Wiki Alt Text Generator

Feedback

# Model Card — Alt Text Generator

**System / Product:** wiki.free.law  
**Model Type:** Out-of-the-box LLM  
**Version:** v1.0  
**Date:** 2026-07-02  
**Owner:** Michael Lissner  
**Status:** [x] Production &nbsp; [ ] Deprecated

---

## 1. Purpose

### 1.1 What this model does
<!-- 
Describe the task the model performs and the output it produces. Be specific enough that someone unfamiliar with the project understands what the model does and does not do.

For generative models: describe the nature of the output (e.g., extractive vs. abstractive, structured vs. free-form).
-->

When writing pages on [the Free Law Project wiki][w], this generates alt text so that inline images have decent descriptions even if wiki users aren't being diligent about writing alt text.

This helps blind and low-vision people to understand the images and helps make the images searchable in the wiki search engine.

### 1.2 Intended users
<!-- 
Who consumes this model's outputs — internal pipelines, external users, downstream systems? If external users, how is the original and model-produced data stored and surfaced to users?

For generative models: how is AI-generated content labeled or distinguished from verified legal data?
-->

Users of the wiki.

### 1.3 Scope and limitations

It only handles standard image types below 4.5MB.

---

## 2. Base Model

| Field | Details |
|---|---|
| Base model name | claude-opus-4-8 |
| Version / checkpoint | 4.8 |
| Provider | Anthropic |
| License | Proprietary |
| Link to production model location | https://www.anthropic.com/claude-opus-4-8-system-card |

### 2.1 Model Modifications
<!-- 
Describe any modifications made to the base model. Check all that apply and provide details for each.

- Continued pretraining: additional pretraining on domain-specific data before task-specific finetuning. Describe the data used, volume, and objective.
- Finetuning: supervised finetuning on labeled task data. Describe the approach (e.g., full finetuning, LoRA, RLHF) and link to training data in Section 3.
- Prompting only (generative models): no weights modified; behavior shaped through prompt design. Document prompt design in Section 3.3.
- Other: describe any other modifications (e.g., quantization, distillation, adapter layers).
-->

None.

---

## 3. Data

N/A

---

## 4. Training & Evaluation

N/A

---

## 5. Known Limitations
<!-- 
List limitations honestly. This section should be updated as new issues are discovered post-deployment. Limitations do not block deployment — undisclosed limitations do. Include bias and fairness considerations and the steps taken to minimize them.

For generative models: pay particular attention to hallucination risks, output consistency, and any guardrails in place.
-->

- **Hallucination risk:** This model may make mistakes. This is mitigated by the human in the loop.

---

## 6. Deployment and Monitoring

### 6.1 Deployment Setup and Data Dependencies
<!-- 
Describe how the model is deployed, what app it lives in, and what the upstream and downstream data and impacts are.

For generative models: note how AI-generated content is labeled and distinguished from verified legal data.
-->

This is deployed via the Wiki's repo.

### 6.2 Monitoring Plan

Internal use of the wiki should reveal any issues.

### 6.3 Re-evaluation Criteria and Process
<!-- 
Describe the conditions that would trigger a formal re-evaluation of the model. Address the following:
- What performance thresholds or drift signals trigger re-evaluation
- What events would prompt an unscheduled review (e.g., data pipeline changes, upstream schema changes, shifts in production distribution)
- The process for conducting re-evaluation — who initiates it, what datasets are used, and what criteria determine whether the model is retrained, adjusted, or deprecated
- Cadence of scheduled re-evaluations, if applicable

For generative models: also consider base model updates and prompt changes as triggers for unscheduled review.
-->

We will reevaluate if we hear consistent concerns about the quality or when the model is obsoleted.

## 7. Version History

| Version | Date | Author | Summary of changes |
|---|---|---|---|
| v1.0 | 2026-07-02 | Mike Lissner | Initial release |

---

## 8. Contacts

| Role | Name | Contact |
|---|---|---|
| Model owner | Michael Lissner | mike@free.law |
| AI team contact | N/A | N/A |
| Legal / compliance _(if applicable)_ | N/A | |

[w]: https://wiki.free.law

Model Card — Alt Text Generator

System / Product: wiki.free.law
Model Type: Out-of-the-box LLM
Version: v1.0
Date: 2026-07-02
Owner: Michael Lissner
Status: [x] Production [ ] Deprecated

1. Purpose

1.1 What this model does

When writing pages on the Free Law Project wiki, this generates alt text so that inline images have decent descriptions even if wiki users aren't being diligent about writing alt text.

This helps blind and low-vision people to understand the images and helps make the images searchable in the wiki search engine.

1.2 Intended users

Users of the wiki.

1.3 Scope and limitations

It only handles standard image types below 4.5MB.

2. Base Model

Field	Details
Base model name	claude-opus-4-8
Version / checkpoint	4.8
Provider	Anthropic
License	Proprietary
Link to production model location	https://www.anthropic.com/claude-opus-4-8-system-card

2.1 Model Modifications

None.

3. Data

N/A

4. Training & Evaluation

N/A

5. Known Limitations

Hallucination risk: This model may make mistakes. This is mitigated by the human in the loop.

6. Deployment and Monitoring

6.1 Deployment Setup and Data Dependencies

This is deployed via the Wiki's repo.

6.2 Monitoring Plan

Internal use of the wiki should reveal any issues.

6.3 Re-evaluation Criteria and Process

We will reevaluate if we hear consistent concerns about the quality or when the model is obsoleted.

7. Version History

Version	Date	Author	Summary of changes
v1.0	2026-07-02	Mike Lissner	Initial release

8. Contacts

Role	Name	Contact
Model owner	Michael Lissner	mike@free.law
AI team contact	N/A	N/A
Legal / compliance (if applicable)	N/A

5 views Last updated 4 days, 21 hours ago

Creator:

mike