEG-native with per-occurrence agent_id_hash — the lens already enumerates (agent_id_hash, agent_name) pairs internally (api/scoring_api.py:241) but only exposes template aggregates (/capacity/{agent_name}) and fleet (/capacity/fleet).
Ask: GET /scoring/capacity/agent/{agent_id_hash} — THIS occurrence's externally-evaluated score, served only when the occurrence's consent:community_trust:v1 grant is in force (the CEG wire artifact written at opt-in; withdraws/recants = stop serving + the deletion cascade). Serving an occurrence-identified score is itself a privacy surface, so the consent gate applies to reads, not just ingest.
Agent-side consumption is designed and waiting: the /v1/my-data/capacity endpoint will add scope=occurrence alongside fleet/local — the trust story becomes: local self-vitals (epistemically capped at 0.95/0.60 by trust edge), occurrence cohort-grade score (external evaluation, consent-gated), fleet aggregate. cc CIRISAgent#866, CIRISLensCore#43.
Two coupled items from the CIRISAgent 2.9.6 health/capacity work (the agent-side erroneous-1.0 fix landed as CIRISAgent e09e3eec0; this is the cohort half).
1. Bug: missing evidence defaults to perfect scores
Live prod payload (
GET /lens-api/api/v1/scoring/capacity/Ally, 2026-06-11):composite_score: 1.0, all five factors 1.0, attrace_count: 18,confidence: "low"— andfragility_index: 0.999rendered alongsidecategory: "high_capacity".Default-to-perfect sites in
api/ciris_scoring.py:score = 1.0(documented in the docstring!)I_replay = 1.0placeholder ("no replay detection inputs in v0.5.x")Q_deferral = 1.0placeholder ("requires WBD tracking")i_chain = 1.0 - gaps/max(total,1)— an 18-row chain with 0 gaps reads as perfect integrityThis is the same bug class the agent just fixed locally (a health probe defaulting absent evidence to healthy=True): absence of evidence is not evidence of perfection. A cold-start/thin corpus should produce an epistemically discounted score (or explicit Indeterminate, which lens-core's ManifoldConformity already models as
CohortColdStart), never 1.0. Suggested shape, mirroring the agent fix: multiply a coverage/maturity term into the composite (evidence_count / target_count, capped <1 until the corpus matures), and replace placeholder-1.0 sub-terms with placeholder-EXCLUDED (renormalize) so unimplemented detectors don't inflate the score. A self-consistency check (fragility 0.999 + high_capacity should be impossible to emit together) would have caught this at the boundary.2. Ask: occurrence-scoped capacity, gated on the CEG consent grant
Post-fold (CIRISAgent#866), consented traces flow CEG-native with per-occurrence
agent_id_hash— the lens already enumerates(agent_id_hash, agent_name)pairs internally (api/scoring_api.py:241) but only exposes template aggregates (/capacity/{agent_name}) and fleet (/capacity/fleet).Ask:
GET /scoring/capacity/agent/{agent_id_hash}— THIS occurrence's externally-evaluated score, served only when the occurrence'sconsent:community_trust:v1grant is in force (the CEG wire artifact written at opt-in; withdraws/recants = stop serving + the deletion cascade). Serving an occurrence-identified score is itself a privacy surface, so the consent gate applies to reads, not just ingest.Agent-side consumption is designed and waiting: the
/v1/my-data/capacityendpoint will addscope=occurrencealongside fleet/local — the trust story becomes: local self-vitals (epistemically capped at 0.95/0.60 by trust edge), occurrence cohort-grade score (external evaluation, consent-gated), fleet aggregate. cc CIRISAgent#866, CIRISLensCore#43.