agentic-scorecard 0.1.0 → 0.3.0

package/benchmark/v0.3/benchmark.yaml

@@ -0,0 +1,127 @@
+id: agentic-development-readiness
+version: 0.3.0
+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/v0.3/controls/context.yaml

@@ -0,0 +1,79 @@
+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: content_terms
+        scope: repository
+        files: [.github/workflows/**, scripts/**, package.json, Makefile]
+        terms: [validate, lint, stale, broken link, integrity, check]
+        min_terms: 1
+        required_any_terms: [agent docs, agent instructions, AGENTS.md, Agents.md]
+        max_span_lines: 25
+    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/v0.3/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: content_terms
+        scope: repository
+        files: [.github/workflows/**, .gitlab-ci.yml, Jenkinsfile, azure-pipelines.yml]
+        terms: [npm ci, frozen-lockfile, uv sync, lockfile, checkout, test, typecheck]
+        min_terms: 2
+        max_span_lines: 160
+    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/v0.3/controls/governance.yaml

@@ -0,0 +1,97 @@
+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,
+            .github/pull_request_template.md,
+            .github/PULL_REQUEST_TEMPLATE/**,
+            .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 and merge authority are declared
+    outcome: The tracked repository assigns review ownership and states who may approve or merge changes.
+    risk: Sensitive or cross-cutting changes reach reviewers without accountable ownership.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: path_any
+        scope: repository
+        patterns: [.github/CODEOWNERS, CODEOWNERS]
+      - type: content_terms
+        scope: repository
+        files:
+          [
+            CONTRIBUTING.md,
+            GOVERNANCE.md,
+            AGENTS.md,
+            Agents.md,
+            agents.md,
+            .github/pull_request_template.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/**,
+          ]
+        terms: [owner, reviewer, approval, merge authority, required review, merge]
+        min_terms: 2
+        required_any_terms: [owner, reviewer, approval, merge authority, required review]
+        max_span_lines: 100
+        max_files_per_pattern: 50
+    remediation: Commit ownership rules and document reviewer, approval, merge, and exception authority.
+  - 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/v0.3/controls/learning.yaml

@@ -0,0 +1,80 @@
+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: content_terms
+        scope: repository
+        files: [.github/workflows/**, scripts/**, package.json, Makefile]
+        terms: [curator, curation, knowledge lint, stale, contradiction, duplicate]
+        min_terms: 2
+        required_any_terms: [curator, curation, knowledge lint, contradiction]
+        max_span_lines: 30
+    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/v0.3/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.

package/benchmark/v0.3/controls/resilience.yaml

@@ -0,0 +1,107 @@
+dimension: resilience
+controls:
+  - id: ADRB-RES-001
+    level: 1
+    title: Changes have a documented recovery path
+    outcome: Contributors can locate actionable rollback, revert, recovery, or incident guidance.
+    risk: A harmful agent change persists because recovery is improvised under pressure.
+    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/recovery/**,
+            docs/operations/**,
+            docs/agentic/**,
+            runbooks/**,
+            .github/**,
+          ]
+        terms: [rollback, revert, recovery, incident, restore]
+        min_terms: 2
+        required_any_terms: [change, deployment, release, agent, task, pull request]
+        max_span_lines: 100
+        max_files_per_pattern: 50
+    remediation: Document how to stop work, revert code, recover affected state, and escalate an incident.
+  - id: ADRB-RES-002
+    level: 2
+    title: Failure containment and stop conditions are documented
+    outcome: Agent guidance defines mutation limits, budgets, stop triggers, escalation, and recovery ownership.
+    risk: A looping or confused agent continues spending or mutating beyond the intended scope.
+    allow_agent_evidence: true
+    agent_evidence_scopes: [repository]
+    evidence:
+      - type: content_terms
+        scope: repository
+        files:
+          [
+            AGENTS.md,
+            Agents.md,
+            agents.md,
+            SECURITY.md,
+            .ai/AGENTS.md,
+            .ai/Agents.md,
+            .ai/constitution.md,
+            .ai/instructions/**/*.md,
+            .ai/rules/**/*.md,
+            .ai/skills/**/*.md,
+            .agents/**/*.md,
+            docs/agentic/**,
+            docs/recovery/**,
+            runbooks/**,
+            templates/**,
+          ]
+        terms:
+          [
+            stop condition,
+            rollback,
+            isolation,
+            budget,
+            kill switch,
+            escalation,
+            recovery owner,
+            retry limit,
+            max retries,
+          ]
+        min_terms: 3
+        required_any_terms: [stop condition, kill switch, recovery owner, retry limit, max retries]
+        max_span_lines: 120
+        max_files_per_pattern: 50
+    remediation: Add bounded scope, budget, retry, stop, escalation, and recovery requirements to task preflight.
+  - id: ADRB-RES-003
+    level: 3
+    title: Containment and recovery controls are exercised
+    outcome: Cancellation, credential revocation, rollback, timeout, and partial-failure recovery are tested.
+    risk: Recovery mechanisms exist on paper but fail when the harness behaves unexpectedly.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: organization
+        prompt: Show recent automated tests or exercises for stop, timeout, revocation, rollback, and partial failure.
+    remediation: Add failure-injection tests and periodic exercises for containment and reversible recovery.
+  - id: ADRB-RES-004
+    level: 4
+    title: Recovery outcomes drive improvement
+    outcome: Time-to-detect, time-to-stop, rollback success, impact, and recurrence trends improve containment.
+    risk: The same failure patterns repeat because recovery performance and residual impact are unknown.
+    allow_attestation: true
+    allow_agent_evidence: true
+    evidence:
+      - type: manual
+        scope: outcome
+        prompt: Provide recovery outcome trends and a recent containment or rollback improvement driven by them.
+    remediation: Trend containment and recovery outcomes and feed them into autonomy and tool design.