bubbles-anti-fabrication

name: bubbles-anti-fabrication description: Enforce the Bubbles Anti-Fabrication Policy. Use when about to claim work is done, mark a DoD item complete, write a report.md evidence section, set state.json status, or assert that tests/commands passed. Triggers include marking [x], writing "Exit Code: 0", claiming "all tests pass", recording evidence, or transitioning a scope or spec to a terminal status.

Bubbles Anti-Fabrication

Goal

Never claim work happened that did not actually execute in the current session. Evidence must come from real tool/terminal output, not from inference, expectation, or summary.

When to use

• Before marking ANY DoD checkbox [x]
• Before writing evidence into report.md
• Before transitioning a scope to Done or a spec to done
• Before asserting a build, test, lint, deploy, query, or curl succeeded
• Before composing a result envelope claiming completed_owned

Non-negotiable rules

1. Execute first, record second. Run the command in this session, observe the output, then paste the actual output. Never write evidence before the command ran.
2. Raw terminal output, ≥10 lines. Evidence sections shorter than ~10 lines are presumed to be summaries and are rejected. Copy/paste the literal terminal text, including the command line and the exit code.
3. Session-bound proof. Evidence from a prior session, a prior tool call earlier in the same task that was not re-executed for the current claim, or evidence reused verbatim across multiple DoD items is fabrication.
4. No narrative substitutes. Strings like "all tests pass", "verified", "endpoint works", "UI looks correct" are not evidence. Show the command and output.
5. No template stubs. Placeholder text like [ACTUAL output, ≥10 lines] left in place is fabrication.
6. No batch checking. Marking many DoD items [x] in one edit without per-item execution is fabrication regardless of how confident the agent is.
7. No fabricated specialists. Claiming bubbles.audit, bubbles.chaos, bubbles.validate, or any specialist agent ran without an actual sub-agent invocation in this session is fabrication.

Self-check (run before every status transition)

1. Did I actually run a tool/terminal command for this DoD item in this session?
2. Did I see the output in a terminal or tool response?
3. Am I about to paste the actual output, not what I expected to see?
4. Is the evidence ≥10 lines of raw output?

If any answer is "no", do not mark the item [x]. Execute first.

Authoritative governance modules

This skill is a discovery shim. The full enforceable policy lives in:

• agents/bubbles_shared/critical-requirements.md — Honesty Incentive, hard non-negotiables
• agents/bubbles_shared/evidence-rules.md — Evidence Provenance Taxonomy, ≥10-line rule, Uncertainty Declaration Protocol
• agents/bubbles_shared/completion-governance.md — sequential completion, deferral blocking, red/green traceability
• bubbles/scripts/state-transition-guard.sh — mechanical enforcement that runs in pre-push

Read the module that matches the specific question. Do not rewrite these rules inside agent prompts.

Mechanical detection

The framework auto-detects fabrication via:

• state-transition-guard.sh Checks 2B, 3G, 5B–5C, 7A–7B
• artifact-lint.sh (template + section presence)
• done-spec-audit.sh (recertification-only; advisory by default — see bubbles-status-transition skill for grandfather rules)

Bypassing these scripts or asking for --force flags is itself a policy violation. They do not exist.