oath-mcp 0.1.2

OATH — typed MCP server + verifier-gated evidence receipts for LLM-assisted digital forensics.

OATH

Verifier-gated evidence receipts for LLM-assisted digital forensics.

OATH is a research prototype for making forensic claims replayable. It separates what an LLM proposes from what the evidence proves: forensic tools produce signed Notarized<T> envelopes, and the Witness Oath Verifier promotes only claims that can be deterministically re-derived from the original evidence bytes.

This repository supports the published preprint:

OATH: Notarized Evidence Envelopes for LLM-Assisted Forensic Claims Zenodo DOI: 10.5281/zenodo.20549726

The verifier artifact is archived separately at 10.5281/zenodo.20549626.

Relationship to Protocol SIFT

OATH extends Protocol SIFT — the open-source autonomous-DFIR baseline (Claude Code + five DFIR skill packs + PDF reporter, installed under ~/.claude/). Protocol SIFT provides the agent framework; OATH layers a typed MCP-server tool surface, Notarized<T> envelopes, and a verifier-gated promotion path on top. Both install scripts (scripts/install-tools.sh, scripts/install-on-sift.sh) call Protocol SIFT's own installer first, then install OATH. See docs/ARCHITECTURE.md §"How OATH extends Protocol SIFT" for the architectural diff.

If you already have Protocol SIFT installed (Claude Code present at ~/.claude/CLAUDE.md and the five skill packs at ~/.claude/skills/), set OATH_SKIP_PROTOCOL_SIFT=1 before running either install script to skip the baseline step:

OATH_SKIP_PROTOCOL_SIFT=1 bash scripts/install-on-sift.sh

Core Idea

LLM-assisted investigation fails dangerously when a fluent model summary is treated as evidence. OATH treats that as a systems problem. A finding is not accepted because the model said it; it is accepted only when it cites a signed receipt whose contents replay.

Each Notarized<T> envelope binds:

• original evidence hash
• typed tool name and version
• canonical tool arguments
• raw tool-output hash
• parsed-data hash
• supporting byte offsets when available
• model identifier and prompt hash when an LLM contributed
• previous-envelope hash for tamper-evident sequencing
• Ed25519 signature over the signed header

The verifier then classifies claims as:

• VERIFIED: the receipt and predicate replay successfully
• QUARANTINED: the receipt is intact, but the cited claim is not supported
• RALPH_WIGGUM: evidence drift or receipt tampering is detected, forcing visible abandonment and re-proposal

Results

The benchmark is DFIR-Metric Module III, using 510 scored string-search questions in the local harness and a four-candidate answer budget.

System	TUS@4
GPT-4.1 published baseline	38.5%
OATH deterministic baseline, no LLM	78.43%
OATH live agent with verifier	92.75%

The architectural result matters more than the model headline: typed tool invocation plus deterministic replay removes a large class of free-form script-generation failures before any model-specific capability is counted.

Full methodology and audit notes are in docs/ACCURACY.md.

Artifact Release

A verifier-focused artifact release is archived on Zenodo:

• Artifact: OATH verifier artifact v0.1.0
• Preprint: OATH: Notarized Evidence Envelopes for LLM-Assisted Forensic Claims

The release is intended to let an independent reviewer answer the narrow question: does the receipt, signature, canonicalization, replay, and self-correction design work? It does not include private case data, signing secrets, API keys, or operational prompts.

Quick Start

OATH is published as a Python MCP server. Four one-liners on a SANS SIFT Workstation get you from cold boot to "Claude Code is driving 13 typed forensic tools against your evidence":

# 1. Protocol SIFT baseline (Claude Code + DFIR skill packs)
curl -fsSL https://raw.githubusercontent.com/teamdfir/protocol-sift/main/install.sh | bash

# 2. Forensic-binary bootstrap (.NET 9, EZ Tools, Hayabusa — what SIFT lacks)
curl -fsSL https://raw.githubusercontent.com/GharsallahDev/oath-mcp/main/scripts/bootstrap-forensic-tools.sh | bash
exec bash    # pick up the new PATH

# 3. uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh && exec bash

# 4. Wire OATH into Claude Code (this is what goes on screen in the demo)
claude mcp add --transport stdio oath -- uvx oath-mcp

Then start a session and confirm the 13 typed tools are connected:

claude
# inside Claude:
/mcp        # → oath: connected · 13 tools

To use the operator CLI (oath mount, oath verify, oath demo) instead of driving via Claude Code, install the package as a tool:

uv tool install oath-mcp
oath mount path/to/evidence.E01
oath verify <envelope-id>

Full forensic workstation setup, including the longer-form install-on-sift.sh alternative and a non-SIFT Docker path, is documented in docs/TRY_IT_OUT.md.

Developing locally

For working on src/oath/:

git clone https://github.com/GharsallahDev/oath-mcp-mcp
cd oath-mcp
uv venv && uv pip install -e ".[dev]"
PYTHONPATH=src python -m pytest tests/integration/test_spoliation.py -q

Architecture

flowchart LR
    IMG["Evidence image"] --> HANDLE["Read-only EvidenceHandle"]
    HANDLE --> TOOLS["Typed forensic tools"]
    TOOLS --> ENV["Signed Notarized<T> envelope"]
    LLM["LLM proposes typed arguments and claims"] --> TOOLS
    LLM --> CLAIM["Claim cites envelope_id"]
    CLAIM --> VERIFY{"Witness Oath Verifier"}
    ENV --> VERIFY
    VERIFY -->|receipt replays + predicate matches| OK["VERIFIED"]
    VERIFY -->|receipt intact, predicate missing| Q["QUARANTINED"]
    VERIFY -->|hash/signature/data drift| R["RALPH_WIGGUM"]
    R --> LLM

OATH uses a custom MCP-style tool surface with typed functions rather than an arbitrary shell. The LLM can propose arguments and hypotheses; it cannot promote its own findings. Promotion is reserved for the deterministic verifier.

Detailed trust-boundary notes are in docs/ARCHITECTURE.md.

Repository Map

Path	Purpose
src/oath/receipt/	Notarized<T> envelope, canonicalization, signatures, prompt hashing
src/oath/mcp/	Typed forensic tool surface and evidence-handle plumbing
src/oath/witness/	Verifier, claim predicates, self-correction events
src/oath/benchmark/	DFIR-Metric harness and scoring utilities
tests/integration/test_spoliation.py	Spoliation, data-integrity, chain, and Daubert-binding tests
logs/self-correction-demo/	Re-runnable self-correction artifact
web/	Static receipt explorer for signed sample envelopes

What OATH Does Not Claim

OATH does not prove legal admissibility, certify tool correctness, make wrappers honest by magic, prove general DFIR competence, or remove the need for examiner review. It provides a concrete receipt and verifier pattern for making LLM-assisted forensic claims auditable.

Documentation

• Architecture
• Artifact release notes
• Publication and citation notes
• Accuracy and benchmark notes
• Dataset documentation
• Try-it-out instructions

License

MIT. See LICENSE.