Executive summary
Supervisory attention is moving from post-hoc monitoring alone toward lifecycle controls, human oversight, and evidence that decisions can be traced to approved anchors before they bind to systems of record. This page pairs two explainer videos with the repository skeleton for The State of Admissibility 2026 — use the Markdown for PDF export and auditor mapping tables.
Observability vs. admissibility
Archetypal comparison — named vendors (e.g. Arize, LangSmith, Honeycomb) are examples of the observability column, not a feature-by-feature audit. See white paper Appendix E.
| Feature | Traditional observabilitye.g. Arize, LangSmith, Honeycomb | BiDigest Trustee-tierAdmissibility |
|---|---|---|
| Primary goal | Model Performance: Debugging, evaluation, and drift detection. | Execution Integrity: Enforcement, SoD, and forensic admissibility. |
| Operational timing | Reactive: Analyzes traces after the LLM has already responded. | Proactive: Intercepts intent before the state transition (Commit Boundary). |
| Control mechanism | Alerting: Notifies a human if a “guardrail” is breached. | Mechanical Gating: Fails-closed at the DB layer if the “double-lock” is missing. |
| Trust foundation | Statistical: Based on confidence scores and semantic similarity. | Deterministic: Based on cryptographic hashes and versioned Anchor Prose. |
| Audit artifact | Logs/Traces: High-fidelity records of what the model said. | Forensic Receipt: Merkle-sealed proof of who authorized the action and why. |
| SoD enforcement | None: Does not distinguish between Proposer and Approver at the API level. | Hard-Coded: Schema-level constraints prevent self-approval of sensitive actions. |
| Regulatory fit (illustrative) | NIST (Measure): Good for “measuring” model risk. | MAS/FCA/EU AI Act (Manage): Designed for “lifecycle control” and “human oversight.” |
Primary goal
Observability
Model Performance: Debugging, evaluation, and drift detection.
BiDigest admissibility
Execution Integrity: Enforcement, SoD, and forensic admissibility.
Operational timing
Observability
Reactive: Analyzes traces after the LLM has already responded.
BiDigest admissibility
Proactive: Intercepts intent before the state transition (Commit Boundary).
Control mechanism
Observability
Alerting: Notifies a human if a “guardrail” is breached.
BiDigest admissibility
Mechanical Gating: Fails-closed at the DB layer if the “double-lock” is missing.
Trust foundation
Observability
Statistical: Based on confidence scores and semantic similarity.
BiDigest admissibility
Deterministic: Based on cryptographic hashes and versioned Anchor Prose.
Audit artifact
Observability
Logs/Traces: High-fidelity records of what the model said.
BiDigest admissibility
Forensic Receipt: Merkle-sealed proof of who authorized the action and why.
SoD enforcement
Observability
None: Does not distinguish between Proposer and Approver at the API level.
BiDigest admissibility
Hard-Coded: Schema-level constraints prevent self-approval of sensitive actions.
Regulatory fit (illustrative)
Observability
NIST (Measure): Good for “measuring” model risk.
BiDigest admissibility
MAS/FCA/EU AI Act (Manage): Designed for “lifecycle control” and “human oversight.”
Commit boundary (Slice A)
Mechanical segregation of duties: intent is anchored, human-approved with hash match, then executed—fail-closed by design.
Source asset: docs/Posting/Img&Vids/Commit boundary.mp4 · mirrored under public/ for web delivery.
Triple Lock — Legal · Risk · Engineering
How decision rights show up at the intercept: the same verdict semantics governance expects, not a rubber stamp.
Source asset: docs/Posting/Img&Vids/Video_Triple_Lock.mp4 · mirrored under public/ for web delivery.
Full narrative & appendices
Chapters on IFQ, Merkle forensic batches, jurisdictional sovereignty, glossary, and bibliography live in docs/whitepapers/STATE_OF_ADMISSIBILITY_2026_SKELETON.md. Sync videos after edits: npm run whitepaper:sync-admissibility-videos.