多路径推理与学术验证引擎

Reasoning Engine

An MCP server that brings difficulty-adaptive, multi-path reasoning to Claude Code. It implements the Actor-Critic-Planner-Reflexion (ACPR) pipeline for deep research synthesis, and verifies the claims it produces against real academic evidence.

Version 0.2.0 is a single static Rust binary. The Python package is deprecated (see Migrating from Python).

What It Does

You give it a research question. It decides how hard the question is, allocates compute accordingly, explores multiple reasoning paths, scores each path, self-corrects weak paths, synthesises the best results, and then checks every claim against retrieved evidence before it reaches you.

"What is a Process Reward Model?"
  -> difficulty 0.21 -> single pass -> done in seconds

"How do PRMs interact with MCTS for test-time compute scaling?"
  -> difficulty 0.71 -> forest strategy -> 8 branches, 3 reflexion rounds

Architecture

Claude Code (LLM-powered orchestrator)
  |  runs the ACPR loop sequentially in one context — no subagents
  |  calls MCP tools for algorithmic decisions and evidence retrieval
  v
reasoning-engine (Rust, stdio MCP by default)
  - Difficulty estimation, DORA budget allocation, UCB1 branch selection
  - Dual-signal PRM scoring (Promise + Progress), Reflexion memory
  - Real academic search: OpenAlex, Crossref, arXiv, Semantic Scholar,
    Europe PMC, DOAJ, DBLP (free, no auth), fetched concurrently,
    deduplicated and ranked
  - Verifiable pipeline: claim extraction, lexical / hybrid / NLI
    verification, quality gate, attested run-pack export
  - SQLite state (same schema as the Python version), content sanitisation
  - Optional: local Streamable HTTP server and read-only audit UI

No API key is needed for the reasoning loop or academic search.

Measured with scripts/bench-startup.sh on one Linux machine, so read them as orders of magnitude rather than promises.

How It Works

The ACPR Pipeline

Difficulty-Adaptive Scaling

Installation

Pick one.

Release binary

Download the archive for your platform from Releases: linux-x86_64 and linux-aarch64 (static, any distribution), macos-universal, or windows-x86_64. Then verify it and put reasoning-engine on your PATH:

sha256sum --check --ignore-missing SHA256SUMS
gh attestation verify reasoning-engine-v0.2.0-linux-x86_64.tar.gz --repo Raoof128/reasoning-engine
tar -xzf reasoning-engine-v0.2.0-linux-x86_64.tar.gz reasoning-engine

Release binaries include the local HTTP server (local-host, off until you enable it), signed, replayable run packs (trust) and speculative retrieval (tasks).

From source

Needs Rust 1.89 or newer (1.92 for semantic).

git clone https://github.com/Raoof128/reasoning-engine.git
cd reasoning-engine
cargo install --locked --path crates/re-cli                          # stdio MCP only
cargo install --locked --path crates/re-cli --features local-host,trust,tasks,semantic,fulltext,graph,learn,fa,wasm-adapters   # everything

Docker

The image serves MCP over HTTP and always requires a bearer token:

docker build -t reasoning-engine .
docker run --rm -p 8765:8765 -v re-data:/data \
  -e REASONING_ENGINE_HTTP_TOKEN="$(openssl rand -hex 32)" reasoning-engine

Tagged releases also publish ghcr.io/raoof128/reasoning-engine for linux/amd64 and linux/arm64. The image is distroless and runs as a non-root user. It signs run packs with /data/signing.key, created on first run, so keep the volume.

Setup

reasoning-engine setup

setup asks where to keep data (default ~/.reasoning-engine), which verifier to use, and whether to enable the local HTTP server and audit UI. Both are off unless you say yes. Every question has a flag (setup --help). Without a terminal, setup never prompts, so scripts add --yes: it keeps the defaults for anything not given as a flag (setup --yes --http --ui).

Configure Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "reasoning-engine": {
      "command": "reasoning-engine",
      "args": ["mcp"]
    }
  }
}

Or register it for every project, in Claude Code or Codex:

claude mcp add reasoning-engine -s user -- reasoning-engine mcp
codex mcp add reasoning-engine -- reasoning-engine mcp

Then install the deep-research skill. Both clients load a skill from its own folder as SKILL.md:

mkdir -p ~/.claude/skills/deep-research ~/.codex/skills/deep-research
cp skill/deep-research.md ~/.claude/skills/deep-research/SKILL.md
cp skill/deep-research.md ~/.codex/skills/deep-research/SKILL.md

Features

Verifier tiers

The numbers come from reasoning-engine bench verify on 147 labelled claim/evidence pairs in tests/fixtures/verify/labelled_pairs.json. The labels were AI-written and have not yet been reviewed by a person. Lexical cannot detect contradiction at all, and embeddings cannot see negation, so hybrid mainly serves as NLI's shortlist. high_stakes runs always use nli when the build and models allow it. Models are pinned by revision and SHA-256 and checked on every load.

SciFact scoreboard

A public yardstick, not one we labelled ourselves. reasoning-engine bench verify --dataset scifact scores every built tier on SciFact's dev split (Wadden et al. 2020): 300 expert-written claims, turned into 340 claim/abstract pairs with the whole abstract as the evidence. The dataset (claims CC BY 4.0, abstracts ODC-By 1.0) is downloaded once into ~/.reasoning-engine/scifact/, checked against a pinned SHA-256, and never committed. A weekly workflow (.github/workflows/bench.yml) uploads a fresh table as an artifact.

Verifiers at commit fd4c0a940c7a, measured 2026-10-19 on 4 CPUs.

This set is much harder than the labelled pairs above. A whole abstract is a long premise for a sentence-level NLI model, so nli almost never reaches its 0.7 entailment bar: when it says supported it is right every time, but it finds only 9% of the supported pairs. It is still the only tier that catches contradictions. Treat these numbers as the baseline the next verifier has to beat.

Academic search and the verifiable pipeline

Search is live by default and fans out concurrently across seven free sources:

reasoning-engine scholar search "MCP prompt injection" --limit 3

The full pipeline searches, extracts claims, verifies them, runs the quality gate and exports an attested run pack. The gate returns blocked when a cited claim is unsupported, unverified or contradicted. It returns pass_with_warnings when a cited claim is only partially_supported (read those before trusting the report) or evidence gaps remain. Otherwise it returns pass.

reasoning-engine research "Retrieval augmented generation reduces hallucination" \
  --draft "Retrieval augmented generation reduces hallucination in large language models."
reasoning-engine verify-runpack ~/.reasoning-engine/runs/<run_id>

Environment variables

An empty value counts as unset. Tokens are never stored in SQLite or run packs.

Trust and reproducibility (trust)

Signed run packs. Every export is signed with Ed25519 over its run_pack_hash, which already covers every artifact. The key is created on first use, with mode 0600. Each signed export is also appended to a hash-chained transparency log in the database.

reasoning-engine keys show                      # your public key
reasoning-engine verify-runpack <pack> --public-key <base64>

verify-runpack checks the artifact hashes, the signature (pinned to a key when you pass one) and inclusion in the log. It exits:

• 0: ok;
• 1: tampered (an artifact, the signature, the key or the log disagrees);
• 2: unknown (the hashes match, but the pack is unsigned).

Someone who edits a pack and re-signs it with their own key fails both the pinned key and the log.

Replayable runs.

reasoning-engine research --record "..." --draft "..."
reasoning-engine replay <run_id>     # offline; exits 1 if any artifact differs

A recorded run pins its clock and ID seed, and keeps every provider response in the pack (http/ and replay.json). replay re-runs the pipeline with no network and compares every artifact hash with the attestation; a response that was not recorded fails closed.

Recordings never contain:

• API keys or your contact address (they are stripped from URLs before keying);
• secret response headers;
• any configured secret that a provider echoes back in a body.

Providers may still echo request metadata, though: DBLP's JSON includes the caller's IP address. So share recorded packs the way you would share request logs.

References in your profile's style. Packs gain references.<style>.md, rendered with CSL and listed in the attestation:

sources.bib stays for compatibility.

Numbers must match. A quantitative claim cannot be supported unless one of its numbers appears in the evidence it cites:

• matching uses 2% relative tolerance and compatible units (%, percentage points, ms/s, x);
• ranges, confidence intervals and Persian digits are understood;
• years and confidence levels never count as the match.

Otherwise the claim is capped at partially_supported and flagged for review.

Speculative retrieval (tasks)

Prefetch. Call plan_research_angles_tool with prefetch: true and it answers at once with the angles, each carrying a search_query. One background task then searches the bare query and every angle query concurrently.

• Clients that support the MCP tasks extension can poll progress with tasks/get or cancel it.
• Every other client, Claude Code included, still gets an ordinary result and a warm cache.
• The /deep-research skill passes prefetch when the server offers it.

Evidence cache. Each source's results are cached in the database per source, normalised query and limit: arXiv for 6 hours, everything else for 24.

• Only successful results are cached.
• A hit is rebound to the asking run but keeps its original retrieved_at.
• Recorded runs never use the cache, so replay stays exact.

Cost ledger. Every search appends a retrieval_cost provenance event with:

• bytes fetched;
• OpenAlex credits;
• sources fetched and cache hits;
• estimated tokens (characters / 4);
• elapsed time.

Each pipeline run adds a cost_ledger total before export, so the costs land in provenance.jsonl.

Full text and source quality (fulltext)

Verifying against the paper, not the abstract

With fulltext, the pipeline reads the open-access full text of the five best-ranked citable results that have one:

1. Europe PMC's JATS XML when the record has a PMCID (no PDF parsing needed);
2. otherwise the open-access PDF URL that OpenAlex reported (pdf-extract).

Each paper is split into labelled sections (abstract, introduction, methods, results, discussion, limitations, conclusion, other) and then into chunks of about 60 words.

• For every claim, the verifier sees the three chunks per paper that share the most terms with it. Records without full text keep their snippet.
• Background is not a finding. If support comes only from a paper's abstract or introduction, the claim is checked again against the rest of the paper. When nothing else supports it, it is capped at partially_supported, and the rationale names the section.
• Support from the body names its section in the rationale (Full text: results.).
• Each run records what was read in a full_text provenance event.
• The numeric check searches the whole paper.

Fetching safeguards:

• bodies are capped at 20 MB;
• PDF URLs and redirects to loopback, private, link-local or local-only hosts are refused;
• a PDF parser panic is caught and recorded as malformed_response;
• parsed papers are cached in ~/.reasoning-engine/fulltext/.

Recorded runs fetch full text through the same tape, so replay stays offline and byte-identical. Offline mode (REASONING_ENGINE_OFFLINE=1) fetches nothing.

Benchmark: reasoning-engine bench verify --full-text tests/fixtures/verify/fulltext_pairs.json. It uses 43 claims about three CC BY papers, each scored three ways: abstract only, the abstract split into chunks, and full text.

Abstract-only NLI predicts nothing as supported or partly supported (its partially_supported precision is 0). Splitting the abstract into chunks barely helps, so the gain comes from reading the paper.

Two caveats:

• The gold labels treat background-only claims as partially_supported. That is the same policy the cap enforces, so part of that class's score holds by construction.
• The claims were written and labelled by the implementer and have not yet been reviewed by a person.

Source quality

OpenAlex and Europe PMC report what kind of work each result is. With fulltext, that drives ranking per research profile:

• Each result's source_type becomes systematic_review, clinical_guideline, preprint, peer_reviewed_article, technical_report, dataset or book. Results a provider says nothing about keep their old type.
• Types the profile prefers gain 0.1; types it treats as weak lose 0.2. Results are re-ranked.
• Retracted work is flagged (risk_flags: ["retracted"]). It stays in the evidence ledger but never supports a claim.
• Recorded runs store the setting in replay.json, so replay reproduces them exactly.

Without the feature, labels and scores match the Python server.

Scrutiny (graph)

Four checks that ask "is this claim as solid as it looks?".

Evidence graph and retraction cascade. Each record with a DOI is looked up in OpenAlex (one batched request per 50 DOIs, cached for a week). Records are linked when one cites another or they share an author.

• Independence. A claim's independence is the number of separate clusters among its supporting records. Two papers from the same group, or one citing the other, count once.
• Mermaid diagram. The report draws one for every claim with more than one supporter.
• Retraction cascade. If every supporter of a claim is retracted, or cites a retracted record in the run, the claim is capped at partially_supported. cascade.json (attested with the pack) lists the citation paths.
• Limit. Only links between the run's own records are seen.

Contradiction search. After verification, each supported claim (up to five) is searched again with a deterministic contradiction query, such as X does not improve Y or X replication failure. The verifier then judges the claim against the results.

• This only runs with tiers that can say contradicted, which today means nli.
• Survivors gain +0.05 confidence.
• Contested claims are flagged for review and listed in the report's ## Counter-view, alongside contradicted and partly supported claims.

Conformal calibration (nli tier).

• reasoning-engine bench calibrate fits a split conformal threshold on the labelled pairs, reports held-out coverage and stores ~/.reasoning-engine/calibration.json.
• Each run then records, in a calibration provenance event, every verdict's prediction set and whether the verdict's label is in it.
• On the 147 labelled pairs, held-out coverage was 0.873 at a 0.90 target, and 0.775 at 0.80.

Planning.

• Adversarial angle. With REASONING_ENGINE_ADVERSARIAL=1, hard queries (beam search or forest) end with an adversarial angle of contradiction queries.
• Budget re-estimation. Under REASONING_ENGINE_DORA=directional, select_next_branches re-estimates the budget once, after the first evaluation round. It uses how much the branches disagree and how well they score, and moves at most one strategy tier.

Recorded runs store every switch, and the calibration itself, in replay.json. OpenAlex lookups and contradiction searches go through the tape, so replay stays offline and byte-identical.

Learning and living reports (learn)

A prior learned from your own sessions. reasoning-engine prior train turns every session with two or more scored branches into training data.

• Input: what was known when each branch was registered: its trace words (hashed), depth, length, sources, whether it has a parent, and the session's difficulty.
• Label: whether the branch ended in the top three by score.
• Model: an L2-regularised logistic regression, stored as ~/.reasoning-engine/prior-v<N>.json with held-out log-loss against the base rate. Older versions are kept.
• Minimum data: it refuses to train on fewer than 30 sessions.

With a prior in place:

• register_branch seeds a new branch's score (and returns prior_score), weighted by min(1, sessions / 200).
• select_next_branches lists prior_disagreement when Claude's score and the prior differ by more than 0.4. That's a cue to reflect.
• prior show describes the prior in use.

Evidence memory. Every retrieved source is remembered across runs, along with how the final verdicts used it (supported, partially_supported, contradicted or uncited). recall_memory_tool with include_evidence: true returns {memories, evidence}, ranked by BM25. Without the flag, the result keeps its usual shape.

Living reports. reasoning-engine watch <run_id> re-runs a finished run's research, using the claims in its pack, and compares the new verdicts with the old.

• It writes runs/<run_id>/watch/<date>.diff.md, covering changed verdicts, sources newly retracted or flagged, sources gained or lost, and new retraction cascades.
• It exits 3 when anything changed, so it slots into cron or CI.
• --every 7d keeps watching.

Persian (fa)

Ask in Persian, verify against the English literature.

• Spelling. Arabic yeh and kaf, short vowels, tatweel, Persian and Arabic-Indic digits, and stray or missing half-spaces (ZWNJ) are normalised, so ميشود, میشود and می‌شود match. The lexical tier then counts Persian terms as well as English ones.
• Cross-lingual embeddings. With semantic too, any claim set containing Persian is embedded with bge-m3 (int8 ONNX, 570 MB, downloaded once and pinned by SHA-256), so a Persian claim and an English abstract share one vector space. English-only text stays on BGE-small, unchanged. bge-m3 uses its own cut-offs (0.70 supported, 0.62 partial).
• NLI. The NLI model reads English only. For a pair with Persian on either side, nli falls back to the hybrid decision and says so in the rationale.
• Report. A Persian query gets report.md in Persian, right to left, with statuses and the gate verdict translated. Set REASONING_ENGINE_PERSIAN_DIGITS=1 to show the query and claims in Persian digits.

On 30 Persian claims checked against English evidence (tests/fixtures/verify/fa_pairs.json), hybrid finds half of the supported pairs, at precision 1.000. That is up from 0.2 recall with the English cut-offs. No contradicted or neutral pair is marked supported. lexical cannot match across languages at all. The claims are AI translations of pairs from the English set, not yet reviewed by a person, and the cut-offs were chosen on this same set.

Persian databases (SID, Magiran, ISC) are not built in: none of them could be reached from the build environment to record test fixtures. A WASM adapter can add any of them.

Third-party sources (wasm-adapters)

Add a search source without forking the engine: drop a WebAssembly module into ~/.reasoning-engine/adapters/ (or REASONING_ENGINE_ADAPTERS_DIR) and name it in REASONING_ENGINE_SOURCES, like a built-in.

cp my-index.wasm ~/.reasoning-engine/adapters/
REASONING_ENGINE_SOURCES=openalex,my-index reasoning-engine research "..."

Each adapter runs in a fresh sandbox per search:

• Fuel: 10 million instructions, then it is cut off.
• Memory: 64 MiB.
• No WASI: no filesystem, clock, environment or sockets. The host offers only http_get and log.
• Network: http_get fetches over HTTPS, only from the public hosts the adapter declares in allowed_hosts, through the engine's own client. That means recorded runs tape it and replays stay offline.

A broken or missing module is skipped with a warning. The ABI and a working example (in WAT, so no toolchain is needed) are in examples/adapters/hello-adapter.

Local HTTP server and audit UI

stdio is the default and needs no network. The HTTP server is opt-in (local-host feature, then setup --http, or --http on the command line):

reasoning-engine serve --http                 # http://127.0.0.1:8765/mcp
reasoning-engine serve --http --ui            # plus http://127.0.0.1:8765/ui/

Safety defaults:

• The server binds to loopback only, and accepts only loopback Host headers (DNS-rebinding protection).
• The MCP endpoint refuses any request that carries a browser Origin header, so web pages cannot drive the tools.
• A public bind needs --unsafe-bind-public and a bearer token, which serve --bearer-token-env VAR reads from the environment.
• Tokens are compared in constant time and never logged.
• Request bodies are capped at 1 MiB.

The audit UI is read-only and shows:

• recent sessions, each with its branch tree, scores and reflections;
• research runs, each with its claims and their verification status, evidence with risk flags, gaps and provenance.

When a token is set, the UI asks for it: the browser shows a sign-in prompt, and you enter the token as the password with any username. Pages carry a strict Content-Security-Policy with no scripts.

For Notion AI Custom MCP through a Cloudflare tunnel, run ./run-notion-mcp-laptop.sh (or double-click run-notion-mcp-laptop.command on macOS). See Notion Laptop MCP Tunnel.

MCP Tools

Tool names, arguments, results and error messages match the Python server byte for byte after normalising IDs and timestamps. CI checks this on every change with tests/parity/run_parity.py.

Migrating from Python

1. Install the binary (above).
2. In .mcp.json, change command from /path/to/.venv/bin/reasoning-engine to reasoning-engine. The arguments stay ["mcp"].
3. Keep your data. The Rust engine opens the Python database unchanged. Python's default location was inside the package (src/reasoning_engine/reasoning.db), so either move that file to ~/.reasoning-engine/reasoning.db or set REASONING_ENGINE_DB to its path.

serve --transport http still works as an alias for serve --http. The Python package (pip install -e ., Dockerfile.python) stays for 0.2.x with a deprecation warning, and is removed in 0.3.0.

Project Structure

reasoning-engine/
  crates/
    re-core/        # Pure algorithms: difficulty, DORA, UCB1, sanitiser, ranking
    re-model/       # Records shared by storage and verification
    re-store/       # SQLite (Python-compatible schema), FTS5 memory recall
    re-retrieval/   # Seven academic adapters + concurrent aggregator
    re-verify/      # Claims, verifier tiers, quality gate, run packs
    re-mcp/         # The 20 MCP tools; stdio and Streamable HTTP
    re-ui/          # Read-only audit UI (local-host)
    re-cli/         # The reasoning-engine binary
  tests/            # Golden vectors, recorded HTTP, parity harness, labelled pairs
  src/              # Deprecated Python implementation (removed in 0.3.0)
  skill/            # deep-research skill (Claude Code and Codex)

Background

• Process Reward Models score individual reasoning steps rather than only final answers, which gives tree search dense feedback.
• DORA (Direction-Oriented Resource Allocation) uses score variance to switch dynamically between exploring many paths and exploiting the best one.
• Reflexion feeds textual critiques back into the prompt, so the model can self-correct without weight updates.
• UCB1 selection balances trying promising branches against exploring under-visited ones.
• ReAct / Self-RAG style evidence checks keep retrieval and verification explicit before synthesis.

Documentation

• Architecture Overview
• API Reference
• Usage Examples
• Rust rewrite plan
• Contributing
• Security Policy
• Code of Conduct

License

MIT