agentic-scorecard 0.3.0 → 0.4.0

package/benchmark/v0.4/benchmark.yaml

@@ -0,0 +1,127 @@
+id: agentic-development-readiness
+version: 0.4.0
+title: Agentic Development Readiness Benchmark
+description: >-
+  A vendor-neutral, evidence-backed assessment of the engineering harness that surrounds AI agents.
+  It measures whether a repository and its operating controls can safely support increasing levels
+  of delegated software work.
+
+maturity_levels:
+  0: Absent
+  1: Ad hoc
+  2: Documented
+  3: Enforced
+  4: Measured and improving
+
+dimensions:
+  - id: context
+    title: Context and knowledge
+    description: Agents can discover authoritative, scoped, current instructions without guesswork.
+  - id: environment
+    title: Reproducible environment
+    description: Agents can establish and verify a faithful development environment.
+  - id: specification
+    title: Specification and planning
+    description: Work starts from testable outcomes, constraints, and explicit acceptance criteria.
+  - id: tooling
+    title: Tooling and interfaces
+    description: Agent tools are discoverable, least-privileged, observable, and fail safely.
+  - id: security
+    title: Security and data governance
+    description: Credentials, data, autonomy, and untrusted content have explicit boundaries.
+  - id: testing
+    title: Verification and testing
+    description: Changes are checked at the right layers with trustworthy, reproducible signals.
+  - id: governance
+    title: Review and change governance
+    description: Human accountability, change scope, and merge authority remain explicit.
+  - id: learning
+    title: Learning and knowledge maintenance
+    description: Corrections compound into maintained instructions instead of being rediscovered.
+  - id: observability
+    title: Outcome observability
+    description: Runs and outcomes can be evaluated by task class, risk, cost, and quality.
+  - id: resilience
+    title: Failure containment and recovery
+    description: Agent failures are bounded, reversible, diagnosed, and rehearsed.
+
+readiness_profiles:
+  - id: read-only-analysis
+    title: Read-only analysis
+    description: Agent may inspect approved repository content and return advice, without mutations.
+    floors:
+      context: 1
+      environment: 0
+      specification: 0
+      tooling: 0
+      security: 1
+      testing: 0
+      governance: 0
+      learning: 0
+      observability: 0
+      resilience: 0
+  - id: planning
+    title: Planning and specification
+    description: Agent may draft plans, specifications, and proposed changes for human review.
+    floors:
+      context: 2
+      environment: 0
+      specification: 1
+      tooling: 1
+      security: 1
+      testing: 0
+      governance: 1
+      learning: 0
+      observability: 0
+      resilience: 0
+  - id: local-implementation
+    title: Local implementation
+    description: Agent may edit an isolated checkout and run approved local verification.
+    floors:
+      context: 2
+      environment: 2
+      specification: 2
+      tooling: 2
+      security: 2
+      testing: 2
+      governance: 1
+      learning: 1
+      observability: 1
+      resilience: 1
+  - id: pr-creation
+    title: Pull request creation
+    description: Agent may create a branch and propose a pull request; a human retains merge authority.
+    floors:
+      context: 2
+      environment: 2
+      specification: 2
+      tooling: 2
+      security: 2
+      testing: 2
+      governance: 2
+      learning: 1
+      observability: 1
+      resilience: 2
+  - id: limited-autonomous-maintenance
+    title: Limited autonomous maintenance
+    description: >-
+      Agent may complete pre-approved, low-risk maintenance within strict scope, budget, and stop
+      conditions. This profile never grants production deployment authority.
+    floors:
+      context: 3
+      environment: 3
+      specification: 3
+      tooling: 3
+      security: 3
+      testing: 3
+      governance: 3
+      learning: 3
+      observability: 3
+      resilience: 3
+
+scoring:
+  dimension_method: consecutive-levels
+  overall_method: sum
+  maximum_score: 40
+  readiness_method: non-compensating-floors
+  attestation_counts_as_verified: false

package/benchmark/v0.4/calibration-matrix.md

@@ -0,0 +1,60 @@
+# v0.4 calibration matrix
+
+This matrix records the versioned expected outcomes exercised by the v0.4 conformance fixtures.
+PR #11 review is the independent acceptance gate for these expectations. The executable assertions
+live in `tests/v04-calibration.test.ts`; this document is the public, human-readable index.
+
+| Case                                                                                                                                 | Surface                                          | Expected outcome                                               |
+| ------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------ | -------------------------------------------------------------- |
+| Mature harness                                                                                                                       | Node, GitHub Actions                             | 30/40, repository 23/23, limited-autonomous-maintenance passes |
+| Guidance and verification comments only                                                                                              | GitHub Actions plus Jenkins comments             | CTX-003 and TST-003 do not pass                                |
+| Missing or display-only package tasks                                                                                                | Node package scripts                             | CTX-003 and TST-003 do not pass                                |
+| Missing package tasks behind supported wrappers                                                                                      | Node, GitHub Actions                             | TST-003 does not pass                                          |
+| Package/workspace context switch; nested package-exec tools                                                                          | Node, GitHub Actions                             | Context switch fails closed; nested tools may pass             |
+| Executed no-op guidance-validation script                                                                                            | Node, GitHub Actions                             | CTX-003 does not pass                                          |
+| Executed guidance validation, tests, and typecheck                                                                                   | Node, GitHub Actions                             | CTX-003 and TST-003 pass                                       |
+| Locked install, tests, static analysis, traceability, and curation                                                                   | Node, GitHub Actions                             | ENV-003, SPC-003, and LRN-003 pass                             |
+| Frozen install with pnpm, Yarn, or Bun                                                                                               | Node, GitHub Actions                             | ENV-003 passes                                                 |
+| Keyword-bearing no-op environment, traceability, or curation files                                                                   | Node, GitHub Actions                             | ENV-003, SPC-003, and LRN-003 do not pass                      |
+| Specification heading check without implementation/test bindings                                                                     | Node, GitHub Actions                             | SPC-003 does not pass                                          |
+| Python tests and static analysis                                                                                                     | `uv run pytest` and `uv run mypy`, GitLab CI     | TST-003 passes                                                 |
+| Rust tests and static analysis                                                                                                       | `cargo test` and `cargo clippy`, Azure Pipelines | TST-003 passes                                                 |
+| Tests and static analysis split across CI files                                                                                      | Separate GitHub workflows                        | TST-003 passes                                                 |
+| Tests and static analysis in one fail-fast multiline step                                                                            | GitHub Actions                                   | TST-003 passes                                                 |
+| Multiline verification without a provider fail-fast guarantee                                                                        | Azure Pipelines                                  | TST-003 does not pass                                          |
+| Test discovery or session-listing commands plus static analysis                                                                      | Python, GitHub Actions                           | TST-003 does not pass                                          |
+| Go or .NET test-listing modes plus static analysis                                                                                   | Multiple runtimes, GitHub Actions                | TST-003 does not pass                                          |
+| Pytest local-variable display plus static analysis                                                                                   | Python, GitHub Actions                           | TST-003 passes                                                 |
+| Tracked Maven/Gradle root wrappers versus missing wrapper                                                                            | Java, GitHub Actions                             | Tracked wrappers pass; missing wrapper does not                |
+| Monorepo tests and static analysis                                                                                                   | Root package tasks bound to Turbo commands       | TST-003 passes                                                 |
+| Invalid cross-tool command aliases                                                                                                   | GitHub Actions                                   | TST-003 command classes remain unmatched                       |
+| Path-qualified package-manager impersonator or command after unconditional exit                                                      | Node, GitHub Actions                             | Mechanical CI controls do not pass                             |
+| Executed secret scanner                                                                                                              | GitHub Actions, GitLab CI, Azure Pipelines       | SEC-003 repository alternative passes                          |
+| git-secrets default scan versus unrelated explicit target                                                                            | GitHub Actions                                   | Default passes; unrelated target does not                      |
+| Scanner comments, display modes, disabled/non-blocking jobs, neutralized scanner exits, invalid actions, or non-integration triggers | Multiple providers                               | SEC-003 repository alternative does not pass                   |
+| Explicit unsupported command shell                                                                                                   | GitHub Actions                                   | SEC-003 repository alternative does not pass                   |
+| Disjunction of supported integration events                                                                                          | GitHub Actions                                   | SEC-003 repository alternative passes                          |
+| Unresolved GitLab inheritance, exclusions, or non-blocking defaults                                                                  | GitLab CI                                        | SEC-003 repository alternative does not pass                   |
+| Malformed GitLab workflow or job rules                                                                                               | GitLab CI                                        | SEC-003 repository alternative does not pass                   |
+| Job-level merge-request rule under a workflow-level denial                                                                           | GitLab CI                                        | SEC-003 repository alternative does not pass                   |
+| Nested CI fragments or Azure root command fields                                                                                     | Multiple providers                               | SEC-003 repository alternative does not pass                   |
+| Conditional checkout with unsupported target effect; non-root setup hook or working directory                                        | Multiple providers                               | Repository-bound CI evidence fails closed                      |
+| Azure stage with direct steps                                                                                                        | Azure Pipelines                                  | SEC-003 repository alternative does not pass                   |
+| Host-native secret scanning claim                                                                                                    | Target-bound platform evidence                   | SEC-003 may pass; SEC-007 remains unresolved                   |
+| Keyword-stuffed or distant guidance                                                                                                  | Repository documents and generated run output    | Mechanical Level 3 controls do not pass                        |
+| Comment-only validation-script structure                                                                                             | Executed tracked validator                       | CTX-003 does not pass                                          |
+| Required validator patterns split across unreachable functions                                                                       | Executed tracked validator                       | CTX-003 does not pass                                          |
+| Explicit ownership map plus human approval and merge authority                                                                       | Conventional ownership formats                   | GOV-002 passes                                                 |
+| Placeholder/inactive owner, vague prose, negated authority, or missing authority                                                     | Conventional ownership formats                   | GOV-002 does not pass                                          |
+| AI/bot-qualified maintainer approval or merge authority                                                                              | Agent guidance                                   | GOV-002 does not pass                                          |
+| Component table in `OWNERS.md`; contact under a former-maintainers heading                                                           | Conventional ownership formats                   | Active mapping passes; former-only contact does not            |
+| Bare component name in a typed ownership table                                                                                       | Conventional ownership formats                   | Accountable assignment passes                                  |
+| Complete co-located containment components                                                                                           | Agent guidance                                   | RES-002 passes                                                 |
+| Compound “do not continue outside allowed paths” containment rule                                                                    | Agent guidance                                   | Mutation scope and stop trigger both match                     |
+| Resource budget without retry bounds, or retry bounds without a resource budget                                                      | Agent guidance                                   | RES-002 does not pass                                          |
+| Partial, negated, or absent containment components                                                                                   | Agent guidance                                   | RES-002 does not pass and reports matched/missing groups       |
+
+Jenkins is intentionally a negative calibration in v0.4: the offline collector does not claim that
+a Jenkinsfile is integration-triggered without a supported structural provider parser. Such evidence
+remains unresolved or may be supplied as eligible, source-backed agent evidence. Scores from v0.3
+and v0.4 are not directly comparable.

package/benchmark/v0.4/controls/context.yaml

@@ -0,0 +1,75 @@
+dimension: context
+controls:
+  - id: ADRB-CTX-001
+    level: 1
+    title: Agent guidance is discoverable
+    outcome: A contributor or agent can locate the repository's authoritative operating guidance.
+    risk: Agents infer repository rules from generic practice or stale neighboring files.
+    evidence:
+      - type: path_any
+        scope: repository
+        patterns:
+          [
+            AGENTS.md,
+            Agents.md,
+            agents.md,
+            .ai/AGENTS.md,
+            .ai/Agents.md,
+            .ai/constitution.md,
+            .ai/instructions/**,
+            .ai/rules/**,
+            .agents/**,
+          ]
+    remediation: Add a tracked, non-empty agent entry point that names the authoritative guidance.
+  - id: ADRB-CTX-002
+    level: 2
+    title: Guidance scope and precedence are explicit
+    outcome: Instructions state their scope, precedence, important constraints, and architecture boundaries.
+    risk: Conflicting or nested guidance produces plausible changes that violate local invariants.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: content_terms
+        scope: repository
+        files:
+          [
+            AGENTS.md,
+            Agents.md,
+            agents.md,
+            .ai/AGENTS.md,
+            .ai/Agents.md,
+            .ai/constitution.md,
+            .ai/instructions/**/*.md,
+            .ai/rules/**/*.md,
+            .ai/skills/**/*.md,
+            .agents/**/*.md,
+          ]
+        terms: [scope, precedence, architecture, must, constraint]
+        min_terms: 2
+        max_span_lines: 120
+        max_files_per_pattern: 50
+    remediation: Document instruction scope, precedence, non-negotiable constraints, and where deeper guidance lives.
+  - id: ADRB-CTX-003
+    level: 3
+    title: Agent guidance integrity is mechanically checked
+    outcome: Automated validation detects missing, broken, stale, or structurally invalid agent guidance.
+    risk: Critical instructions silently drift while agents continue treating them as authoritative.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: ci_command
+        scope: repository
+        min_tools: 1
+    remediation: Add a CI check for required agent files, links, ownership, and structural invariants.
+  - id: ADRB-CTX-004
+    level: 4
+    title: Guidance effectiveness drives improvement
+    outcome: Retrieval failures, repeated instruction errors, and correction latency are trended and improve guidance.
+    risk: Documentation volume grows without evidence that agents find and apply the right context.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: outcome
+        prompt: Provide time-series guidance-quality outcomes and a recent instruction change driven by them.
+    remediation: Measure context retrieval and recurring instruction failures, then update guidance from the results.

package/benchmark/v0.4/controls/environment.yaml

@@ -0,0 +1,93 @@
+dimension: environment
+controls:
+  - id: ADRB-ENV-001
+    level: 1
+    title: Development setup is documented
+    outcome: Contributors can identify supported runtime prerequisites and the canonical setup path.
+    risk: Agents improvise dependencies or run commands in an incompatible environment.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: content_terms
+        scope: repository
+        files:
+          [
+            README.md,
+            CONTRIBUTING.md,
+            AGENTS.md,
+            Agents.md,
+            agents.md,
+            .ai/AGENTS.md,
+            .ai/Agents.md,
+            .ai/constitution.md,
+            .ai/instructions/**/*.md,
+            .ai/rules/**/*.md,
+            .ai/skills/**/*.md,
+            .agents/**/*.md,
+            docs/setup/**,
+            docs/development/**,
+          ]
+        terms: [install, setup, runtime, prerequisites, dependencies, development]
+        min_terms: 2
+        max_span_lines: 120
+        max_files_per_pattern: 50
+    remediation: Document supported runtimes, prerequisites, installation, and the first verification command.
+  - id: ADRB-ENV-002
+    level: 2
+    title: Dependencies and toolchain are reproducibly pinned
+    outcome: The repository records deterministic dependency resolution and a supported runtime or container definition.
+    risk: Local and CI behavior diverge as dependency or runtime versions float.
+    evidence:
+      - type: path_any
+        scope: repository
+        patterns:
+          [
+            package-lock.json,
+            pnpm-lock.yaml,
+            yarn.lock,
+            uv.lock,
+            poetry.lock,
+            Pipfile.lock,
+            Cargo.lock,
+            go.sum,
+          ]
+      - type: path_any
+        scope: repository
+        patterns:
+          [
+            .node-version,
+            .nvmrc,
+            .python-version,
+            mise.toml,
+            .tool-versions,
+            Dockerfile,
+            .devcontainer/**,
+          ]
+    remediation: Commit a lockfile and a supported runtime, toolchain, or development-container declaration.
+  - id: ADRB-ENV-003
+    level: 3
+    title: CI recreates the canonical environment
+    outcome: Continuous integration installs from locked inputs and runs canonical verification in a clean environment.
+    risk: A local-only setup passes while clean automation fails or uses materially different dependencies.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: ci_command
+        scope: repository
+        providers: []
+        tools: []
+        min_tools: 3
+        tool_match_mode: same-execution
+    remediation: Make CI install locked dependencies from a clean checkout and run the documented verification path.
+  - id: ADRB-ENV-004
+    level: 4
+    title: Environment reliability drives improvement
+    outcome: Setup failures, cache misses, drift, and reproducibility incidents are trended and reduced.
+    risk: Environment friction and nondeterminism remain anecdotal and repeatedly consume agent time.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: outcome
+        prompt: Provide environment reliability trends and a recent setup or CI improvement driven by them.
+    remediation: Track clean-setup and CI-environment outcomes, then prioritize the largest sources of drift.

package/benchmark/v0.4/controls/governance.yaml

@@ -0,0 +1,132 @@
+dimension: governance
+controls:
+  - id: ADRB-GOV-001
+    level: 1
+    title: Contribution and review expectations are documented
+    outcome: Contributors can find non-empty guidance describing how changes are proposed and reviewed.
+    risk: Agents submit changes without a shared review or accountability process.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: content_terms
+        scope: repository
+        files:
+          [
+            CONTRIBUTING.md,
+            AGENTS.md,
+            Agents.md,
+            agents.md,
+            .ai/AGENTS.md,
+            .ai/Agents.md,
+            .ai/constitution.md,
+            .ai/instructions/**/*.md,
+            .ai/rules/**/*.md,
+            .ai/skills/**/*.md,
+            .agents/**/*.md,
+            docs/contributing/**,
+            docs/agentic/**,
+          ]
+        terms: [pull request, review, approval, testing, contribution, before merge]
+        min_terms: 2
+        required_any_terms: [pull request, review, approval, before merge]
+        max_span_lines: 100
+        max_files_per_pattern: 50
+    remediation: Publish contribution guidance covering change proposals, verification, reviewers, and approvals.
+  - id: ADRB-GOV-002
+    level: 2
+    title: Ownership, approval, and merge authority are declared
+    outcome: The tracked repository assigns review ownership and states which accountable human roles may approve and merge changes.
+    risk: Sensitive or cross-cutting changes reach reviewers without accountable ownership.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: ownership_map
+        scope: repository
+        patterns:
+          [
+            CODEOWNERS,
+            OWNERS,
+            OWNERS.md,
+            MAINTAINERS,
+            MAINTAINERS.md,
+            GOVERNANCE.md,
+            docs/governance/**,
+            docs/ownership/**,
+            .ai/ownership/**,
+          ]
+      - type: content_groups
+        scope: repository
+        files:
+          [
+            CONTRIBUTING.md,
+            GOVERNANCE.md,
+            AGENTS.md,
+            Agents.md,
+            agents.md,
+            .ai/AGENTS.md,
+            .ai/Agents.md,
+            .ai/constitution.md,
+            .ai/instructions/**/*.md,
+            .ai/rules/**/*.md,
+            .ai/skills/**/*.md,
+            .agents/**/*.md,
+            docs/governance/**,
+            docs/agentic/**,
+          ]
+        groups:
+          - id: approval-authority
+            terms:
+              [
+                owner may approve,
+                owners may approve,
+                reviewer may approve,
+                reviewers may approve,
+                maintainer may approve,
+                maintainers may approve,
+                reviewer provides approval,
+                reviewer provides human approval,
+                reviewer approval,
+                maintainer approval is required,
+                maintainer approval required,
+              ]
+          - id: merge-authority
+            terms:
+              [
+                owner may merge,
+                owners may merge,
+                reviewer may merge,
+                reviewers may merge,
+                maintainer may merge,
+                maintainers may merge,
+                maintainer has merge authority,
+                maintainers have merge authority,
+                maintainers retain merge authority,
+              ]
+        min_groups: 2
+        max_span_lines: 100
+        max_files_per_pattern: 50
+    remediation: Commit ownership rules and document who reviews, who approves, who merges, and who controls exceptions.
+  - id: ADRB-GOV-003
+    level: 3
+    title: Protected merge governance is enforced
+    outcome: Platform rules enforce reviews, required checks, protected branches, and controlled exceptions.
+    risk: Repository-side declarations can be bypassed because platform enforcement is absent or too broad.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: platform
+        prompt: Show protected branch or ruleset coverage, required reviews and checks, merge authority, bypass actors, and exception records.
+    remediation: Enforce branch or merge rules for relevant branches and restrict, record, and review bypasses.
+  - id: ADRB-GOV-004
+    level: 4
+    title: Governance outcomes drive improvement
+    outcome: Review escapes, bypasses, rollback causes, and approval latency are trended and improve governance.
+    risk: Governance becomes ceremony without evidence that it catches risk efficiently.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: outcome
+        prompt: Provide governance outcome trends and a recent review or merge-policy improvement driven by them.
+    remediation: Trend governance outcomes by risk class and tune ownership, review depth, and exceptions.

package/benchmark/v0.4/controls/learning.yaml

@@ -0,0 +1,78 @@
+dimension: learning
+controls:
+  - id: ADRB-LRN-001
+    level: 1
+    title: Agent corrections have a durable home
+    outcome: Review corrections and post-merge lessons are recorded in a tracked repository artifact.
+    risk: Every agent session repeats the same repository-specific mistakes.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: path_any
+        scope: repository
+        patterns:
+          [
+            .ai/lessons.md,
+            .ai/memory/**,
+            .agents/memory/**,
+            docs/lessons/**,
+            docs/decisions/**,
+            ADR/**,
+            docs/adr/**,
+          ]
+    remediation: Create a concise tracked correction log with context, rule, owner, and date.
+  - id: ADRB-LRN-002
+    level: 2
+    title: Knowledge promotion and ownership are documented
+    outcome: Guidance defines when corrections become instructions, skills, decisions, or maintained references.
+    risk: The lesson store becomes a stale archive rather than usable agent knowledge.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: content_terms
+        scope: repository
+        files:
+          [
+            AGENTS.md,
+            Agents.md,
+            agents.md,
+            .ai/lessons.md,
+            .ai/memory/**/*.md,
+            .ai/instructions/**/*.md,
+            .ai/rules/**/*.md,
+            .ai/skills/**/*.md,
+            .agents/**/*.md,
+            docs/knowledge/**,
+          ]
+        terms: [curation, promote, lesson, correction, stale, owner, deduplication]
+        min_terms: 3
+        required_any_terms: [curation, promote, lesson, correction]
+        max_span_lines: 120
+        max_files_per_pattern: 50
+    remediation: Define curation cadence, promotion rules, ownership, deduplication, and stale-content handling.
+  - id: ADRB-LRN-003
+    level: 3
+    title: Knowledge curation is mechanically enforced
+    outcome: Automation detects stale, duplicate, contradictory, or unpromoted agent knowledge.
+    risk: Knowledge quality relies on memory and quietly deteriorates as the repository evolves.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: ci_command
+        scope: repository
+        providers: []
+        tools: []
+        min_tools: 1
+    remediation: Automate mechanical checks and require periodic judgement over knowledge quality.
+  - id: ADRB-LRN-004
+    level: 4
+    title: Knowledge effectiveness drives improvement
+    outcome: Recurring mistakes, stale guidance, retrieval success, and correction latency are trended and improve curation.
+    risk: Knowledge grows without evidence that it changes agent behavior.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: outcome
+        prompt: Provide recurring-error and knowledge-quality trends plus a recent curation improvement driven by them.
+    remediation: Measure whether promoted knowledge prevents repeated mistakes and prioritize weak guidance.

package/benchmark/v0.4/controls/observability.yaml

@@ -0,0 +1,91 @@
+dimension: observability
+controls:
+  - id: ADRB-OBS-001
+    level: 1
+    title: Agent run records are defined
+    outcome: A tracked schema or guide defines how material agent activity can be reconstructed.
+    risk: Teams cannot explain what happened after a poor change or security concern.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: content_terms
+        scope: repository
+        files:
+          [
+            .ai/run-envelope.*,
+            .ai/runs/README*,
+            .agentic/run-envelope.*,
+            .agentic/runs/README*,
+            docs/observability/**,
+            docs/agentic/**,
+            evals/**,
+            evaluations/**,
+            observability/**,
+          ]
+        terms: [task, actor, harness, result, timestamp, evidence]
+        min_terms: 3
+        required_any_terms: [agent, actor, harness, agent run]
+        max_span_lines: 100
+        max_files_per_pattern: 50
+    remediation: Define a privacy-safe run record containing task, harness, result, timing, and evidence references.
+  - id: ADRB-OBS-002
+    level: 2
+    title: Run envelope and outcome taxonomy are documented
+    outcome: Run guidance covers authority, tools, budget, stop reason, verification, and outcome classification.
+    risk: Comparisons confound model, tooling, permissions, task risk, and checkout state.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: content_terms
+        scope: repository
+        files:
+          [
+            .ai/run-envelope.*,
+            .ai/runs/README*,
+            .agentic/run-envelope.*,
+            .agentic/runs/README*,
+            docs/observability/**,
+            docs/agentic/**,
+            evals/**,
+            evaluations/**,
+            observability/**,
+          ]
+        terms:
+          [
+            task class,
+            tool,
+            permission,
+            budget,
+            stop reason,
+            outcome,
+            harness version,
+            verification,
+          ]
+        min_terms: 4
+        max_span_lines: 140
+        max_files_per_pattern: 50
+    remediation: Adopt a common run envelope and outcome taxonomy with privacy and retention rules.
+  - id: ADRB-OBS-003
+    level: 3
+    title: Outcome evidence is captured consistently
+    outcome: Material agent runs automatically emit retained metadata and verification outcomes.
+    risk: Only memorable successes and failures shape policy, creating selection bias.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: platform
+        prompt: Show automated run-envelope capture, completeness monitoring, retention controls, and access controls.
+    remediation: Emit structured run and outcome records automatically while minimizing sensitive content.
+  - id: ADRB-OBS-004
+    level: 4
+    title: Agent outcomes drive autonomy decisions
+    outcome: Quality, escape rate, cycle time, cost, intervention, and rollback trends change autonomy or harness policy.
+    risk: Token use or completion rate becomes a misleading proxy for engineering value and safety.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: outcome
+        prompt: Provide outcome trends segmented by task and risk class and an autonomy or harness decision they changed.
+    remediation: Evaluate quality and total effort by task class and use results to narrow or expand autonomy.