Conversation Title Generator
System / Product: Litigant Portal
Model Type: Prompted language model for short text generation
Version: 1.0
Date: 2026-11-18
Owner: Free Law Project, individual owner pending confirmation
Status: Draft. Live deployment and model selection await confirmation.
1. Purpose
1.1 What this model does
The portal gives a conversation a short title so the litigant can find it again. When a thread has no description at the end of a chat turn, a separate model call reads the start of the exchange and proposes a label. Title generator and caller
1.2 Intended users
Litigants see the title in the chat toolbar and activity list. Authorized staff can search thread descriptions when investigating a conversation. Thread list · Audit search
1.3 Scope and limitations
The title is a navigation aid. It can summarize the topic, but has no role in deciding eligibility, recording facts, or filling a form. Only the first 4,000 characters of visible user and assistant text enter the title request. Later turns can change the focus after the title has been stored.
2. Base Model
| Field | Details |
|---|---|
| Base model name | Site-selected fast model. The default chooses an available Economy model. GPT Luna is first in the documented order. |
| Version / checkpoint | The eligible release is resolved through the site’s Bedrock catalog. The November 9 benchmark recorded openai.gpt-6-luna through Bedrock Mantle in us-east-1 as its fast model. The deployed release is unverified. |
| Provider | Amazon Bedrock. The model developer depends on the selected family. |
| License | Model-specific terms. Confirm the deployed model and applicable terms with the owner. |
| Model location | Public reference for the deployed model awaits confirmation. Fast-model selection · Model curation |
2.1 Model Modifications
Prompting only. FLP uses provider-hosted weights and controls the instructions sent with each request.
3. Data
3.1 Pretraining Dataset
FLP used the hosted model as supplied. No FLP pretraining dataset applies.
3.2 Finetuning Dataset
No FLP finetuning dataset applies.
3.3 Prompt Design and Versioning
The versioned instruction asks for at most six words, with no enclosing quotation marks or trailing punctuation. The service joins eligible messages in order with speaker roles, takes the first 4,000 characters, and sends them with that instruction. Raw file blocks, hidden messages, and tool audit rows are outside this input. Details repeated in visible text can still enter it. Prompt and input selection
3.4 Validation and Test Dataset
Controlled tests exercise model selection, usage accounting, blank and length-limited responses, title-generation failure after a chat reply, and the default activity label. No collection of titles rated by litigants for usefulness, accuracy, or privacy was identified. Title model tests · Chat failure test · Activity label test
3.5 Label Documentation
No FLP training labels apply.
4. Training & Evaluation
4.1 Design Decisions and Rationale
Using the fast-model role lets the site choose a model independently of its legal-information assistant. The call happens after the substantive reply. A title failure is caught so a completed answer remains available. A later turn can try again if the thread still has no description. Title generation and failure handling
4.2 Metrics and Evaluation
The chat benchmark records auxiliary title failures separately from answer quality. Its rubric grades the delivered legal-information answer. It supplies no title-accuracy, sensitive-detail, or findability score. Evaluation reporting
4.3 Failure Analysis
An empty response leaves the description blank and the activity list can display “New chat.” A nonempty response cut short by the model’s output limit can still be stored. The service caps the stored result at 255 characters. The six-word limit is a prompt instruction without a word-count check. Title code · Static activity list
5. Known Limitations
Early text can contain a name, allegation, or sensitive circumstance that the model repeats in a highly visible label. A title can also imply an outcome or miss a later correction. The prompt includes no explicit instruction to remove identifying details. Differences in usefulness across languages or writing styles have not been evaluated in the cited tests.
6. Deployment and Monitoring
6.1 Deployment Setup and Data Dependencies
The chat service sends the bounded text to Bedrock, saves a nonempty result as the thread description, and sends it to the open chat. Both the interactive and server-rendered activity lists use the description. Chat service · Server-rendered list
6.2 Monitoring Plan
The service records a title-call metadata row with model and reported usage on a successful response. Empty or truncated results produce a warning, and generation exceptions are logged. These signals can reveal blank or failed titles. They cannot detect a misleading title by themselves. Production review ownership, cadence, and thresholds await confirmation. Diagnostics
6.3 Re-evaluation Criteria and Process
Proposed triggers are a change to the fast model, input window, title prompt, or display surface, or a report of a misleading or identifying title. Inspect a permitted set of short, long, sensitive, and topic-changing conversations against the titles shown in the activity list.
7. Version History
| Card version | Date | Author | Change |
|---|---|---|---|
| 0.1 | 2026-11-18 | Codex, prepared for FLP review | Initial draft against portal revision b20fec4. |
8. Output Disclosure
The toolbar and activity list render the generated phrase as the conversation title without an AI-origin label. A topic label may carry little substantive information, but a title that asserts something about a litigant or case can be mistaken for a user-supplied fact. The title on its own has no attribution. A proposed improvement is to identify automatically generated titles in the activity list and toolbar and review whether sensitive details belong in them. Toolbar · Interactive list
9. Contacts
| Role | Name | Contact |
|---|---|---|
| Feature owner | Litigant Portal team. | Public contact pending confirmation. |
| AI team contact | John Gardner | john@free.law |