PROJECTS / EVIDENCE ENGINEERING

Built to question
the conclusion.

Retrieval finds the material. Deterministic analysis exposes relationships and gaps. Role-separated review challenges the interpretation. These are distinct components, not an automatic vulnerability detector.

01 / SECURITY RAG

PUBLIC SYNTHETIC EXPORT

Find the artifact.
Keep the provenance.

Evidence ingestion and case-scoped retrieval that retain source hashes, artifact IDs, and chunk references. The public offline demo regenerates a synthetic ownership-check failure, compares a fixed control, and checks the source bytes.

Case isolation is a retrieval control, not multi-user authorization. The public demo’s semantic mode uses sparse local term weights; it is not a neural embedding model.

Review public code ↗Read the runnable case →

02 / CIPHER EVIDENCE ENGINE

PRIVATE RESEARCH TOOLING

Exact relationships.
Explicit uncertainty.

Cipher adds read-only, source-referenced analysis over the maintained evidence index. It compares chronology, identifier partitions, missing stages, contradictory assertions, controls, and exact-value provenance candidates.

Input

Explicit case scope, source references, recorded observations, and bounded evidence packets.

Analysis

Deterministic joins and comparisons. Exact matches preserve candidate origins; time proximity creates no causal edge.

Output

Source-linked observations, contradictions, gaps, and a narrow claim for human adjudication.

Source support: merged implementation inspected at bf932c08d0b8 on 27 October 2026. Private code and research evidence are not published here. The synthetic case below illustrates the method; it is not a real-product finding.

Explore the synthetic case ↗

03 / LOCAL RESEARCH TEAM

Four roles.
One accountable workflow.

Analyst

Identify literal observations and the first proof gap.

Challenger

Test alternate explanations independently.

Synthesizer

Reconcile the recorded reports and failed checks.

Verifier

Check the synthesis against frozen sources.

These are separate role passes on a shared local inference backend. Checked outputs, bounded corrective retries, and version-bound method scores guide later method selection. This changes prompts and recorded methods, not neural-network weights. Mechanical scores do not establish research truth or measured semantic improvement.

04 / RESEARCH GATE

The model cannot promote
its own answer to evidence.

Claim and research gates preserve evidence status and reporting limits. Human adjudication remains the final step.

  1. 01Acquire & preserve
  2. 02Hash & retrieve
  3. 03Analyze with Cipher
  4. 04Challenge & correct
  5. 05Gate the claim
  6. 06Human adjudication

Evidence Ledger public repository ↗ · Local lab architecture →