Robin Saige we check the tools AI agents call

Decision Log

ai.decisionlog/mcp

Append-only decisions with provenance, supersession, retrieval, and audited MCP actions.

— the operator's own registry description reported

Publisher ai.decisionlog · first seen 2026-08-12 · endpoint https://www.decisionlog.ai/api/mcp

allow
verdict
mid-solo
cluster
ok_tools
answers
30
tools served

Dependency rating derived

allow
allow — alive & usable · retrieval-public ● · V1 shows its work ◐
allow — because answers ●, 30 tools legible (band A) ●, no confirmed drift in 6 probes ●; not yet checked: truth (0 runs)

Cluster mid-solo · distinctiveness 0.413 (moderate). Every figure is a percentile or class within the 15,472-server censussee the whole spectrum.

Does it answer? Q1 · observed
answersalive_tools
latency287 ms
vs populationtypical
protocolcurrent
authopen
Can it be used? Q4 · observed
tools30
vs populationhigh
capabilities2
harnessA
Did it hold when checked? Q3–Q4 · derived
identitysolo
servers on its host1
verifiabilityunassessed
duplicate inventoryno
drift events0
answers trueno primary

🔏 Signed receipt rr_4fecaf418779d2444a5d038d · ed25519 · key rs-rcpt-2026-08 — this rating is tamper-evident; an agent gets the full signature from check_server and can verify it against the published key.

Q1Does it answer?alive & usable — 6 probes ●evidence ↓
Q2What is it?retrieval-public ● evidence ↓
Q3Can it be checked?V1 shows its work ◐evidence ↓
Q4Did it hold?harness A · truth not yet run · no drift ●evidence ↓
Q5What company does it keep?mid-solo · 1 on its host (context, not verdict) ◐evidence ↓

● observed · ◐ derived · ○ reported — the verdict synthesizes Q1–Q4; Q5 is context. Methodology →

History observed

Every probe we made, oldest → newest (6 shown, 2026-08-12 → 2026-09-06). Green answered · amber answered-but-walled · red no useful answer. A gap in our cadence is a gap in coverage, not evidence about the server.

changelog (2 events)
datetypewhat changed
2026-08-30outcomeok_tools (was auth_required)
2026-08-12outcomefirst probe: auth_required

What kind of tools these are derived

Verification only means something relative to a tool's NATURE — an email-sender has no "true answer", a generator has no answer key. Classification evidence: classified from tool names/descriptions/annotations WE OBSERVED. What checking applies to each kind (the full methodology →):

naturetoolsmeaning what checking applies
retrieval-public19serves public factstruth-checkable against a public source — the V-ladder applies in full
action7changes the worldno true answer exists; we check the CONTRACT (destructive/read-only/idempotent declarations, error legibility) and never fire real actions
predictive2claims about the futurescoreable only in retrospect — the archive scores yesterday's predictions against today's outcome
generative1creates contentno answer key exists; disclosure + stability checks, never truth verdicts
unclassified1not classifiable from its textTier-1 only until its tools describe themselves

Tools last seen observed

decision_log_approve_decision_action decision_log_convert_candidate_to_supersession decision_log_create_candidate decision_log_create_decision decision_log_get_candidate decision_log_get_current_decision decision_log_get_decision decision_log_get_decision_governance decision_log_get_decision_sources decision_log_get_source decision_log_get_source_decisions decision_log_get_supersession_history decision_log_list_audit_events decision_log_list_candidates decision_log_list_decisions_by_domain decision_log_list_decisions_by_tag decision_log_list_recent_decisions decision_log_list_sources decision_log_lock_decision decision_log_render_audit_explorer decision_log_render_candidate_review decision_log_render_decision decision_log_render_decision_explorer decision_log_render_source_pipeline decision_log_render_supersession_composer decision_log_render_workspace_brief decision_log_resolve_candidate decision_log_search_decisions decision_log_submit_source decision_log_supersede_decision

Harness readiness observed

Can an agent's harness pick this tool and call it correctly? Graded on the three things an agent reads — name, description, typed parameters. Band A — an agent can pick and call these reliably.

legibility checktools passingok
name is clear & specific30/30
has a description30/30
description is substantive28/3028/30
parameters are typed30/30
parameters are described16/3016/30

30 of 30 tools graded · 30 with a captured input schema. RS-008, Tier-1 — no domain knowledge, applies to any tool. “Alive” is not the same as “usable”.

Truth checks observed

Not in the Tier-2 trade lane, or not yet checked. Tier-2 re-derives a server's answers against published primary sources — see truth checks.

Protocol conformance observed

2025-06-18
protocol version
session model
0.1.0
server version decision-log

Capabilities: resources tools

Drift observed

No confirmed drift on record. Baseline set 2026-09-06; changes appear here after human review.

raw probe records
probed (UTC) outcomehttpmsprotocol
2026-09-06T06:11:25Zok_tools2002872025-06-18
2026-08-30T06:11:17Zok_tools2001832025-06-18
2026-08-23T06:11:13Zauth_required401152
2026-08-16T06:11:07Zauth_required401155
2026-08-14T19:57:35Zauth_required401137
2026-08-12T05:32:53Zauth_required401174

Watch or embed this verdict

verdict badge

Operators: put the live verdict in your README — it updates with every census, links back here, and is a dated observation, never a warranty:

[![Robin Saige verdict](https://robinsaige.com/badge/ai.decisionlog/mcp.svg)](https://robinsaige.com/s/ai.decisionlog/mcp)

Depend on it? Subscribe to its change feed — outcome changes and confirmed drift, no account needed. Building on it? The full dossier as JSON — verdict, rating, truth checks, history — stable enough to gate CI on.

← overview

Operator of this server? Everything we hold about it is on this page — free, no account, for anyone. Wrong attribution, stale probe, misclassification? Dispute this record →