aget-check-facts

name: aget-check-facts description: Verify factual claims using a stratified 3-pass protocol (data, logic, propagation). Layer 2 Pre-Assertion Gates for the anti-confabulation architecture.

/aget-check-facts

Verify factual claims using a stratified 3-pass protocol. Operationalizes the conversational assertion provenance gap (L803) as structured verification.

Purpose

V-tests gate artifacts and ground-truth checks gate handoffs, but nothing gates mid-conversation factual claims. Unverified assertions propagate through a chain: conversation -> session record -> L-doc -> CLAUDE.md -> fleet-wide behavior. Cross-fleet evidence (4+ agents across 2+ fleets) independently identified this as a structural deficiency.

Evidence: L803 (Conversational Assertion Provenance Gap), L801 (Cross-Fleet False Positive — asserted "12/17 L-doc citations are fabricated" without verifying it was checking the wrong L-space), L756 (AGCI Verification Protocol), L802 (Convergent Architecture across 4+ agents), L807 (Incident Density = Urgency, 3x in one session), L782/remote (Stratified 3-pass discovery). CCB INTEL-002 (21 documented incidents).

Input

$ARGUMENTS

Parameters:

Examples:

• /aget-check-facts — Self-check recent conversation assertions
• /aget-check-facts the fleet has 34 agents — Verify specific claim
• /aget-check-facts --file sessions/SESSION_2026-04-10_1400.md — Verify claims in artifact
• check facts — Natural language trigger
• verify that — Alias for targeted check

Execution

Step 1: Identify Claims

Extract factual claims from the scope:

A "claim" is any assertion containing: entity names, dates, counts, identifiers, or causal relationships.

Step 2: Pass 1 -- Data Verification

For each claim, trace every noun-claim to its source:

Gate question: "Can I cite file:line, commit, or user statement for every entity name, date, count, and identifier?"

Record provenance for each datum: {file:line | commit:hash | user-stated | inferred}

Step 3: Pass 2 -- Logic Verification

For claims that passed data verification, check inferences:

Gate question: "Does my conclusion follow from the data, or am I building a story?"

Step 4: Pass 3 -- Propagation Assessment

For claims that passed logic verification, assess downstream impact:

Gate question: "If this assertion is wrong, where will it propagate?"

Step 5: Report

=== /aget-check-facts: [scope] ===

Claims checked: [N]

| # | Claim | Data | Logic | Propagation | Status |
|---|-------|:----:|:-----:|:-----------:|:------:|
| 1 | [claim summary] | PASS | PASS | PASS | OK |
| 2 | [claim summary] | FAIL | -- | -- | WARN |
| 3 | [claim summary] | PASS | PASS | FAIL | WARN |

Downgraded assertions:
- [claim]: "[inferred, unverified]" -- [reason]

Overall: [OK | WARN | CRITICAL]

Conversational Micro-Gate (L803)

After verification, apply the micro-gate to any future assertions about the checked topics:

1. Scope: Am I the authoritative source for this claim?
2. Provenance: Can I cite the source?
3. Recency: Is my source current?

If any check fails, downgrade: "I believe X [inferred, unverified]" rather than "X is true."

Requirements

Constraints

• C1: Read-only — MUST NOT modify any KB artifacts, session files, or L-docs during verification
• C2: Passes MUST execute in order (data -> logic -> propagation) — later passes depend on earlier results
• C3: If Pass 1 fails for a claim, skip Passes 2-3 for that claim (data failure makes logic/propagation checks meaningless)
• C4: MUST NOT produce false confidence — "skill ran" does NOT mean "claims are verified" (L284 Delegation Theater guard)

Traceability

aget-check-facts v0.1.0 Category: Quality Anti-Confabulation Architecture: Layer 2 (Pre-Assertion Gates)