AUDIT-recursive-audit

arifos-recursive-audit v2.0 (O_Ω Constitutional Layer)

v2.0 Upgrade (2026-08-08)

Cross-surface methodology added. Now audits ALL 10 agent surfaces (AAA canonical, kimi, opencode, grok, claude, codex, hermes, hermes-asi, openclaw-ws, openclaw-bundled), not just one directory. Contrast template: /root/AAA/skills/CONTRAST_ANALYSIS_2026-08-08.md

Cross-Surface Inventory (NEW)

# 1. Dump each surface (example for AAA vs kimi)
comm -23 <(ls /root/AAA/skills/ | sort) <(ls /root/.kimi-code/skills/ | sort)   # orphans: AAA has, kimi lacks
comm -13 <(ls /root/AAA/skills/ | sort) <(ls /root/.kimi-code/skills/ | sort)   # drift: kimi has, AAA lacks

10 Surfaces

Classification Matrix (NEW)

arifOS-ACT Embedding

Before using this skill on any mutating, irreversible, or high-blast-radius task:

1. ART — Attune (what is the real task?), Recognize (what class of power?), Test (fit · authority · evidence · blast · reversible).
2. Kernel — Route to arifOS for F1–F13 judgment if action class is Maker/Messenger/Mutator/Destroyer/Sovereign.
3. ACT — Apply narrow, Constrain scope, Trace witness, STOP before corruption.
4. Receipt — Leave evidence of what changed, why, and under whose authority.

Purpose

Audit all installed skills for overlap, stale docs, prompt bloat, trigger ambiguity, and broken references.

Use When

1. Reviewing the overall capabilities, naming conventions, and performance of the active skill portfolio.
2. A new skill is drafted or proposed for installation into any workspace or global directory.
3. An existing skill's SKILL.md is updated or modified.
4. Executing portfolio maintenance checks to clean up outdated APIs, dead links, or legacy documentation.
5. Tuning the triggering accuracy of skills when experiencing trigger drift.

Do Not Use When

1. Creating a single new skill from scratch (use skill-creator for capture and bootstrapping instead).
2. Linting individual skill trigger statements (use skill-trigger-linter instead).
3. The task is a general system performance audit unrelated to L2 Skills.

Inputs

• Skill Directories: Path to /root/.agents/skills/ or individual skill directories.
• Dependency Files: package.json, pyproject.toml, wrangler.jsonc configurations.
• Audit Manifest: List of currently configured tools and their execution limits.

Operational Age Thresholds

• Freshness Threshold: Any skill whose SKILL.md has not been modified for >30 days MUST be cross-checked against its dependent packages/APIs to verify compatibility.
• De-activation Candidate: Any skill that has not been active in the last 90 days must be classified as unused-rot and flagged for graceful de-activation or archive.

Rot Classification Schema

Each checked skill must receive a specific Rot Rating in the audit report:

• doc-rot: The skill references external URLs, guidelines, or paths that are no longer accessible or have been superseded.
• api-rot: The SDK or CLI packages that the skill coordinates (e.g. Wrangler, Agents SDK) have moved past the compatibility versions listed in the skill.
• trigger-rot: The skill's triggering criteria overlap semantically with other skills, leading to multiple activations or trigger failures.
• unused-rot: The skill is structurally valid but has not registered a telemetry execution record within the threshold window.
• archive-void-rot: (Added 2026-08-04) The skill has been physically moved to .archive-20260804/ but remains discoverable by the skill loader. Physical archive without deactivation (SKILL.md → SKILL.md.VOID) creates drift between filesystem and registry. This is a new rot class discovered during the first autonomous VOID audit.
• drift-rot: (Added 2026-08-08) Skill exists on agent surface but NOT in AAA canonical, or vice versa. Classify per matrix: PROMOTE / ARCHIVE / HARNESS-NATIVE / MIRROR.
• dual-name-rot: (Added 2026-08-08) Same skill has different names on different surfaces (e.g. FORGE-act-federation-ingress vs FORGE-sct-federation-ingress). Fix via alias table or symlink.

Procedure

1. Portfolio Scan: Map all active SKILL.md files and resolve their frontmatter definitions. Include .archive-* directories in the scan — archived skills still discoverable by the loader are archive-void-rot.
2. Staleness Analysis: Apply Age Threshold rules. Check all embedded documentation URLs and reference scripts against the latest system files and packages.
3. Prompt Bloat Check: Calculate the token density of each skill. Flag files where instruction length exceeds 500 lines.
4. Collision Auditing: Run semantic cross-checks on skill descriptions to flag potential trigger overlaps.
5. Rot Classification tagging: Assign rot ratings (doc-rot, api-rot, trigger-rot, unused-rot, archive-void-rot) to flagged skills.
6. VOID Candidate Surfacing: Rank candidates by: zero invocations + zero outcome_delta + archive-void-rot. Surface top 3. The VOID pattern: rename SKILL.md → SKILL.md.VOID (F1-reversible), ingest as VERIFY into arifFlow, seal to VAULT999. This is the FQ predator feed — VOIDing dead skills is a VERIFY act that breaks execute-heavy skew.
7. Report Generation: Write a structured markdown analysis identifying critical flaws, warnings, and remediation paths.

Postconditions

1. All skills are checked for age thresholds and direct system dependency freshness.
2. A complete rot matrix mapping overlap risk and prompt size is successfully outputted.
3. Any identified broken file reference is cataloged in the telemetry report.

Failure Modes & Escalation

• Infinite Loop Detect: Circular references between skills (e.g. Skill A loads Skill B, which loads Skill A). Action: Immediately break recursion, output a warning block, and log the path loop.
• Workspace Access Error: Missing permissions on global or shared folders. Action: Degrade gracefully and audit only local workspace skills, raising a warning flag in the report.

Telemetry per Run

{
  "skill_name": "arifos-recursive-audit",
  "version": "1.1.0",
  "trigger_phrase": "{{trigger_phrase}}",
  "selected_reason": "{{selected_reason}}",
  "latency_ms": 0,
  "token_in": 0,
  "token_out": 0,
  "commands_run": 0,
  "artifacts_written": 0,
  "postcondition_pass": false,
  "human_approval_required": false,
  "hold_code": "{{hold_code}}"
}

Recursive Scorecard

• Activation Precision: [0.0 - 1.0] (Target: >0.90)
• Task Completion Rate: [0.0 - 1.0] (Target: >0.95)
• Rollback Safety: [0.0 - 1.0] (Target: 1.00)
• Context Efficiency: [0.0 - 1.0] (Target: >0.95)
• Doc Freshness: [0.0 - 1.0] (Target: 1.00)
• Cross-Skill Collision Rate: [0.0 - 1.0] (Target: 0.00)
• Human Trust Score: [0.0 - 1.0] (Target: >0.95)

Independent Verification Lane (WAJIB 2 — added 2026-07-19)

A-FORGE planning, execution, AND verification in the same trust chain is the primary substrate defect identified in the readiness audit (ARIFOS-READINESS-2026-07-20). This skill documents the required separation so the next agent can implement it.

The defect (current state)

Hermes/A-FORGE plans
    ↓
A-FORGE executes
    ↓
A-FORGE verifies  ← same chain → cannot self-attest
    ↓
A-FORGE summarizes success

"Completed and verified" is currently too easy to self-attest. A deployment that returns exit code 0 but leaves the public service broken is reported as success.

The required separation

A-FORGE executes mutation
    ↓
Independent observe-lane verifier reads resulting reality
    ↓
Kernel checks evidence against original success criteria
    ↓
Only then may completion be recorded to VAULT999

Verifier contract (minimum)

verification_request:
  original_intent_hash: <sha256>
  success_criteria: <structured>
  mutation_receipt: <VAULT999 entry id>
  executor_identity: <actor_id>
  target_state: <expected world state after mutation>
  permitted_observation_tools: [<list>]   # read-only
  freshness_requirement: <max staleness in ms>

verification_result:
  state: VERIFIED | MISMATCH | INCONCLUSIVE | STALE
  raw_evidence_refs: [<list>]
  method: <how evidence was obtained>
  verifier_identity: <actor_id>
  verifier_independence_proof: <signed assertion>
  observed_at: <unix_ms>
  residual_uncertainty: <OBS/DER/INT/SPEC label + confidence>

Verifier hard rules

The verifier MUST:

• NOT have performed the mutation
• Use independently obtained state (its own MCP reads, not the executor's claim)
• Receive the original success criteria, NOT the executor's rewritten summary
• Be unable to modify the state it is checking
• Record raw evidence and method (not just a verdict)
• Distinguish command success (exit code 0) from outcome success (real world matches criteria)
• Return one of: VERIFIED, MISMATCH, INCONCLUSIVE, STALE
• NEVER issue constitutional approval — only the kernel may SEAL

Kernel rejection rules

The kernel MUST reject "completion" when:

• Verifier identity == executor identity
• Evidence originated ONLY from the executor (no independent observation)
• Verifier had mutation permission over the target
• Evidence is older than freshness_requirement
• Original success criteria are missing from the request
• Results cannot be independently reproduced (no second source)

Audit-discoverable evidence

This skill, when loaded, MUST trigger the audit checklist:

• Does the proposed verification lane have a distinct identity (not just a flag)?
• Does the kernel reject self-verification at the contract layer?
• Are original success criteria preserved through the receipt chain?
• Can the verifier be invoked without executor cooperation?

Authority scope

WAJIB 2 is T3 (F13 ratification required). This skill section documents the protocol only. Implementation requires the verification lane to be a new constitutional role, not a mode of A-FORGE.