Legal AI that cites — or declines.
ThesisLogic is an open-source legal AI workbench built around one uncompromising rule: no unverified citation ever reaches an attorney. Every authority in every answer is checked against your jurisdiction's validated corpus. What can't be proven is withheld — visibly, auditably, every time.
Fluent is not the same as true.
Courts across the United States have sanctioned attorneys for filings that cited cases which do not exist — invented by AI tools that were rewarded for sounding confident, not for being right. Bar associations now treat generative-AI competence as a professional responsibility issue.
The standard industry answer is a disclaimer. ThesisLogic's answer is architecture: verification isn't a review step someone might skip — it's a gate the output physically cannot bypass.
Generates from model memory. Citations sound plausible. A footnote asks you to "verify important information." The failure is discovered in court.
Generates only from a validated evidence package. Every citation is parsed and checked against the corpus. Unverified → the answer is downgraded, visibly, and logged.
Answers every question, including ones its sources can't support — the harder the question, the smoother the hallucination.
Measures whether retrieved evidence is actually responsive. Weak evidence → it declines, in plain language, and tells you why.
Five stages. One guarantee.
Research and drafting run a pipeline where the model is the least trusted component. Document workflows — summaries, chronologies, comparisons, privilege review — are fully deterministic and never call a model at all.
Typed retrieval
Exact-citation, case-name, weighted full-text, and semantic search over your jurisdiction pack return authority records and holding-level spans — never anonymous text chunks.
Evidence package
Ranked authorities, their support-eligible language, and the exact set of citations any generation is allowed to use. Case captions can never count as support.
Deterministic answer
A complete memorandum is built from the evidence alone — before any model runs. If the model fails, this is what you get. It's always safe.
Constrained generation
Your model — local or cloud — writes under a citation contract: allowed authorities only, exact strings only, approved decline language for unsupported points.
The proof gate
Every citation-shaped string in the draft is verified against the corpus. One unverified citation → the draft is rejected, corrected once, or downgraded. Everything is audited.
Trust features, not trust promises.
Local AI or cloud AI — your call
Run llama.cpp, Ollama, or vLLM entirely on your own hardware so client material never leaves the building; or use the Anthropic API under your firm's data terms; or run with no model at all — every workflow still works.
Any jurisdiction, by design
All jurisdiction knowledge lives in a data pack: authorities, citation formats, practice-area taxonomy, disclaimers. Scaffold a pack for any state in minutes; a step-by-step adoption guide walks your firm through migration.
Forensic audit trail
Every answer records the model, the retrieval mix, the evidence package, the proof-gate verdict, and any downgrade — reconstructable per matter, displayed beside every answer.
Matter isolation
Documents and results are scoped to user + matter, enforced server-side from the session. Case knowledge never bleeds across matters that merely share a practice area.
Deterministic-first workflows
Chronologies, comparisons, privilege flags, and summaries are computed, not generated — reproducible outputs with no model in the loop.
Shadow mode for rollout
Evaluate any model safely: attorneys see only deterministic answers while live output accumulates in the audit trail for review before you ever promote it.
Audit the tool that audits your AI.
Trust infrastructure for the legal profession shouldn't be a black box. ThesisLogic is Apache-2.0 licensed — every rule the proof gate enforces is readable, testable code.
- Single Python service + SQLite — a firm's IT generalist can run and back it up
- No telemetry; network calls go only to the model endpoints you configure
- Validation prompt suite included — 22 automated checks across every workflow
- There is deliberately no flag that lets unverified citations through
# Quickstart — any US jurisdiction git clone https://github.com/alentra-dev/thesislogic pip install -e thesislogic # 1 · scaffold your jurisdiction pack thesislogic pack scaffold my-state \ --jurisdiction "My State" thesislogic pack build my-state # 2 · pick your AI posture (or none at all) export THESISLOGIC_GENERATION_PROVIDER=openai_compatible export THESISLOGIC_GENERATION_BASE_URL=http://127.0.0.1:8080 # 3 · verify, then serve thesislogic doctor thesislogic serve # → workspace at :8600
Professional duties, mapped to running code.
Responsible AI in a regulated industry isn't a policy document — it's a set of engineering constraints. ThesisLogic is a working demonstration of that thesis, and a template for applying it anywhere the cost of a confident error is unacceptable: law, healthcare, finance, compliance.
| Professional duty | Enforcing mechanism |
|---|---|
| Candor — no false authority | Proof gate: unverified citations are structurally blocked; downgrades are visible and logged |
| Competence with technology | Provenance rail: every answer shows its mode, model, evidence, and verification result |
| Confidentiality | Fully local operation supported; zero telemetry; cloud AI is an explicit, revocable choice |
| Supervision of AI assistance | Shadow mode, professional-review notices, review-only privilege flags |
| Recordkeeping & accountability | Append-only audit trail keyed by request — every output is reconstructable |
| Honesty about limits | Retrieval-confidence floor: weak evidence triggers a plain-language decline, not a fluent guess |
Built by a practitioner, not a press release.
Udonna Eke-Okoro
I design and ship AI systems for environments where being wrong has consequences — legal, and other highly regulated domains. My working thesis: the path to trustworthy AI is verifiable architecture, not vibes. ThesisLogic is that thesis in production form: retrieval you can inspect, generation you can constrain, and verification you can't turn off.
I take on a small number of consulting engagements — AI strategy and architecture for regulated industries, responsible-AI reviews of existing systems, and bespoke builds on the patterns you see here. If your organization needs AI it can defend in front of a regulator, a court, or a board:
contynu.com · github.com/alentra-dev/contynu
Let's talk about AI your organization can defend.
Consulting engagements, speaking on responsible AI in regulated industries, ThesisLogic deployment help, or press — I read everything sent through this form.
Prefer email? Reach me at reveal address. This site sets no cookies and uses privacy-respecting, aggregate-only analytics.