rei-verify
rei-verify
반증 기계 (refutation machine) — 생성이 아닌 부정을 전문으로 하는 검증 infrastructure + MCP server.
Version: 0.1.0a1 (2026-08-19) — 4 primitives + 4 refutation tools + 8 MCP tools + integration demo. test 198/0 PASS.
왜 「반증 기계」인가
생성은 포화된다. refutation은 포화되지 않는다.
현행 LLM은 유창하다. 그럴듯한 증명의 논리, 그럴듯한 code, 그럴듯한 정리의 이름을, 사실인지 여부와 독립적으로 출력할 수 있다. benchmark가 96%까지 포화되어도, 이 구조는 변하지 않는다. 세계에 부족한 것은 「그럴듯한 것을 만드는 기계」가 아니라, 「그럴듯한 것을 확실히 죽이는 기계」 쪽이다.
반증 기계의 core promise:
주장을 받으면, 반례 탐색에 계산 자원을 할당한다. 증명 시도는 나중으로 미룬다.
반례를 찾지 못한 경우, 「찾지 못한 탐색 공간의 형태」 를 명시적으로 return (침묵을 성공으로 위장하지 않음).
출력에 반드시 「그 주장이 거짓이라면 무너지는 곳」이 첨부된다. Lean 4의
sorry제로는 이것의 가장 엄격한 특수 케이스.「반증하지 못함」 과 「옳음」 을, 타입 level에서 별개의 것으로 취급한다.
Related MCP server: leanscreen
4-value verdict (「절대 거짓말하지 않는다」 core discipline)
class Verdict(str, Enum):
CONFIRMED = "confirmed" # post-condition PASS + marker 空
REFUTED = "refuted" # 具体的な counter-witness が 得られた
HOLDING = "holding" # counter-witness 未発見 かつ marker 非空
INCOMPLETE_FRAME = "incomplete_frame" # 主張自体が well-formed でないBinary TRUE/FALSE로 하지 않음 = 반증되지 않음 ≠ 옳음. IUT 12년 holding discipline의 타입화.
「침묵을 성공으로 위장하지 않음」 타입적 보장: CONFIRMED 이외의 모든 verdict에 IncompleteMarker (dimension + what_was_tried + what_was_not_tried + reason)가 1개 이상 필수 (dataclass invariant, 깨지지 않음).
4 primitives (rei_verify)
primitive | 역할 |
| 4값 enum |
| dimension 4종 어휘 ( |
| sha256 hash-chained append-only JSONL + tamper detection ( |
| pre-check + action + post-check + audit를 atomic하게 묶는 context |
4 refutation tools (rei_verify.*)
반증 기계의 심장부. 모든 tool이 VerdictWithMarkers (4값 verdict + markers + audit_hashes)를 return하는 일관된 shape.
tool | module | 의미 | verdict pattern |
|
| Lean 4 source를 실행, sorry / native_decide / disallowed axiom을 verify | CONFIRMED / REFUTED / HOLDING / INCOMPLETE_FRAME |
|
| iterable space + callable predicate로 반례 탐색 | REFUTED / HOLDING / INCOMPLETE_FRAME (never CONFIRMED) |
|
| N labeled cases × 개별 logic의 망라 검사 | REFUTED / HOLDING / INCOMPLETE_FRAME (never CONFIRMED) |
|
| 선언적 HOLDING 생성 (「보류의 타입화」) | HOLDING / INCOMPLETE_FRAME (only) |
★ CONFIRMED를 tool이 내는 것은 refute_lean_source만 (Lean 4 kernel이 sorry-free로 인정한 case만). 다른 3 tool은 항상 REFUTED 또는 HOLDING = 「absence of counter-example is not proof」 discipline의 타입 level 보장.
8 MCP tools
Claude Desktop / Cursor / Cline 등의 LLM client에서 직접 호출 가능:
tool | 용도 |
| named audit chain 생성 |
| raw entry 추기 |
| integrity walk + tamper 검출 |
| 4값 verdict + markers를 단순 추기 (invariant enforced) |
| Lean 4 source 검증 |
| 반례 탐색 ( |
| 망라 검사 ( |
| 선언적 HOLDING |
MCP-safe expression은 restricted eval = __import__ / exec / eval / open / __ prefix 사전 reject, _SAFE_BUILTINS whitelist (abs/min/max/sum/len/int/float/str/bool/round/any/all/range)만 허용.
Installation
pip install rei-verify # core primitives (no external deps)
pip install rei-verify[mcp] # + MCP server또는 소스에서:
git clone https://github.com/fc0web/rei-verify.git
cd rei-verify
pip install -e .[mcp]Python 3.10+ 필요 (dataclass + Enum + typing 신기능). core primitives는 표준 library만으로 동작 (mcp package 없어도 import OK).
Usage — library
VerifiedExecution (custom)
from pathlib import Path
from rei_verify import (
Verdict, IncompleteMarker, PostCheckResult,
VerifiedExecution, AuditChain,
)
audit = AuditChain(Path("./reasoning.jsonl"))
ve = VerifiedExecution(
claim="1 + 1 == 2",
pre_check=lambda: True,
post_check=lambda r: PostCheckResult(refuted=(r != 2), markers=[]),
audit=audit,
)
result = ve.run(lambda: 1 + 1)
# result.verdict == Verdict.CONFIRMED
# result.audit_hashes == [h1, h2, h3, h4, h5] # 5 phase entryrefute_lean_source
from rei_verify.refute import refute_lean_source
result = refute_lean_source(
claim="trivial True holds",
lean_source="theorem trivial_true : True := trivial\n",
audit=audit,
theorem_name="trivial_true",
timeout_sec=60,
)
# result.verdict == Verdict.CONFIRMED (axiom-free, ~1200 ms)Default allow_axioms = Mathlib base [propext, Classical.choice, Quot.sound]. sorry / native_decide / disallowed axiom은 HOLDING으로 routing.
search_counterexample
from rei_verify.search import search_counterexample
result = search_counterexample(
claim="no n in [1,100] equals 42",
predicate=lambda x: x == 42,
space=range(1, 101),
audit=audit,
space_description="range(1, 101)",
)
# result.verdict == Verdict.REFUTED (witness marker: n=42)exhaustion → HOLDING (search_space marker, 「absence ≠ proof」)
time/sample budget → HOLDING (compute_budget marker)
assert_breakpoints
from rei_verify.breakpoint import Breakpoint, assert_breakpoints
result = assert_breakpoints(
claim="Collatz t1=1 orbits descend",
breakpoints=[
Breakpoint("n=27", assertion=lambda: descent(27), context={"n": 27}),
Breakpoint("n=703", assertion=lambda: descent(703), context={"n": 703}),
Breakpoint("n=6171", assertion=lambda: descent(6171), context={"n": 6171}),
],
audit=audit,
stop_on_first_failure=True, # False で 全 breakpoint 実行 (集計目的)
)임의 breakpoint False → REFUTED (label + context를 witness)
모두 pass → HOLDING (「listed checkpoints exhausted ≠ 모든 case cover」)
hold_verdict
from rei_verify.hold import hold_verdict
result = hold_verdict(
claim="my analytical claim under investigation",
markers=[
IncompleteMarker(
dimension="search_space",
what_was_tried="5 counterexample approaches",
what_was_not_tried="structural refutation via categorical semantics",
reason="categorical angle deferred to next session",
),
],
audit=audit,
notes="manual reasoning pause",
require_multi_dimension=True, # 単一 dim なら augmentation marker 追加
)
# result.verdict == Verdict.HOLDING (audit chain 4 phase entries + caller markers)Usage — MCP (Claude Desktop)
claude_desktop_config.json:
{
"mcpServers": {
"rei-verify": {
"command": "python",
"args": ["-m", "rei_verify"]
}
}
}또는 (설치된 script):
{
"mcpServers": {
"rei-verify": {
"command": "rei-verify"
}
}
}MCP expression 예 (x bind for search, ctx bind for breakpoints):
{
"tool": "search_counterexample_explicit",
"arguments": {
"chain_id": "chain-abc123",
"claim": "no perfect square in [1,100] equals 42",
"samples": [1, 4, 9, 16, 25, 36, 49, 64, 81, 100],
"predicate_expr": "x == 42",
"space_description": "perfect squares up to 100"
}
}{
"tool": "assert_breakpoints_explicit",
"arguments": {
"chain_id": "chain-abc123",
"claim": "Collatz t1=1 descent",
"breakpoints": [
{"label": "n=27 case", "assertion_expr": "ctx['descent'] < 0",
"context": {"n": 27, "descent": -0.5}},
{"label": "n=703 case", "assertion_expr": "ctx['descent'] < 0",
"context": {"n": 703, "descent": 0.2}}
]
}
}Integration demo
examples/collatz_t1_ones_lyapunov_demo.py — Collatz 홀수 n with trailing_ones(n)=1의 Lyapunov α-descent scan을 assert_breakpoints로 실행.
python examples/collatz_t1_ones_lyapunov_demo.pySample output (1,048,575 samples / 76.7 ms):
α=0.5〜0.85: WITNESS n=9 r(n)=0.885622 → α refuted ✓
α=0.9: WITNESS n=17 r(n)=0.905315 → α refuted ✓
α=0.93: WITNESS n=57 r(n)=0.930288 → α refuted ✓
α=0.95: WITNESS n=313 r(n)=0.950121 → α refuted ✓
α=0.97: WITNESS n=14,601 r(n)=0.970001 → α refuted ✓
α=0.99: NO WITNESS in range max r=0.981135 < 0.99 → α NOT refuted in sample
VERDICT: REFUTED (α=0.99 が サンプル範囲 で 未 refute = finite absence report)
audit chain: 6 entries、 sha256 hash chain intactWitness n이 α tight화와 함께 증가 (n=9 → n=14,601) = **r(n) → 1 as n → ∞**의 유한 반영을 직접 관측, tool이 「finite absence를 CONFIRMED로 자동 승격하지 않음」 discipline을 준수한 실례. 자세한 honest scope는 demo 내 주석 참조.
Test coverage
누적 198/0 PASS (6 test files):
file | assert | 내용 |
| 37 | Verdict + IncompleteMarker + PostCheckResult + VerdictWithMarkers + AuditChain + VerifiedExecution invariants + 4-verdict paths |
| 30 | tool 직접 invoke + validation + tamper detection + smoke registration |
| 22 | parse_lean_axioms + classify_axioms + pre-check + live smoke (Lean 4.33) |
| 37 | 4 exit path + per-sample error + restricted eval safety (8 hostile expr reject) + MCP tool |
| 33 | pre-check + verdict paths + stop_on_first_failure + time budget + var_name extension |
| 39 | pre-check + valid HOLDING + require_multi_dimension + invariant + MCP + 4-tool shape consistency |
# individual
python test/test_skeleton.py
python test/test_refute.py # requires 'lean' on PATH for live smoke
# all
for f in test/test_*.py; do PYTHONIOENCODING=utf-8 python -u "$f" | tail -3; doneFor maintainers: PyPI Trusted Publisher setup
본 repo의 .github/workflows/publish.yml은 v tag push로 PyPI 본반 publish* + workflow_dispatch로 TestPyPI dry-run. 사용 전에 PyPI / TestPyPI 양쪽의 Trusted Publisher 등록 + GitHub Environment 생성이 필요.
자세한 절차는 TRUSTED_PUBLISHER_SETUP.md 참조.
⚠️ tag push 전에 workflow.yml + Trusted Publisher 등록의 double-check를 (「tag = release trigger」 사고 방지).
Design
See DESIGN.md for full rationale (8절):
설계의 출발점 (Rei stack 4 원칙)
4 primitives 상세
Verdict rule table (단일 source of truth)
반증 기계 3+1 tool mapping
framing-drift-detector와의 관계
비목표 (out of scope for skeleton)
의존 (external dep 제로의 의도)
honest scope
Related
fc0web/rei-automator-mcp — Windows PC 자동화 MCP, AuditChain의 유래 (STEP 1340 AuditLogWriter의 범용 추출)
fc0web/grounded — 산문 grounding checker (2-tier verification)
fc0web/grounding-check — SCPI hardware grounding checker
Author
후지모토 노부키 (Nobuki Fujimoto)
License
MIT (v0.x irrevocable). v1.0+에서 AGPL-3.0 + commercial dual 가능성. LICENSE 참조.
Honest scope (양보할 수 없는 선)
(i) skeleton의 refutation tools는 Lean 4와의 직접 연계 (single-file
lean실행)만, Mathlib 의존 proof는 별도 iter (lake project 경유)(ii)
IncompleteMarker.dimension어휘는 초기 4종만, 확장은 operational 경험에서(iii) hash chain은 tamper detection용, cryptographic signing (Sigstore 등)은 별도 concern
(iv) restricted eval은 AST-level analysis (asteval 등)보다 약함, 고신뢰 요건은 별도 iter에서 의존 추가
(v) 「반증 기계」의 신규성 주장 제로 (
[[feedback-world-uniqueness-claim-controllable]]) = property-based testing (Hypothesis) + Lean 4 sorry-check + Coq / Isabelle 계 industry 표준의 통합 discipline layer만, novelty는 「4값 verdict + marker invariant + hash chain의 타입적 통합 + MCP wrapper」의 조합 discipline만(vi) integration demo (Collatz t1=1)는 후지모토 씨 실제 리야프노프 해석의 재현이 아님 — 단순화 V = log2(n)과 유한 sample에서의 TOOL 동작의 실증만, 진정한 reproduction은 후지모토 씨 실제 V + 조건 + Lean 4 formalization 경유로 별도 iter
(vii)
refute_lean의 "sorry-free" 판정은#print axioms의존 = Lean 자체의 kernel bug가 있으면 verify되지 않음 (kernel bug는 Rei scope 밖)
This server cannot be deployed
Maintenance
Related MCP Connectors
38 MCP tools: Rust computation, Discovery, composition, counterexample scans and signed checks.
Read-only MCP server checks and quote verification, free and keyless; results signed and publicly lo
MCP-native AI evaluation: rubric audits, eval suites, and proof reports for AI/LLM output.
Physics-based validation of simulation results: receipts with per-check verdicts, via MCP.
Related MCP Servers
- AlicenseNot gradedqualityAmaintenanceMCP server that gives LLMs access to formal verification via Z3 and SWI-Prolog, plus tree-sitter-based source code analysis. Translates natural language problems into formal logic using a template-based pipeline, verifies results with mathematical certainty, and analyzes call graphs for reachability, dead code, and impact analysis.89 npm218Apache 2.0
- FlicenseAqualityBmaintenanceA calibrated faithfulness screen for informal↔Lean 4 statement pairs, served over MCP. It provides deterministic checks and deep LLM-based analysis to help draft Lean statements.27-
- AlicenseAqualityBmaintenanceAn MCP server that provides tools for certificate verification, equivalence proving, and pre-registration sealing, enabling AI agents to re-derive verdicts from artifacts rather than trust assertions.9Apache 2.0
- AlicenseAqualityCmaintenanceMCP server that provides six verification tools (Lean proof checking, axiom audit, bound, gridlock check, certificate verification, residency check) with honest status reporting (ok/failed/unavailable) to prevent agents from claiming unchecked proofs passed.10Apache 2.0