Callparity Claimkill

ClaimKill

CallParity spends the next CALL-E call to FALSIFY a claim. The planner writes call N+1 as the cheapest question that would flip an untested or supported node to contradicted. Questions that disclose party A's answer are dropped. We do not book appointments, cascade-until-yes, or one-shot-verify a person. We reconcile one freight fact.

preview() compiles that plan from committed fixtures and places zero calls. Live CALL-E outbound is not required to run the tests. CONFIRMED is an alias of SUPPORTED.

When to use

Use this skill when two systems already disagree about one location or SKU, Party A has a quoted claim, and you need the cheapest Party B question that could flip that claim to contradicted without leaking Party A's answer.

When not to use

Do not use this skill to book appointments, run cascade-until-yes outreach, survey a person, verify identity, or place a live call without explicit approval and stored consent.

Workflow

1. Load a claim graph. Each node has id, text, evidence.span.quote, observable_of, and status in SUPPORTED, UNTESTED, CONTRADICTED, UNREACHABLE, ABSTAIN.
2. Score every RefutationQuestion with leak_score. Discard a question that discloses Party A's answer.
3. Keep observable questions Party B could have seen. On FR-1842, a driver question about whether dock 3 was empty stays. Any question that contains warehouse said dock 3 is dropped.
4. Write call N+1 as the cheapest kept question that could flip an UNTESTED or SUPPORTED node to CONTRADICTED, within the spoken-time budget.
5. Run scripts/claimkill.py preview --fixture fixtures/FR-1842.json. Report the plan and stop. No call is placed.
6. After a host returns transcript quotes, run the merger. No falsifier leaves a node UNTESTED. Voicemail or a missing transcript becomes UNREACHABLE. Confidence below 0.45 becomes ABSTAIN.

Read references/planner.md for the leak-drop algorithm. Read references/schemas.md for node, edge, and status fields.

Demo ticket

FR-1842 is the only demo SKU. Before any call, the warehouse quote pallet left dock 3 disagrees with the driver's report that dock 3 was empty. ClaimKill asks one refute question, not a survey. After merge, dock is CONTRADICTED, arrival is SUPPORTED, and seal is UNTESTED.

FR-1900 and FR-1888 are extra pytest fixtures for abstain and unreachable. Do not narrate them as a second product story.

Preview

python3 scripts/claimkill.py preview --fixture fixtures/FR-1842.json
python3 scripts/claimkill.py merge --fixture fixtures/FR-1842.json
python3 -m pytest tests/test_claimkill.py -q

calls_placed is always 0 on this path. Phone numbers in the plan are masked.

Safety

Read references/safety.md before compiling a plan that a host might later dial.

• Phone calls are real-world side effects.
• Require explicit user intent before any live call.
• Use E.164 numbers only. Mask them in summaries.
• Do not expose credentials.
• Do not create hidden recurring schedules or duplicate jobs for the same ticket.
• Cancel by withholding approval. This skill never places the call itself.
• Stay inside one operational freight fact. Refuse medical, legal, financial, or emergency content.

Files

• scripts/claimkill.py: leak-drop planner, claim graph, merger, and preview().
• fixtures/FR-1842.json, fixtures/FR-1900.json, fixtures/FR-1888.json.
• tests/test_claimkill.py: leak-drop and merger regressions.

Read references/examples.md for FR-1842 preview and merge output.