Public assessment

knowledgepa3/gia-mcp-server (gia-mcp-server)

gia-mcp-server · v0.4.4 · scanned Aug 3, 2026

What changed in the harness

Selection accuracy 100→100, token cost up 5%, unconfirmed writes 0%→0%.

Category breakdown

Where the score comes from.

Earned points across the four signals Gradable measures. Safety and Legibility are scored out of 30; Economics and Discoverability are scored out of 20.

01Safety

0.0 / 30

0.0 out of 30

02Legibility

19.5 / 30

19.5 out of 30

03Economics

9.2 / 20

9.2 out of 20

04Discoverability

13.7 / 20

13.7 out of 20

Highest-impact fix

Estimated gain +30 points

Add explicit identity and permission preflight tools

Expose machine-readable principal/tenant confirmation and a non-mutating permission check so agents can verify both before destructive actions.

Description evidence

Defects and rewrites.

30 defects found across the exposed tool descriptions. Suggested rewrites make purpose, inputs, boundaries, and returns easier for an agent to understand.

Tool	Defect types	Suggested rewrite
`verify_ledger`	no_return_description	Check the internal self-consistency of the hash-chained forensic audit ledger by walking the in-memory chain reconstruction (not persisted DB rows; use verify_ledger_v2 for persisted rows). Returns a report with overall integrity status (consistent/inconsistent), whether the hash linkage is intact, and details of any broken or reordered links. This is internal self-consistency only, not third-party content verification. Classification: INFORMATIONAL — read-only, no side effects.
`verify_ledger_v2`	no_return_description	Verify the PERSISTED forensic ledger rows in PostgreSQL (not an in-memory reconstruction). Epoch-aware: epoch-2 rows are content-verified against persisted columns with Ledger Canonical v2; legacy epoch-1 rows are linkage-verified only and labeled honestly. Returns a verification report with per-epoch and overall integrity status (PASS/FAIL), counts of content-verified vs linkage-only rows, and any integrity findings. Findings append a MANDATORY LEDGER_INTEGRITY_VIOLATION entry (never repaired, never UPDATEd). Classification: INFORMATIONAL read; MANDATORY escalation on findings.
`generate_report`	name_restates_behavior no_return_description	Produce a governance status report for the current engine state. Returns a report object (format: summary/detailed/executive, default summary) containing system health, Storey Threshold status, compliance framework coverage, and operational metrics such as decision counts and gate activity. Use for periodic governance reviews or executive briefings.
`record_value_metric`	no_return_description	Record a workflow value metric (time saved, risk blocked, success rate, autonomy level, task complexity) for ROI reporting. Persists the metric with full audit trail and returns a confirmation record containing the metric ID, the recorded values, the workflow attribution, and the ledger reference.
`record_governance_event`	no_return_description	Record a governance event (gate triggered, drift prevented, violation blocked, red team resolved, human intervention) for impact reporting. Appends the event to the forensic audit ledger and returns a confirmation record with the event ID, event type, details, and ledger reference.
`generate_value_report`	no_return_description	Generate a DRAFT ledger-anchored economic value report over real runtime sessions. Returns a report object with a three-scenario value range, per-value MEASURED/MODELED provenance flags, the assumption-set hash recorded on the forensic chain, and a draft status indicating it is not yet released. Client-facing reports FAIL CLOSED unless every rate is externally cited (BLS OEWS basis) or client-declared with citations. Release is human-ISSO-only — agents propose, humans release.
`seal_memory_pack`	no_return_description	Create and hash-seal a new Governed Memory Pack (GMP), producing an immutable, TTL-bound institutional knowledge artifact with trust level enforcement. Returns the sealed pack record including its pack ID, SHA-256 integrity hash, trust level, effective TTL, and sealing timestamp and creator.
`load_memory_pack`	params_unexplained no_return_description	Load a Governed Memory Pack into an agent's execution context. Before loading, validates that the pack TTL has not expired, the trust level is permitted, the operator's role has access (operator_role is the requester's role, checked against the pack's allowed_roles), the context class is compatible, and the pack hash is intact. Returns the loaded pack content with its trust level, integrity hash, and an audit confirmation of the load.
`transfer_memory_pack`	no_return_description	Transfer a memory pack from one agent to another through a governed knowledge corridor, creating a derived pack with transfer provenance. ALWAYS requires a MANDATORY gate — no silent transfers. Returns the derived pack ID, its provenance trail (source pack to target agent), and the gate approval record.
`compose_memory_packs`	params_unexplained no_return_description	Compose two or more memory packs into a single unified execution context. Merge rules: highest risk level wins, shortest TTL wins, allowed roles intersect, and trust level contaminates downward (composed pack inherits the lowest source trust). operator_role is the composing agent's role and must satisfy the merged role set. Returns the composed pack ID, the effective merged trust/risk/TTL/roles, and the audit record of the composition.
`promote_memory_pack`	params_unexplained no_return_description	Promote a memory pack to a higher trust level after human review (how distilled EPHEMERAL packs become CASE or ORG packs). Requires a MANDATORY gate approval before the promotion is applied. approver_role is the reviewing human's authority level and is required alongside approved_by for the gate. Returns the pack's new trust level, promotion timestamp, and the gate approval audit record.
`phoenix_snapshot`	no_return_description	Create a governed state snapshot capturing the current platform operational state: ledger chain head, active gates, contracts, budgets, MAI state, intelligence counts, and memory packs. Each snapshot is SHA-256 hashed over its captured state for integrity (snapshot-to-snapshot chaining is not yet persisted). Returns the snapshot ID, capture timestamp, SHA-256 integrity hash, and a summary of the captured state. Classification: INFORMATIONAL — read-only capture, no mutations.
`phoenix_recovery_health`	no_return_description	Assess Phoenix recovery readiness. Returns a health report with an overall readiness determination (READY / DEGRADED / NOT_READY) and per-area status for snapshot availability, ledger chain integrity, intelligence layer status, and compliance posture, plus any gaps that would block recovery. Classification: INFORMATIONAL — read-only assessment.
`request_context`	no_return_description	Request governed internal context (policies, SOPs, architecture docs, compliance rules, or playbooks) — agents declare intent and GIA decides what to serve. Every retrieval is role-bound, tenant-bound, hash-verified, and ledgered. Returns the matched context items (capped by max_results per source) with each item's trust level, classification, and integrity hash; denied requests return a refusal with reason. Set include_compliance to overlay compliance mappings.
`board_convene_session`	params_unexplained	Convene a governed deliberation session for a charter. Each seat (an AI model with a specific role) deliberates on the topic according to the charter's mode (parallel/chain/adversarial/roundtable/auto). institution_id is the governing body's ID from board_list_institutions and identifies which institution owns the charter (match it with charter_id from board_list_charters). Returns a session_id — use board_get_session to retrieve the output once complete. Typical runtime: 30–120 seconds depending on seat count and mode.
`board_get_session`	params_unexplained	Retrieve the status and output of a governed deliberation session. Locate the session by passing institution_id (the governing body's ID from board_list_institutions), charter_id (the board or committee ID from board_list_charters), and session_id (returned by board_convene_session). Returns each seat's position, the synthesized consensus output, dissenting views, and the governance record. If still running, status will be "running" — poll again in 10–15 seconds.
`board_install_kit`	no_return_description	Install a prebuilt Institution Kit — a complete governed org chart template with sealed charters ready to convene. Three kits available: "engineering-suite" (ARB + RAB for tech teams), "federal-ai-board" (ARMB + AEOB, NIST/CMMC/FedRAMP/EO14110 aligned), "eu-ai-risk-council" (HRAAB + CAB, EU AI Act + GDPR aligned). Returns the created institution ID and the list of its pre-sealed charter IDs ready for board_convene_session.
`agent_citizenship_status`	no_return_description	Query agent citizenship tier and merit score (Colony Layer 5), or trigger reassessment. Returns per action: status = the agent's citizenship tier, merit score, and component metrics (deliberation quality, behavioral health, gate approval rate, rights exercise); assess = confirmation and the updated merit assessment; leaderboard = the top agents by merit, capped by limit. Agents earn trust through deliberation quality, behavioral health, gate approval rates, and responsible rights exercise.
`branch_authority_status`	no_return_description	Colony Layer 4 — Separation of Powers. Query constitutional branch authority across three branches: legislative (creates law), executive (executes law), judicial (interprets law). Returns per action: status = the user's assigned branches and authorities; roster = all authority holders, capped by limit; violations = the branch-violation log entries, capped by limit.
`colony_convene_request`	no_return_description	Colony Autonomy: Request, list, or review agent-initiated session convene requests. Actions: request (citizen+ can request a governed session), list (view pending/all requests for a charter), review (elder+ approve/reject a request). Returns the convene request ID and resulting status for request/review, or the matching request records for list. Tier-gated: agents earn the right to request and approve sessions through demonstrated merit.
`colony_suggestion`	params_unexplained no_return_description	Colony Autonomy: Suggest, list, review, or upvote charter amendment suggestions — the petition mechanism for governed agents. Actions: suggest (citizen+ propose changes), list (view suggestions), review (elder+ promote to formal amendment or decline), upvote (citizen+ signal support). charter_id identifies the charter the suggestion belongs to or is scoped to; status_filter narrows list results to open, under_review, promoted, declined, or withdrawn suggestions. Returns the suggestion ID and its resulting status for suggest/review/upvote, or the matching suggestion records for list.
`colony_health`	no_return_description	Colony Autonomy: View colony health score, its trend over time, or trigger an on-demand health snapshot. Returns per action: snapshot = the latest health score with component breakdown (agent distribution, merit averages, deliberation quality, gate efficiency, constitutional compliance); trend = the historical scores over the requested days; pulse = the freshly computed snapshot. Health score computed from agent distribution, merit averages, deliberation quality, gate efficiency, and constitutional compliance.
`context_revive`	no_return_description	Governed context compaction — detects context pressure and restores capacity under GIA governance. Three tiers: sparkling (20-30%, INFORMATIONAL), electrolyte (40-55%, ADVISORY), iv (60-75%, MANDATORY gate). Returns per action: status = current context pressure, tier recommendation, and gate requirements; compact = confirmation with restored capacity and tier; verify = integrity result of the last compaction; history = the compaction audit trail for the session. Force may override a recommendation but NEVER bypasses MANDATORY gates. Cooldown prevents duplicate compaction per session+tier. No autonomous compaction loop.
`governed_sample`	no_return_description	Request a governed LLM completion via MCP Sampling. The client performs the model call — the server governs when, how, and under what constraints sampling is allowed. Every request is classified, policy-checked, optionally gated, and recorded in the forensic ledger. Returns the model's completion response together with the governing classification, whether a gate was required or bypassed, and the forensic ledger reference.
`srt_diagnose`	no_return_description	Run the SRT Diagnostician on an incident. Matches the finding to known playbooks, identifies root cause, and proposes a staged repair plan. Returns the diagnosis with matched playbook, identified root cause, confidence, and the proposed repair steps (which require srt_approve_repair before execution). Classification: ADVISORY — read-only analysis, no mutations.
`srt_approve_repair`	no_return_description	Approve or reject a pending SRT repair plan — the human-in-the-loop gate for SRT repairs. Repair plans CANNOT execute without explicit human approval. Pass action="approve" to approve or action="reject" to reject (reason is required when rejecting). Returns the updated incident status and whether the repair plan was authorized for execution, recorded in the audit chain. Classification: MANDATORY.
`srt_generate_postmortem`	no_return_description	Generate a structured postmortem report for a completed SRT incident. Returns a report containing the incident timeline, root cause, what worked/failed, prevention actions, real timing metrics (TTD/TTDiag/TTR), an ESTIMATED ROI (humanTimeSaved/costAvoided from a severity-bucket heuristic — not measured savings, flagged via roiEstimated/roiBasis), and optional playbook delta. Classification: ADVISORY.
`gia_list_packs`	no_return_description	List available governed operations packs. Returns the matching packs with their ID, intent (remediation/patrol/hardening/audit), category, risk level, and trust level, optionally filtered by intent, category, risk, or trust_level. Classification: INFORMATIONAL.
`gia_apply_pack`	no_return_description	Execute a remediation or hardening pack with MANDATORY gate approval. Requires inputsHash from gia_dry_run_pack (what-you-approved-is-what-ran binding); rejects patrol/audit packs (use gia_run_patrol for those). Hardening packs run preflight checks before execution. Returns the execution result with per-step status, the gate approval record, and the forensic ledger entry. Classification: MANDATORY — human approval required.
`gia_ingest_document`	no_return_description	Governed document ingestion — ingest a document into governed retrieval. Accepts raw text OR a base64-encoded file (PDF, DOCX, TXT, images). Hybrid extraction: text path first for cost efficiency, automatic Claude vision fallback for scanned/image-only PDFs (< 30 words/page triggers fallback). Content is chunked, embedded, hash-verified, and stored with full audit trail. Returns the document ID, the number of chunks created, each chunk's SHA-256 integrity hash, and the audit/ledger reference. Classification: ADVISORY — creates governed content, audited.

Selection evidence

Confusable tool pairs.

14 pairs where similar names or overlapping descriptions may send an agent toward the wrong tool.

Tool A	Tool B	Confidence	Why they collide
`verify_ledger`	`verify_ledger_v2`	high	Both tools verify the forensic audit ledger; only the in-memory reconstruction vs persisted PostgreSQL distinction separates them, and a plain task like 'verify the ledger' gives the agent no cue which to choose.
`board_approve_gate`	`approve_gate`	medium	Both approve MANDATORY gates using gate_id/approved_by/rationale; a task like 'approve the pending gate' does not reveal whether it is a board-deliberation session gate or a generic MANDATORY gate.
`evaluate_threshold`	`evaluate_routing_threshold`	medium	Both 'evaluate a threshold' over rate-based metrics; a task like 'evaluate the threshold' could target the Storey escalation rate or the model-routing health thresholds.
`generate_report`	`generate_impact_report`	medium	generate_report is the default governance status report, but the impact report also covers governance and economics; 'generate the governance impact report' plausibly routes to either tool.
`generate_impact_report`	`generate_value_report`	medium	Both generate economic+governance reports (illustrative ROI estimate vs ledger-anchored value), so a task like 'generate the economic value report' is ambiguous between the two.
`generate_report`	`generate_value_report`	low	A bare 'generate the report' defaults to the status report, but with value/economics phrasing (e.g. 'generate the report on savings') the value-report description also matches.
`board_list_institutions`	`board_list_charters`	medium	'List the boards/committees' is ambiguous because institutions are named after boards (Architecture Review Board) while charters are the actual boards/committees/subcommittees.
`load_memory_pack`	`compose_memory_packs`	medium	Both put memory packs into agent context; 'load these two packs into my context' could match load_memory_pack (single pack) or compose_memory_packs (multiple packs into unified context).
`load_memory_pack`	`transfer_memory_pack`	low	'Get this pack to agent B' could mean loading it into agent B's context or transferring it between agents; descriptions differ on provenance/gate but both deliver a pack to an agent.
`gia_dry_run_pack`	`gia_apply_pack`	medium	'Run the remediation pack' is ambiguous between previewing execution (dry run, ADVISORY) and actually executing it (apply, requires inputsHash and MANDATORY approval).
`gia_dry_run_pack`	`gia_run_patrol`	low	Shared 'check/run/pack' wording makes 'run the pack check' ambiguous between previewing a remediation pack and executing a read-only patrol/audit pack.
`colony_convene_request`	`colony_suggestion`	low	Both colony-autonomy petition tools share request/list/review actions and charter/institution params, so 'review the colony's pending requests' could route to session convene requests or amendment suggestions.
`verify_ledger`	`phoenix_verify_integrity`	low	phoenix_verify_integrity explicitly checks ledger hash-chain integrity, so 'verify ledger integrity' plausibly matches both; verify_ledger is narrowly the chain check while phoenix is the broader verification.
`monitor_agents`	`agent_citizenship_status`	low	'Check agent status' could mean the all-agents health/repair monitor or an individual citizenship/merit lookup; both describe querying agent status.

Compare the field

One score is useful.
The evidence makes it actionable.

Back to the leaderboard