A reviewable Howey worksheet from transaction evidence: LiTiL-Howey-2b
A structured Howey-test model that maps numbered transaction evidence to the four factors for review.
The useful output in a Howey workflow is a factor-by-factor record tied to the evidence behind it. A paragraph that sounds legal but loses the source facts is hard to check and even harder to build around.
LiTiL-Howey-2b takes numbered facts about a transaction or offering and routes them across the four Howey factors. For each factor, it returns supported, unsupported, or unclear through a fixed connection ID, plus the evidence ID it selected. The entire response is strict JSON.
The model stops at factor routing. A lawyer or compliance team still reviews the evidence, the governing authority, and the conclusion.
What goes in and what comes out
The input combines a fixed authority card with short, uniquely numbered evidence statements such as E1, E2, and E3. Each statement should contain one fact or one explicit statement of uncertainty.
The output covers:
- investment of money;
- common enterprise;
- expectation of profits; and
- efforts of others.
Every route has one allowed connection ID and one evidence ID from the request. The model also returns authority IDs from a closed list. HOWEY is required; FORMAN, EDWARDS, and TELEGRAM can appear when the stated rule on the authority card matches the selected evidence.
This shape makes the result easy to validate. Code can reject an unknown factor, an unsupported connection ID, a missing evidence reference, or an authority outside the allowed set before the output reaches a review screen.
Where it fits
LiTiL-Howey-2b belongs in the middle of the workflow.
An upstream process collects the relevant transaction facts, breaks them into discrete evidence statements, and preserves their source locations. The model then creates the four factor routes. A downstream interface can display each factor beside the exact fact the model used, while a validator checks the schema and a reviewer accepts, changes, or rejects the route.
That structure supports a worksheet, an intake escalation, or an evidence-gap view. If a factor is unclear, the system can ask for the missing fact instead of filling the gap with an assumption. If two pieces of evidence conflict, the application can surface both before the reviewer makes the final call.
What was trained
The release is a PEFT LoRA adapter for Qwen/Qwen3.5-2B. Training used completion-only supervised fine-tuning with 384 rows drawn from 48 factor-combination families. A separate 48-row tuning set covered 12 families, and the development set contained 21 cases from 21 disjoint families.
The input and output vocabulary are deliberately controlled. The authority card is versioned as howey-card-historical-2026-07-10, and the model is trained to select only the published connection IDs and evidence IDs.
The exact post-training inputs were inspected. They contained the fixed historical authority card, generic evidence sentences, constrained identifiers, and structured targets. No private post-training content was present.
What the measurements say
The saved development outputs covered 21 synthetic cases and 84 factor slots. The LiTiL adapter produced:
- 84/84 correct factor connections;
- 84/84 correct evidence IDs;
- 84/84 exact factor-and-evidence pairs;
- 21/21 cases with all four routes correct; and
- 21/21 valid JSON responses.
The matched Qwen3.5-2B base produced 34/84 correct factor connections, 69/84 correct evidence IDs, 33/84 exact pairs, and one complete four-route case out of 21 cases. Both arms produced valid JSON on all 21 cases.
This is a strong result for the published closed-universe task. The development cases use synthetic evidence packets and a controlled vocabulary, so teams should measure their own paraphrases and evidence packets before widening the workflow.
What to build with it
Build the evidence worksheet first. Keep the authority-card version, each evidence statement, each route, and the reviewer action in the same record. The interface should make changing a route as easy as inspecting why the model chose it.
The next useful layer is gap detection. When the route is unresolved, ask for the specific missing evidence. That turns the model from a free-form legal writer into a structured part of the intake and review process.
The adapter, prompt contract, saved example, and evaluation details are available in the LiTiL-Howey-2b repository.