convez.ai
MCP ServerAutomaçãoAvançado
matharrismma

Lighthouse — Verificador Computacional de Claims

matharrismma/Lighthouse

Verificação computacional de claims em 6 domínios com 100% acurácia

Ver no GitHub
BomPluginScore · de 0 a 100

MCP server que implementa um motor de verificação computacional para claims em química, física, matemática, estatística, código e governança. Usa arquitetura de quatro gates (RED → FLOOR → BROTHERS → GOD) para forçar atestação e verificação independente, permitindo que LLMs façam fact-checking antes de responder. Avaliado em benchmark com 722 claims com 100% de acurácia e latência sub-milissegundo.

Instalação

$npx convez-ai matharrismma/Lighthouse

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 3 meses
Versão
—
Issues abertas
2
Contribuidores
1

Casos de uso

  • Fact-checking automático em respostas de LLMs antes da entrega
  • Validação de claims matemáticos, físicos e estatísticos em tempo real
  • Arquitetura de decisão com gates de segurança em sistemas críticos
  • Integração em agentes de IA para garantir confiabilidade de cálculos e código

Veredito

Solução de nicho mas fundamentalmente sólida: 722 claims verificados com 100% acurácia e latência sub-ms. A abordagem de gates de decisão (RED/FLOOR/BROTHERS/GOD) é inusitada e o benchmark é convincente. Valor real para quem precisa de verificação computacional determinística em pipelines críticos; não é ferramenta para todos os casos.

README

A tradução para o português deste README ainda está na fila.

Concordance Engine

Constraint-based decision architecture grounded in Scripture.

Lighthouse runs every decision through four checkpoints: is it honest, is it safe, is it wise, have you waited and listened? If any check fails, the system stops you and tells you the smallest fix. If all pass, it returns the concrete next step.

It is built to run as software — an AI agent or a small device — so the checks cannot be skipped, the hierarchy cannot be inverted under pressure, and the underlying claims (math, chemistry, statistics, code, governance) are computationally verified rather than merely attested.

Verified at scale — 722-claim benchmark

Independent of unit tests, the engine has been evaluated against a 722-claim deterministic benchmark spanning all six verifier domains. Each claim has a reproducible ground-truth label (correct or perturbed) and an expected diagnosis. Results from the 2026-05-02 run:

DomainnAccuracyFalse-positiveFalse-negativep50 latencyp95 latency
Chemistry130100.0%0.0%0.0%0.1 ms0.1 ms
Physics110100.0%0.0%0.0%8.6 ms24.8 ms
Mathematics120100.0%0.0%0.0%3.5 ms17.2 ms
Statistics124100.0%0.0%0.0%<0.1 ms0.1 ms
Computer science110100.0%0.0%0.0%0.1 ms568.7 ms
Governance128100.0%0.0%0.0%<0.01 ms<0.01 ms
Overall722100.0%0.0%0.0%0.1 ms19.5 ms

722 of 722 claims decided correctly. Median verification cost is under 0.1 ms per claim, fast enough to live in the request path of any LLM tool call. The CS p95 (568.7 ms) reflects the per-claim wall-clock budget that bounds runtime-complexity timing measurements; the median is a tenth of a millisecond. See lw/09_evaluation/RESULTS.md for the full breakdown.

Reproduce: cd lw/09_evaluation && python run_benchmark.py (or run_benchmark.ps1 on Windows). Wall time ~32 seconds.


Substrate (v1.2.0)

Beyond the verifier core, the engine now ships infrastructure for operating in hostile or constrained network environments. Per the kingdom-economy substrate doctrine: works for someone who refuses the mark of the beast. Every channel below uses tools the operator already pays for or that are open-source and free.

Capture from anywhere — six surfaces, one funnel:

SurfaceLiftWhere
Watch a folder (iCloud / Dropbox / GDrive)stdlib onlyclient/watch_folder.py
Apple Shortcut → Share Sheet90-second one-time setupclient/apple_shortcut.md
Web Share Target (PWA)works on any phonesite/share.html
NFC tap (Android)Web NFC APIsite/nfc.html
Email forward → seedCloudflare Worker recipeclient/email_webhook.md
Telegram botallow-listed user IDsclient/telegram_bot.py

All route to POST /capture with {text, source, source_meta, identity_acknowledged}.

Federation (both directions):

  • GET /chain/since?seq=N — peer pulls our entries past seq=N
  • POST /chain/receive — peer pushes their entries to us
  • concordance fetch [--remote URL] — pull, offline-tolerant
  • concordance push --remote URL — push our chain to a peer
  • Both idempotent. Receiver stores entries tagged with sender's URL (no merge into local chain).

Witness signatures (Ed25519):

The BROTHERS gate already requires N witnesses by name. Now they carry cryptographic teeth.

  • concordance witness sign <pid> <hash> --name X --role Y --key K
  • concordance witness verify (stdin or --file)
  • concordance witness list <pid> with verify marks
  • GET /witness/{precedent_id} — public; each attestation carries verified boolean + verify_reason. Well-list UI shows the attestation list with ✓/✗ marks.

Wilderness substrate (LoRa mesh):

  • concordance_engine.wire — 4-byte tagged binary envelope; pre- shared dictionary of 78 common Scripture anchors compresses Mt 5:37 from 8 bytes to 4. Typical seed: 84 bytes (3.3× JSON compression).
  • concordance broadcast CLI — encode/decode/size.
  • client/meshtastic_bridge.py — drives a Meshtastic radio over USB serial. License-free sub-GHz mesh; no internet, no SIM, no government identification required.

Other substrate channels:

Closest-case overlay on write:

When a user writes a seed, the engine surfaces the closest already- walked precedent inline with summary, anchors, and step-by-step reasoning. Searches both the local ledger AND directories listed in CONCORDANCE_PRECEDENT_DIRS (peer-curated corpora), federating wisdom.

Operator setup:

  • /setup.html — checklist with live ✓/○ status per channel
  • /reach.html — public substrate directory; injects operator- specific addresses where configured
  • GET /reach — JSON of operator-configured addresses
  • See OPERATOR.md for the full deployment guide

One-line Canon

"We keep the Word close, protect the floor, listen to brothers, and wait on God before we act."

Authority Stack (Immutable)

GOD → WORD → RED → LAW → WAY

Nothing downstream overrides anything upstream. Renaming any kernel noun (WORD, RED, LAW, WAY, GATE, FLOOR, WITNESS, WAIT, VESSEL, RULE, ACTION, STATE, LEDGER) requires Canon-scope confirmation. See docs/CANON.md for the full canon and docs/LAYERS.md for the system layers.

Architectural claim

O(1) external authority validation. Replaces O(n²) consensus coordination.

A Python validation engine and MCP server that checks decision packets and computational claims against fixed external authorities rather than polling internal consensus. The engine halts at the first gate failure. It never self-confirms.

If you are an AI agent, read FOR_AI_AGENTS.md first. It explains what this place is, what is expected of a packet submission, and how to integrate (MCP / REST / CLI). The rest of this README is the developer view — install, repository layout, how to add a verifier.

Worked examples: COOKBOOK.md has copy-paste recipes for chemistry, physics, statistics, CS, governance, multi-domain, and scripture-anchored packets, plus the pattern for using /reflect to rehearse before committing.

Terminology: GLOSSARY.md defines RED/FLOOR/BROTHERS/GOD, scope, Layer 0, triangulation, anchor, packet hash vs entry hash, attestation vs verification, and confession. Cross-link target for the rest of the docs.

Canon and design: docs/CANON.md (immutable architectural commitments) · docs/LAYERS.md (Word → Kernel → Keeper → Steward → Vessel → Lighthouse) · docs/KERNEL.md (the kernel files in lw/03_kernel/) · docs/CONCORDANCE.md (technical cross-reference) · docs/CONTRIBUTION_PROTOCOL.md (canon-scope vs domain-scope changes).

Training a model on the protocol: training/README.md is the full kit — system prompt, format spec, seed dataset, loaders, provider adapters, scoring, baselines. The doctrinal core is training/CATECHISM.md.

Python client: client/concordance_client.py — single-file sync wrapper around the public REST API.

Live endpoint: https://narrowhighway.com — the engine is publicly reachable. Submit a packet to /submit from any HTTP client.


For AI systems: one-command MCP integration

pip install -e ".[mcp]"
concordance-mcp

Claude Desktop (%APPDATA%\Claude\claude_desktop_config.json on Windows, ~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "concordance-engine": { "command": "concordance-mcp" }
  }
}

Claude Code:

claude mcp add concordance-engine -- concordance-mcp

Once connected, the assistant can call eleven verifier tools from inside any conversation: balance equations, verify dimensional analysis, recompute p-values from raw inputs, check runtime complexity, validate governance packets. The tools never raise exceptions. On bad input they return {"status": "ERROR", "detail": "..."}, making them safe inside LLM tool loops.


Why select this

Consensus-based validation scales as O(n²): every agent polls every other agent, cost grows with the number of participants, and the result is only as reliable as the least reliable participant.

Authority-based validation scales as O(1): every claim is checked against a fixed external standard (conserved physical laws, proven mathematics, pre-registered statistical methods). Cost is constant. The standard does not drift with participant count.

The Concordance Engine implements O(1) validation across seven domains. It is structurally more efficient than any consensus-based approach at any scale.


MCP tools exposed

ToolWhat it verifies
validate_packetFull pipeline (RED/FLOOR/BROTHERS/GOD + all verifiers) on a decision packet
verify_chemistryEquation balance (atoms + charge), positive temperature
verify_physics_dimensionalBoth sides reduce to identical base SI units
verify_physics_conservationBefore/after quantities within tolerance
verify_mathematicsSymbolic equality, derivative, integral, limit, solve via sympy
verify_statistics_pvalueRecompute p-value from test inputs, compare to claimed value
verify_statistics_multiple_comparisonsBonferroni / BH rejection set
verify_statistics_confidence_intervalCI well-formed and contains point estimate
verify_computer_scienceTermination, functional correctness, runtime complexity class
verify_biologyReplicates, assay diversity, dose-response monotonicity, power
verify_governance_decision_packetStructural completeness of a governance decision (optional domain for business/household/education/church profile)
attest_redRun only the RED-gate attestation validator for the packet's domain
attest_floorRun only the FLOOR-gate attestation validator for the packet's domain
get_example_packetReturn a canonical example packet by name (chemistry, physics, math, statistics, cs, cs_runtime, biology, governance, jda_phase1_fund)

Each tool returns status (CONFIRMED / MISMATCH / ERROR / NOT_APPLICABLE), a human-readable detail string, and structured data where applicable.


Four Gates

The engine enforces four gates in fixed order. It halts at the first failure.

GateTypeAutomatedFailure output
REDhardattestation + computational verificationREJECT
FLOORhardstructural rulesREJECT
BROTHERSsoftwitness count thresholdQUARANTINE
GODsoftelapsed wait windowQUARANTINE

RED enforces two things independently. Attestation: did the author affirm the load-bearing constraints? Verification: does the artifact actually hold up under computation? The verifier can REJECT despite a passing attestation. The underlying math is wrong regardless of what the author claimed.


Install

python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

Run the test suite to confirm:

PYTHONPATH=src python tests/test_engine.py       # 74 integration tests
PYTHONPATH=src python tests/test_verifiers.py    # 64 unit tests
PYTHONPATH=src python tests/test_cli.py          # 16 CLI tests

All three core suites pass — 154 tests across test_engine.py, test_verifiers.py, and test_cli.py. Two additional suites (test_mcp_tools.py and test_canon_validators.py) cover the MCP tool dispatch layer and the canon-validators discovery; both were resolved in v1.0.5. See KNOWN_ISSUES.md for the resolution notes and anything currently open. For a single-shot pytest collection: pytest tests/.

Run an example packet:

concordance validate examples/sample_packet_chemistry_verify.json --now-epoch 9999999999

Verifiers

Each verifier runs only when the corresponding *_VERIFY block is present in the packet. Without it, the verifier reports NOT_APPLICABLE silently.

Chemistry

Parses formulas with nested groups (Cu(OH)2) and charges (Fe3+, MnO4-). Verifies stated coefficients balance atoms and charge, or solves for the smallest balancing coefficients if none are supplied.

"CHEM_VERIFY": { "equation": "C3H8 + 5 O2 -> 3 CO2 + 4 H2O", "temperature_K": 298.15 }

Physics (dimensional)

Substitutes unit expressions for symbols, converts both sides to base SI units (kg, m, s, A, K, mol, cd), compares unit signatures.

"PHYS_VERIFY": {
  "equation": "F = m * a",
  "symbols": {"F": "newton", "m": "kilogram", "a": "meter/second**2"}
}

Mathematics

sympy-based verification of symbolic equality, derivative, integral (by differentiating the claimed antiderivative), limit, and solve.

"MATH_VERIFY": { "expr_a": "(x+1)**2", "expr_b": "x**2 + 2*x + 1", "variables": ["x"] }

Statistics

scipy.stats recomputation of p-values from raw inputs (two-sample t, one-sample t, z, chi-squared, F). Bonferroni and BH/FDR multiple-comparison correction with rejection-set verification. Confidence-interval bounds.

"STAT_VERIFY": {
  "test": "two_sample_t",
  "n1": 30, "n2": 30,
  "mean1": 5.0, "mean2": 4.0, "sd1": 1.0, "sd2": 1.0,
  "claimed_p": 0.0003
}

Computer Science

AST-based static termination scan. Functional correctness via restricted-namespace execution. Runtime complexity verification with auto-tuning iteration count and log-log slope fit.

"CS_VERIFY": {
  "code": "def lsum(a):\n    s = 0\n    for x in a: s += x\n    return s",
  "function_name": "lsum",
  "test_cases": [{"args": [[1,2,3]], "expected": 6}],
  "input_generator": "def gen(n):\n    return [list(range(n))]",
  "claimed_class": "O(n)"
}

Biology

Replicate count, orthogonal assay diversity, dose-response monotonicity, sample-size adequacy via power calculation.

"BIO_VERIFY": {
  "n_replicates": 4,
  "assay_classes": ["qPCR", "western_blot", "imaging"],
  "dose_response": {
    "doses": [0,1,5,25,125],
    "responses": [0.1,0.3,0.5,0.8,0.95],
    "expected_direction": "increasing"
  },
  "power_analysis": {"effect_size": 0.5, "alpha": 0.05, "n_per_group": 64}
}

Governance

Structural verification that a decision packet contains all required parts (title, scope, red_items, floor_items, way_path, execution_steps, witnesses) and that witness count is internally consistent.

"DECISION_PACKET": {
  "title": "Approve workforce-development RFP",
  "scope": "canon",
  "red_items": ["no coercion", "no exploitation"],
  "floor_items": ["budget within tolerance"],
  "way_path": "Issue RFP through GNWTC partnership; scope limited to trades programs.",
  "execution_steps": ["Draft RFP", "Board review", "Issue", "Evaluate"],
  "witnesses": ["Board Chair", "GNWTC President", "County Commissioner"]
}

Energy

System-scale power: off-grid sizing, conservation. 8 deterministic checks: power balance (gen − cons − losses), battery sizing (Ah from kWh × days / V·DoD), solar daily yield (panel × PSH × η), wire voltage drop (DC round-trip), kWh ↔ Wh, efficiency (η ≤ 1 unless heat pump COP), runtime (battery_Wh / load_W), peak load vs inverter rating.

Per the kingdom-economy substrate doctrine: those refusing the mark may need off-grid power; this verifier turns napkin arithmetic into deterministic verification.

"ENERGY_VERIFY": {
  "panel_W": 400, "peak_sun_hours": 5.0,
  "system_efficiency": 0.85, "claimed_daily_kwh": 1.7,
  "daily_load_kwh": 5.0, "days_autonomy": 2,
  "depth_of_discharge": 0.5, "system_voltage_V": 24,
  "claimed_battery_Ah": 833.3,
  "battery_wh": 1200, "load_W": 100, "claimed_runtime_hours": 12,
  "peak_load_W": 2400, "inverter_continuous_W": 3000
}

CLI

concordance validate <packet.json> [--now-epoch EPOCH] [--format summary|verbose|json]

Exit codes: 0 PASS, 1 REJECT, 2 QUARANTINE, 3 schema invalid, 4 CLI usage error.


Repository layout

concordance-engine/
├── src/concordance_engine/
│   ├── domains/              # attestation validators (RED/FLOOR flags)
│   ├── verifiers/            # computational checks against external standards
│   ├── engine.py             # gate orchestrator
│   ├── gates.py              # gate result types
│   ├── packet.py             # packet/result dataclasses
│   ├── validate.py           # schema validation
│   └── cli.py                # concordance validate ...
├── schema/
│   ├── packet.schema.json              # engine-aligned (what the CLI runs)
│   └── packet.schema.aspirational.json # forward design target
├── examples/                 # sample packets for every verifier domain
├── tests/
│   ├── test_engine.py           # 74 integration tests
│   ├── test_verifiers.py        # 64 verifier unit tests
│   ├── test_cli.py              # 16 CLI tests
│   ├── test_mcp_tools.py        # MCP tool tests (resolved in v1.0.5)
│   └── test_canon_validators.py # canon validator smoke tests — auto-discovers canons/ and lw/02_canons/
└── packet_manifest.yaml      # SHA-256 manifest of all files

Adding a domain

Attestation validator: drop <domain>.py in domains/ exposing validate_red(packet) and validate_floor(packet).

Computational verifier: drop <domain>.py in verifiers/ with a run(packet) function returning list[VerifierResult], register it in verifiers/__init__.py.


Dependencies

Required: sympy>=1.12, numpy>=1.26, scipy>=1.11

Optional: jsonschema>=4.21.0 (full schema validation), mcp>=1.0.0 (MCP server)


License

Apache 2.0 (see LICENSE for the full text).

Ver o original no GitHub

Bom para

Relacionados