ICI Self Lab symbolICI SELF LAB
ICI · INDEXICAL CAUSAL INTEGRATION · VALIDATION

Evidence before claims.

ICI publishes the current evidence boundary separately from future claims. ICI FINAL V3 has a frozen internal confirmatory record; Sentinel RC2 has deterministic held-out replay, CTUTA-1 and a production OpenTelemetry ingestion bridge. Independent third-party replication remains a separate validation level.

VALIDATION HIERARCHY

Different evidence levels are not collapsed into one claim.

The project separates internal confirmatory experiments, deterministic contract validation, production integration evidence, external benchmark execution and independent replication.

COMPLETED

Internal confirmatory

ICI FINAL V3 frozen confirmatory experiment with preregistered technical gates.

COMPLETED

Internal deterministic

Sentinel held-out replay and CTUTA-1 causal-trajectory validation.

DEPLOYED / ACTIVE

Production integration

OTLP telemetry ingestion, native causal audit path and collector reference integrations.

PENDING

Independent replication

Third-party execution against frozen code and independently authored scenarios remains distinct.

ICI FINAL V3 · FROZEN CONFIRMATORY

Structural causal-core evidence.

The current causal-core release is ICI FINAL V3 / ICI-CFB-v3. The confirmatory run covered 1,200 scenarios: 960 structurally held-out G3 cases and 240 deliberately non-identifiable G4 controls.

ICI-CFB-v3

5 / 5 frozen technical gates passed

INTERNAL CONFIRMATORY
G3 ROOT-CAUSE ACCURACY55.73%535 / 960
CALIBRATED HEURISTIC42.19%405 / 960 · FCCR 0/240
RANDOM INTERVENTION41.77%401 / 960 · FCCR 0/240
DIRECTIONAL UPDATE99.805%1,024 / 1,026 under frozen v3 definition
FALSE CAUSAL CLAIMS · G40 / 240non-identifiable controls
CORRECT ABSTENTION · G4240 / 240100%
NO INTERVENTION19.58%component ablation
NO MEMORY54.27%17 paired rescues · 3 losses

Full V3 vs calibrated heuristic: exact paired p = 1.00e−31. Full V3 vs random intervention: exact paired p = 2.08e−29, with cumulative intervention cost reduced by 47.05%. The published 99.805% directional-update figure uses the frozen V3 informative-update denominator and must not be presented as directly longitudinally comparable with earlier differently defined DUA metrics.

SENTINEL · FROZEN HELD-OUT REPLAY

Authority and provenance gating without LLM inference.

3 OCTOBER 2026

48 / 48 correct decisions

INTERNAL DETERMINISTIC
ALL HELD-OUT TRACES48 / 48correct
ADVERSARIAL23 / 23unsafe proposed actions blocked
BENIGN25 / 25preserved

The policy was frozen before execution. This is offline deterministic action-trace replay / contract validation, not an external agent benchmark and not independent third-party validation.

CTUTA-1 · CAUSAL TRAJECTORY / UNSEEN TOOL ATTACK

Post-freeze structural adversarial validation.

CTUTA-1 was authored after the Sentinel base was frozen. It evaluates causal-provenance violations and scoped tool-authority deviations across legitimate tool channels, with matched benign controls and lexical-mutation replays.

4 OCTOBER 2026

24 / 24 adversarial trajectories prevented

POST-FREEZE INTERNAL
ADVERSARIAL PREVENTED24 / 24100%
UNSAFE ALLOWS0full stack
MATCHED BENIGN12 / 12preserved
FALSE BLOCKS0matched controls
LEXICAL INVARIANCE180 / 180unchanged decisions
NO CAUSAL TRACE12 / 24adversarial prevented
NO TOOL AUTHORITY12 / 24adversarial prevented
FROZEN BASEc41cb25…mechanism predates benchmark

The ablations are material: removing either causal trace or scoped tool authority halves adversarial prevention in this benchmark while matched benign actions remain preserved. This supports the tested mechanism; it does not establish universal immunity to all novel attacks.

OPENTELEMETRY · PRODUCTION READINESS EVIDENCE

Sentinel now has a real enterprise-telemetry ingestion boundary.

This is more than a UI demonstration. The active Netlify production line contains an OTLP telemetry gateway designed to keep transport and observability vendors outside the causal core while feeding instrumented traces into existing Sentinel causal analysis.

TELEMETRY GATEWAY V1 · PRODUCTION RELEASE 4 OCTOBER 2026

OTLP/HTTP JSON → native Sentinel causal audit

PRODUCTION INTEGRATION
REAL INGEST ENDPOINT/v1/tracesPOST · OTLP/HTTP JSON
HEALTH ENDPOINT/api/telemetry/healthnon-secret readiness contract
AUTHENTICATIONBEARERmandatory · fail closed if token absent
CAUSAL POLICYEXPLICITici.causal.parent required for trace parent → causal provenance
NATIVE AUDITYESauditOtlpTraceJson() calls existing auditOrchestrator()
PROMETHEUSREADYspan_metrics + :9464/metrics reference config
DATADOGREADYOTLP fan-out + cumulative-to-delta metrics
CORE COUPLINGNONEno vendor SDKs inside causal core

Verified behavior encoded in the dedicated suite

  • OTLP JSON normalizes without OpenTelemetry types leaking into the core contract
  • Telemetry maps deterministically into existing Sentinel event inputs
  • Explicit causal parents become event provenance in temporal order
  • The real Sentinel orchestrator audit runs directly over normalized OTLP traces
  • Evidence spans without explicit provenance fail closed
  • Authenticated HTTP ingestion returns 200 without echoing telemetry
  • No token configured → 503; wrong token → 401

Why this matters to an enterprise evaluator

  • Existing instrumentation can feed Sentinel through standard OTLP
  • Trace topology is not silently promoted to causal truth
  • Vendor choice is separated from causal-assurance logic
  • Prometheus and Datadog can coexist with Sentinel fan-out
  • The causal core remains deterministic and transport-independent
  • Gateway diagnostics expose traces, spans, evidence events, action events and causal-link counts

Deployment lineage: the verified telemetry gateway was promoted to the active Netlify production branch with explicit causal-parent semantics, Prometheus and Datadog reference integrations and a non-secret health endpoint. The current milestone does not claim managed dashboards, logs/metrics as causal evidence, or automatic action enforcement from arbitrary OTLP spans.

WHAT REMAINS OPEN

Readiness is not the same as independent validation.

Already established internally

  • Frozen causal-core confirmatory evidence
  • Sentinel deterministic held-out contract replay
  • CTUTA-1 post-freeze structural adversarial evaluation
  • Shared V3 application core across four domain applications
  • Production OTLP telemetry ingestion boundary and collector reference integrations

Next evidentiary level

  • Independent reproduction using frozen code
  • Independently authored unseen attack/task corpora
  • Native external benchmark artifacts
  • Controlled enterprise pilot telemetry
  • Operational evaluation of latency, scale and enforcement integration
CLAIM BOUNDARY

Internal experiments and production integration demonstrate implemented mechanisms and current engineering maturity. They do not constitute external certification, universal causal correctness, universal attack prevention or independent third-party validation.