Skip to content

Repository files navigation

받침 (Batchim)

받침 — the final consonant that supports a Korean syllable. Here: the evidence that must support every claim.

A Claude Code plugin for deep research that verifies its claims against evidence — an isolated model judges whether each citation supports its claim, a panel of independent lenses cross-checks the risky ones, and code mechanically enforces that the quoted text really exists and the numbers/dates line up. Anything it can't confirm is quarantined.

/batchim "Did the EU AI Act ban real-time biometric identification in public spaces?"
→ a cited report where every high-risk claim is panel-confirmed + source-matched,
plus an auditable, signed ledger.

Status

MVP gate built (M0–M3 code complete; formal human labeling pending). Forked from fivetaku/insane-research (MIT, see NOTICE); the verification core is essentially all new. Design: docs/PRD.md (v0.4) + docs/M0-PLAN.md. 228 tests green across 16 files.

What's built (skills/batchim-main/scripts/, bench/)

MilestoneBuiltModules
M1 entailment gaterisk classify + atomization, source snapshot+hash, isolated-verifier anchors, §6.7 decision, single joinerclassify_risk · snapshot · anchors · entail_gate · decide · validate_ledger
M1 independencecanonical-URL + simhash clustersdedup
M2 panel (MVP)N=3 prompt-diverse lenses, 2-of-3 consensus → quote-mining defensepanel
M1b durabilitysigned input-closure manifest, single-rename CURRENT commit, replay + cross-version guard, publish gate, body backstop, budget throttlemanifest · commit · replay · backstop · budget
Phase-7 §9 gateleak / missing-proof / span-match / coverage / manifest hard gateeval_report
M0 measurementCohen's κ gate, human-gold scorer, benchmark freeze/signbench/kappa · bench/score_benchmark · bench/freeze
M3 (post-MVP)semantic independence (paraphrase syndication), hybrid RAG retrievalsemantic · retrieval

Orchestration is wired in SKILL.md (Phase 3.5/4.5/4.6). Remaining: real human labelers (≥2, κ≥0.7) + third-party topic lock for the signed launch-gate number — tooling is ready (bench/, RUBRIC.md).

Measured (informal e2e — bench/baseline/m1a_informal_smoke.md)

End-to-end over 3 topics (fabrication / number-swap / quote-mining), false-entail rate:

M1 anchors only+ M2 panel
false-entail1/60/6
false-negative0/60/6

Code anchors block fabrication + number-swap deterministically; the panel closes the hard quote-mine (its refute lens cited EU AI Act Art 5(1)(h) exceptions). Informal — draft topics, no human labels, not the signed benchmark.

Why it's different

insane-research (upstream)받침
Verificationdeterministic gate over LLM-set booleansgate over isolated-verifier verdicts + N=3 panel(verdict = LLMs; span/number checks = code)
Claim↔evidencesrc_id, uncheckedverdict + verbatim-span match in a frozen source snapshot
Quote-miningno defensepanel refute-lens (in the MVP) + a human-labeled precision metric
Independencedistinct-domain count (fakeable)canonical-URL + simhash → semantic (M3) near-duplicate collapse
Auditabilitysource listsigned input-closure manifest + content-addressed run, atomic commit, byte-verified replay
Measurementnonepre-registered benchmark, κ-gated human labels, false-entail by failure-mode

We claim "an auditable, fail-closed, panel-checked entailment gate," not "truth proven in code" — and we measure the difference against the baseline on human labels.

License

MIT (see LICENSE). Derivative work — attribution in NOTICE.

About

받침 (Batchim) — verification-gated deep research for Claude Code: isolated-verifier + N=3 panel entailment, code-enforced span/number anchors, auditable signed ledger.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages