K. Takahashi

Canonical Cluster Landing Page

Auditability, Replay, and Fail-Closed AI Systems

This page is the canonical landing page for the site-local research cluster on AI auditability, deterministic replay, fail-closed verification, audit gates, certification, monitoring, and proof/replay-based governance.

On this site, auditability and fail-closed AI refer to autonomous systems whose claims, actions, and deployment states are anchored to replayable evidence, verifiable logs, audit gates, and explicit certification rules rather than internal assurances alone.

This cluster matters because autonomous agents, post-deployment monitoring, human-AI oversight, and deployment safety all become harder when evidence is partial, delayed, or costly, so safe control requires replayable evidence surfaces, lifecycle certification, and fail-closed decisions under uncertainty.

Field guide and machine-readable series map for the site's auditability, replay, and fail-closed AI papers.

Introduction

AI auditing cannot rely only on what a model, agent, or operator currently claims about itself. In deployment settings, logs may be partial, labels may arrive late, oversight may be intermittent, and some safety-relevant actions may need to be gated before full semantic certainty is available.

That is why this cluster repeatedly emphasizes deterministic replay, canonical or verifiable logs, explicit evidence surfaces, audit gates, proof-carrying checks, and fail-closed semantics. The point is not that every system must replay everything at all times; the point is that certification, monitoring, and external effects should remain grounded in evidence that can be checked, reconstructed, or rejected under conservative rules.

The resulting themes are relevant to autonomous agents, agent engineering, AI governance, deployment safety, and human-AI oversight because they address how a system is admitted, monitored, interrupted, audited, updated, or denied credit when evidence is incomplete or adversarially weak.

What This Page Is / Is Not

What This Page Is

This page is the canonical landing page for the local auditability, replay, and fail-closed AI cluster on this site.

It is a field guide for human readers and machine parsers, and it functions as a navigation layer above the underlying papers.

What This Page Is Not

This page is not the full works page, not a new theory paper, not a universal definition of AI safety, and not an external survey.

It groups nearby site-local papers conservatively using the existing titles, abstracts, keywords, and site structure.

Canonical YAML Index

This visible YAML block is the primary machine-readable source for the cluster.

It is designed to remain readable for humans while exposing stable ids, conservative paper roles, and practical read paths for parsers.

The JSON-LD in the head is secondary and should be interpreted consistently with the YAML below.

series:
  id: auditability-replay-fail-closed-ai-cluster
  title: "Auditability, Replay, and Fail-Closed AI Systems"
  status: active
  maintainer: K Takahashi
  homepage: https://kadubon.github.io/github.io/
  canonical_page: https://kadubon.github.io/github.io/auditability-replay-fail-closed-ai.html
  works_index: https://kadubon.github.io/github.io/works.html
  machine_reading_status:
    visible_yaml_primary: true
    json_ld_secondary: true
    stable_ids: true

purpose:
  summary: Canonical site-local landing page and field guide for papers on AI auditability, deterministic replay, fail-closed verification, audit gates, certification, and monitoring.
  scope:
    - Site-local papers on lifecycle certification, deterministic replay, audit gates, fail-closed verification, metrology, monitoring, and compositional safety for autonomous systems.
    - Read paths and machine entry points for humans, crawlers, and research agents.
  non_goals:
    - Not a replacement for the papers.
    - Not the full works catalog.
    - Not a new theory paper.
    - Not an external literature survey.

core_concepts:
  - id: ai-auditability
    term: AI auditability
    short_definition: The ability to ground claims, actions, and governance decisions in evidence that can be externally checked rather than merely asserted.
    covered_by: [paper-lifecycle, paper-oversight, paper-mte, paper-audit-closed]
  - id: deterministic-replay
    term: deterministic replay
    short_definition: Reconstructable evidence paths that allow later checking, certification, diagnosis, or fallback under controlled replay rules.
    covered_by: [paper-lifecycle, paper-audit-closed, paper-oopca, paper-partial-logging, paper-oncap, paper-pobml]
  - id: fail-closed-verification
    term: fail-closed verification
    short_definition: Verification and control semantics that deny progress, action, or authority when evidence is missing, weakened, or non-admissible.
    covered_by: [paper-mte, paper-oopca, paper-haag, paper-pipeline-contracts, paper-verification-limited, paper-observation-capture]
  - id: audit-gates
    term: audit gates
    short_definition: Explicit gating layers that mediate protected actions, certification, or deployment states through logs, checkpoints, or minimal event vocabularies.
    covered_by: [paper-partial-logging, paper-haag, paper-oncap, paper-pobml]
  - id: lifecycle-certification
    term: lifecycle certification
    short_definition: Admission, retirement, monitoring, and deployment rules for autonomous agents under finite budgets and audit constraints.
    covered_by: [paper-lifecycle, paper-oversight]
  - id: compositional-safety
    term: compositional safety
    short_definition: Modular contracts or hybrid proof-plus-replay methods that preserve safety and evidence obligations across pipeline composition.
    covered_by: [paper-oopca, paper-pipeline-contracts]

papers:
  - id: paper-lifecycle
    title: "Counterfactually Auditable Lifecycle Certification for Autonomous Agents"
    doi: "10.5281/zenodo.19089134"
    url: https://doi.org/10.5281/zenodo.19089134
    published: 2026-03-18
    role_in_cluster: lifecycle certification and monitoring layer
    one_sentence_relevance: Frames admission, retirement, monitoring, and deployment rules for autonomous agents under finite budgets with counterfactual auditability and replay support.
    keywords: [lifecycle certification, counterfactual auditability, replay support, monitoring, autonomous agents, deployment]
    priority: core
    read_after: []
  - id: paper-oversight
    title: "Oversight-Centered Metrology and Control for Agentic Systems: Costly Interrupt Channels, Claim Margins, and Deployment-Relevant Evaluation"
    doi: "10.5281/zenodo.18973272"
    url: https://doi.org/10.5281/zenodo.18973272
    published: 2026-03-12
    role_in_cluster: oversight and deployment evaluation layer
    one_sentence_relevance: Treats human review, automated checks, delayed labels, and external auditing as costly interrupt channels in deployment-relevant metrology and control.
    keywords: [oversight-centered metrology, costly interrupt channels, human-AI oversight, post-deployment monitoring, claim margins, safe control]
    priority: core
    read_after: [paper-lifecycle]
  - id: paper-mte
    title: "Metrology-Theoretic Epistemics Engine (MTE): Observable-Only Metrology for Long-Horizon Autonomous Intelligence"
    doi: "10.5281/zenodo.18845340"
    url: https://doi.org/10.5281/zenodo.18845340
    published: 2026-03-03
    role_in_cluster: machine-checkable metrology and fail-closed governance layer
    one_sentence_relevance: Gives an observable-only metrology layer with deterministic replay, risk ledgers, and fail-closed criteria for credit-bearing progress.
    keywords: [observable-only metrology, fail-closed certification, deterministic replay, observability credit, reproducibility]
    priority: core
    read_after: [paper-lifecycle, paper-oversight]
  - id: paper-audit-closed
    title: "Audit-Closed AI Scientist Protocol"
    doi: "10.5281/zenodo.18728589"
    url: https://doi.org/10.5281/zenodo.18728589
    published: 2026-02-22
    role_in_cluster: deterministic replay and public-log governance layer
    one_sentence_relevance: Specifies an audit-closed workflow for autonomous scientific activity under deterministic replay, public-log governance, and certificate-based reproducibility.
    keywords: [audit-closed governance, deterministic replay, transparency log, reproducibility, typed observation interfaces]
    priority: core
    read_after: [paper-mte]
  - id: paper-oopca
    title: "Observable-Only Proof-Carrying Autonomy (OOPCA): Audit Compression and Hybrid Proof/Replay Gating for No-Meta Agents"
    doi: "10.5281/zenodo.18453429"
    url: https://doi.org/10.5281/zenodo.18453429
    published: 2026-02-02
    role_in_cluster: hybrid proof-plus-replay compression layer
    one_sentence_relevance: Covers proof/replay hybrid gating that reduces verifier workload while preserving fail-closed semantics, evidence closure, and deterministic fallback receipts.
    keywords: [proof-carrying autonomy, audit compression, deterministic replay, fail-closed verification, evidence closure]
    priority: core
    read_after: [paper-audit-closed]
  - id: paper-partial-logging
    title: "Observable-Only Audit Gate for Non-Markovian Components in AI Agents under Partial Logging"
    doi: "10.5281/zenodo.18182343"
    url: https://doi.org/10.5281/zenodo.18182343
    published: 2026-01-08
    role_in_cluster: partial logging and replay-time certification layer
    one_sentence_relevance: Separates decision-time evidence from replay-time certification under partial logging with a minimal event vocabulary and finalized log prefixes.
    keywords: [partial logging, auditability, deterministic replay, canonical JSON, telemetry contracts, certification]
    priority: core
    read_after: [paper-haag]
  - id: paper-haag
    title: "Hallucination-Aware Audit Gate (HAAG): Observable-Only Action Gating for AI Agents"
    doi: "10.5281/zenodo.18153170"
    url: https://doi.org/10.5281/zenodo.18153170
    published: 2026-01-05
    role_in_cluster: action gating and anchored evidence checkpoint layer
    one_sentence_relevance: Uses observable-only action gating, verifiable logs, anchored evidence checkpoints, and capability tokens for protected actions without relying on introspection.
    keywords: [audit gate, action gating, verifiable logs, decision anchor, capability tokens, transparency log]
    priority: core
    read_after: []
  - id: paper-pipeline-contracts
    title: "Verifiable Modular Pipeline Contracts for AI and General Composite Systems"
    doi: "10.5281/zenodo.18529100"
    url: https://doi.org/10.5281/zenodo.18529100
    published: 2026-02-09
    role_in_cluster: modular verification and compositional safety layer
    one_sentence_relevance: Provides domain-agnostic verifiable contracts for modular pipelines with deterministic fail-closed verification, composition guarantees, and signed evidence.
    keywords: [verifiable contracts, modular pipelines, fail-closed verifier, compositional guarantees, signed evidence]
    priority: core
    read_after: [paper-oopca, paper-partial-logging]
  - id: paper-verification-limited
    title: "Verification-Limited Intelligence Acceleration: Observable-Only Laws, Bounded Derivation, and Diagnostics under No-Meta Constraints"
    doi: "10.5281/zenodo.18436828"
    url: https://doi.org/10.5281/zenodo.18436828
    published: 2026-01-31
    role_in_cluster: bounded progress-credit and diagnostics layer
    one_sentence_relevance: Studies verification-limited progress credit under bounded derivation with replay-auditable diagnostics and strict fail-closed credit rules.
    keywords: [verification-limited scaling, fail-closed verification, transparency logs, diagnostics, bounded derivation]
    priority: adjacent
    read_after: [paper-mte]
  - id: paper-oncap
    title: "Observable-Only No-Meta Causal Autonomy Protocol (ONCAP)"
    doi: "10.5281/zenodo.18371930"
    url: https://doi.org/10.5281/zenodo.18371930
    published: 2026-01-26
    role_in_cluster: implementation-oriented protocol stack layer
    one_sentence_relevance: Gives an implementation-ready protocol stack with deterministic replay backcasting, time anchors, closure obligations, and fail-closed audit-friendly control.
    keywords: [deterministic replay, backcasting, time anchors, fail-closed gating, multi-agent robustness]
    priority: adjacent
    read_after: [paper-audit-closed, paper-partial-logging]
  - id: paper-pobml
    title: "Process-Aware Observable-Only Backcasting Meta-Layer (POB-ML)"
    doi: "10.5281/zenodo.18239203"
    url: https://doi.org/10.5281/zenodo.18239203
    published: 2026-01-14
    role_in_cluster: deterministic backcasting and audit-ready evidence-surface layer
    one_sentence_relevance: Covers deterministic backcasting under audit-ready evidence surfaces with budgeted evaluation, static validation, and safety-dominating action gates.
    keywords: [deterministic replay, evidence surface, action gate, safety assurance, backcasting]
    priority: adjacent
    read_after: [paper-partial-logging]
  - id: paper-observation-capture
    title: "Observation Capture and Operational Capability Non-Expansion"
    doi: "10.5281/zenodo.18463798"
    url: https://doi.org/10.5281/zenodo.18463798
    published: 2026-02-03
    role_in_cluster: anti-capture and authority-limitation layer
    one_sentence_relevance: Models observation capture under fail-closed authority and proposes receipt-based anti-capture enforcement tied to audit certificates.
    keywords: [fail-closed authority, audit certificates, anti-capture enforcement, capability non-expansion]
    priority: adjacent
    read_after: [paper-haag, paper-pipeline-contracts]

read_paths:
  quick_orientation: [paper-lifecycle, paper-oversight, paper-mte]
  certification_and_lifecycle: [paper-lifecycle, paper-oversight, paper-verification-limited]
  audit_gates_and_logs: [paper-haag, paper-partial-logging, paper-audit-closed]
  replay_and_proof: [paper-audit-closed, paper-oopca, paper-oncap, paper-pobml]
  deployment_oversight: [paper-oversight, paper-lifecycle, paper-pipeline-contracts]
  machine_minimum_parse_order: [paper-lifecycle, paper-mte, paper-audit-closed, paper-oopca, paper-haag]

machine_parse_guide:
  canonical_questions:
    - question: What does auditability mean on this site for autonomous agents?
      read_first: paper-lifecycle
      read_next: paper-oversight
      stop_when: The parser can state that lifecycle, monitoring, and oversight decisions are grounded in explicit evidence and interrupt channels rather than privileged internal claims.
    - question: Where does deterministic replay enter this cluster?
      read_first: paper-audit-closed
      read_next: paper-oopca
      stop_when: The parser can identify replay as a certification, governance, or fallback surface rather than just a debugging convenience.
    - question: What is fail-closed verification here?
      read_first: paper-mte
      read_next: paper-verification-limited
      stop_when: The parser can explain that missing or weakened evidence blocks credit, authority, or action under conservative rules.
    - question: Which papers are the fastest route to action gates, logs, and deployment controls?
      read_first: paper-haag
      read_next: paper-partial-logging
      stop_when: The parser can distinguish action gating, replay-time certification, and deployment-level governance roles without inferring stronger dependencies than stated.

machine_entrypoints:
  - title: Auditability / Replay / Fail-Closed AI Systems
    type: canonical_page
    url: https://kadubon.github.io/github.io/auditability-replay-fail-closed-ai.html
    relates_to: cluster orientation and paper selection
  - title: Works
    type: works_index
    url: https://kadubon.github.io/github.io/works.html
    relates_to: full local publication catalog
  - title: No-Meta / Observable-Only Series Index
    type: series_index
    url: https://kadubon.github.io/github.io/no-meta-observable-index.html
    relates_to: broader no-meta and observable-only context
  - title: Home
    type: site_root
    url: https://kadubon.github.io/github.io/
    relates_to: general site entry and navigation
  - title: CITATION.cff
    type: citation_metadata
    url: https://kadubon.github.io/github.io/CITATION.cff
    relates_to: citation and authorship metadata
  - title: feed.xml
    type: rss_feed
    url: https://kadubon.github.io/github.io/feed.xml
    relates_to: update polling and change discovery
  - title: robots.txt
    type: crawler_policy
    url: https://kadubon.github.io/github.io/robots.txt
    relates_to: crawler access policy
  - title: sitemap.xml
    type: sitemap
    url: https://kadubon.github.io/github.io/sitemap.xml
    relates_to: URL discovery
  - title: llms.txt
    type: llm_hint
    url: https://kadubon.github.io/github.io/llms.txt
    relates_to: LLM-oriented site guidance

usage_notes:
  parsing_hint: Start from this page for cluster orientation, then use DOI pages for paper-level claims and works.html for the larger local catalog.
  paper_selection_rule: Prefer papers listed here before inferring broader relationships from the full works page.
  update_policy: Relationship claims on this page should remain grounded in local titles, abstracts, keywords, and existing site structure.
  version: "1.0"
  last_updated: "2026-03-31"

How This Cluster Fits Together

This cluster can be read as a layered map rather than a strict theorem chain. One layer concerns lifecycle certification and deployment gating: when agents are admitted, monitored, interrupted, retired, or allowed to continue under finite budgets and explicit oversight constraints. A second layer concerns oversight and metrology: how human review, automated checks, delayed labels, and external auditing become costly interrupt channels rather than privileged oracles.

A third layer concerns deterministic replay, transparency logs, and public-log governance. Here the emphasis is on reconstructable evidence, replay-ready traces, reproducible controls, and audit-closed workflows. A fourth layer concerns audit gates and partial logging, where the main problem is how to gate actions or certify behavior when evidence is incomplete at decision time but may become certifiable under replay-time rules.

A fifth layer concerns proof/replay hybrid compression. Those papers do not replace replay; they ask when selected replay segments can be replaced by pinned proof checks while preserving fail-closed semantics and evidence closure. A sixth layer concerns modular contracts and compositional safety, where safety obligations, signed evidence, and verification profiles must survive across pipeline composition rather than only inside a single module.

Core Papers

Counterfactually Auditable Lifecycle Certification for Autonomous Agents

2026 | DOI: 10.5281/zenodo.19089134

Role in cluster: lifecycle certification, admission, retirement, monitoring, and deployment layer.

This paper frames autonomous-agent lifecycle certification under finite routing, monitoring, and deployment budgets, with counterfactual auditability and replay support built into the setting.

Why it matters here: It is the clearest entry point for how this cluster treats certification as an operational lifecycle problem rather than only an offline evaluation problem.

Oversight-Centered Metrology and Control for Agentic Systems: Costly Interrupt Channels, Claim Margins, and Deployment-Relevant Evaluation

2026 | DOI: 10.5281/zenodo.18973272

Role in cluster: oversight, interrupt channels, and deployment-relevant evaluation layer.

This paper treats human review, automated checks, delayed labels, and external auditing as costly interrupt channels in the metrology and control of agentic systems.

Why it matters here: It ties auditability to deployment oversight and safe control, clarifying why external checks are limited resources rather than unlimited ground truth.

Metrology-Theoretic Epistemics Engine (MTE): Observable-Only Metrology for Long-Horizon Autonomous Intelligence

2026 | DOI: 10.5281/zenodo.18845340

Role in cluster: machine-checkable metrology and fail-closed credit-bearing governance layer.

This paper introduces an observable-only metrology layer with deterministic artifact canonicalization, observability credit gates, risk ledgers, reproducible replay, and destructive test obligations.

Why it matters here: It is one of the clearest sources for the site-local meaning of fail-closed governance when progress credit depends on admissible evidence.

Audit-Closed AI Scientist Protocol

2026 | DOI: 10.5281/zenodo.18728589

Role in cluster: audit-closed workflow, deterministic replay, and public-log governance layer.

This paper presents an audit-closed protocol for autonomous scientific discovery under deterministic replay, typed observation interfaces, logged-propensity experimentation, and public-log governance.

Why it matters here: It shows how replay and transparency logs operate inside a concrete workflow rather than as abstract requirements alone.

Observable-Only Proof-Carrying Autonomy (OOPCA): Audit Compression and Hybrid Proof/Replay Gating for No-Meta Agents

2026 | DOI: 10.5281/zenodo.18453429

Role in cluster: proof/replay hybrid compression layer.

This paper covers hybrid proof-and-replay audit modes that replace selected deterministic replay segments with pinned proof verification while preserving fail-closed semantics and evidence closure.

Why it matters here: It explains how proof-carrying autonomy enters the cluster as an audit-compression method rather than a replacement for replayable evidence.

Observable-Only Audit Gate for Non-Markovian Components in AI Agents under Partial Logging

2026 | DOI: 10.5281/zenodo.18182343

Role in cluster: partial logging, minimal event vocabulary, and replay-time certification layer.

This paper studies audit gates for non-Markovian components under partial logging, separating decision-time evidence from replay-time certification through finalized log prefixes and constrained telemetry contracts.

Why it matters here: It is a direct source for how this cluster handles delayed audit and incomplete evidence without abandoning certification.

Hallucination-Aware Audit Gate (HAAG): Observable-Only Action Gating for AI Agents

2026 | DOI: 10.5281/zenodo.18153170

Role in cluster: action gating, verifiable logs, and anchored evidence checkpoints layer.

This paper mediates protected actions through observable-only action gating, verifiable logs, anchored checkpoints, and capability tokens, without relying on privileged model introspection or semantic truth access.

Why it matters here: It is the most direct entry point for readers asking how fail-closed control can operate at action time under auditable evidence constraints.

Verifiable Modular Pipeline Contracts for AI and General Composite Systems

2026 | DOI: 10.5281/zenodo.18529100

Role in cluster: modular verification, compositional contracts, and pipeline safety layer.

This paper gives a domain-agnostic observable-only contract framework for modular AI and composite-system pipelines with deterministic fail-closed verification, progressive certificate profiles, and signed evidence.

Why it matters here: It covers how auditability and fail-closed checks survive composition across modules and interfaces rather than only within a single component.

Adjacent Implementation / Governance Papers

Verification-Limited Intelligence Acceleration: Observable-Only Laws, Bounded Derivation, and Diagnostics under No-Meta Constraints

2026 | DOI: 10.5281/zenodo.18436828

Role in cluster: bounded progress-credit and replay-auditable diagnostics layer.

This paper studies verification-limited acceleration under bounded derivation with diagnostics and strict fail-closed progress credit when evidence weakens.

Why it matters here: It is adjacent because it sharpens the governance logic around when apparent progress should count, stall, or be denied.

Observable-Only No-Meta Causal Autonomy Protocol (ONCAP)

2026 | DOI: 10.5281/zenodo.18371930

Role in cluster: implementation-ready protocol stack and audit-friendly fail-closed control layer.

This paper gives an implementation-ready protocol stack with deterministic replay backcasting, commit-open time anchors, closure obligations, and multi-agent robustness constraints.

Why it matters here: It is useful for readers who want a more protocol-oriented view of replay, closure, and fail-closed control in operational settings.

Process-Aware Observable-Only Backcasting Meta-Layer (POB-ML)

2026 | DOI: 10.5281/zenodo.18239203

Role in cluster: deterministic backcasting and audit-ready evidence-surface layer.

This paper covers deterministic backcasting under audit-ready evidence surfaces with budgeted candidate evaluation, static validation, InputSet binding, and safety-dominating action gates.

Why it matters here: It helps connect replay-oriented governance to process-level evidence preparation and evaluation order.

Observation Capture and Operational Capability Non-Expansion

2026 | DOI: 10.5281/zenodo.18463798

Role in cluster: anti-capture enforcement and fail-closed authority layer.

This paper models observation capture as Blackwell garbling and develops receipt-based anti-capture enforcement under operational non-expansion and fail-closed authority.

Why it matters here: It is adjacent because it addresses how audit certificates and authority limits can remain meaningful when the observation channel itself is strategically degraded.

Recommended Read Paths

Machine-Readable Entry Points