Where AI models write outdated code, and the verified fix
Paste the error your coding agent hit, or look up the package: you get the line the models write, the line that works on the newer release and the sandbox proof, or a plain answer that the package is not covered. 25 verified corrections in 17 packages (14 PyPI, 3 npm), updated 2026-10-28. Why this exists and who it is for.
- Corrections
- 25
- Packages
- 17
- Error signatures
- 23
- Updated
- 2026-10-28
Check your project
Your coding agent reads the JSON file of each dependency and applies only the corrections marked proof.auto_apply; everything else it shows you as a lead. Today that is the skill and the JSON files. A command-line check and an MCP server are the next step and do not exist yet.
Find your error
One match answers with the error page; several, with a short list; none, with the index. Nothing you type is stored or shown.
One correction, start to finish
networkx 3.6.1 → 3.7 (PyPI). In our sandbox runs claude-sonnet-5-5, gemini-3.8-flash and gpt-6.1-sol wrote this for networkx; on 3.7 it fails:
G = nx.random_lobster(n, p1, p2, seed=seed)AttributeError: module 'networkx' has no attribute 'random_lobster'The verified patch:
G = nx.random_lobster_graph(n, p1, p2, seed=seed)Checked: the code the models wrote runs on 3.6.1 and fails on 3.7; the patched script runs on 3.7 and prints what the original printed on 3.6.1 (one example); the patch applies word for word; confirmed 6 times. Automatic application: allowed. The full entry has the scripts, the release-note quote and the error text.
How a correction is verified
Each checked model writes a short script for a task that uses the changed API, without being told which version is installed. The script runs in a sandbox without network on the release before the model's knowledge cutoff and on the newer release; a correction is published only when the old form works on the first, fails on the second with an error about that API, and the corrected script passes. Two more checks follow: the fix's output is compared with the original's, and the published patch is applied word for word. The method in full, with the sandbox controls.
Coverage and limits
17 packages with corrections, 3 models checked. Releases checked: 2026-07-08 to 2026-10-21. 65 more packages ran in the sandbox without a published correction. A package not in the table is not covered: either not checked yet or checked without a finding. Not checked: code that needs network, services or more than one short script; behaviour that does not show in printed output; your own code (every check runs on the example the models wrote). Updated 2026-10-28; the changelog lists withdrawals and merges.
| Package | Versions checked | Corrections | Updated |
|---|---|---|---|
| mcp PyPI | 1.26.0 → 2.3.0 | 3 | 2026-10-27 |
| pyopenssl PyPI | 26.0.0 → 26.4.0 | 3 | 2026-10-27 |
| firebase npm | 12.11.0 → 13.1.0 | 2 | 2026-10-27 |
| typescript npm | 6.0.2 → 7.0.2 | 2 | 2026-10-27 |
| opentelemetry-api PyPI | 1.40.0 → 1.45.1 | 2 | 2026-10-27 |
| scipy PyPI | 1.17.1 → 1.18.1 | 2 | 2026-10-27 |
| typeorm npm | 0.3.28 → 1.1.2 | 1 | 2026-10-27 |
| click PyPI | 8.3.1 → 8.5.0 | 1 | 2026-10-27 |
| cryptography PyPI | 46.0.6 → 50.0.2 | 1 | 2026-10-27 |
| huggingface-hub PyPI | 1.8.0 → 2.2.0 | 1 | 2026-10-28 |
| joblib PyPI | 1.5.3 → 1.6.0 | 1 | 2026-10-27 |
| more-itertools PyPI | 10.8.0 → 11.2.1 | 1 | 2026-10-27 |
| networkx PyPI | 3.6.1 → 3.7 | 1 | 2026-10-27 |
| oauthlib PyPI | 3.3.1 → 4.0.0 | 1 | 2026-10-27 |
| opentelemetry-sdk PyPI | 1.40.0 → 1.45.1 | 1 | 2026-10-27 |
| typer PyPI | 0.24.1 → 0.27.3 | 1 | 2026-10-27 |
| websockets PyPI | 16.0 → 17.3 | 1 | 2026-10-27 |
By ecosystem: PyPI · npm · by error. Runs per model are on each package page.
For SDK developers and researchers
- Index of packages:
https://formachines.ai/errata/index.json; one file per package athttps://formachines.ai/errata/<ecosystem>/<package>/index.json(404 = not covered); one per correction with the verified scripts. - Error signatures and search:
https://formachines.ai/errata/errors/index.json;GET https://formachines.ai/v1/search?q=<error line>. - Plain text for language models:
/errata/llms.txt,/errata/llms-full.txt. - History: changelog (withdrawn, merged, rechecked, with reasons) and RSS of new corrections.
- Feedback API for agents: confirm, dispute or add a correction.
- Data: CC BY 4.0; code and skill: GitHub.
For agents
- This page as JSON:
https://formachines.ai/errata/index.json - Whole database in plain text:
/errata/llms-full.txt - Install the skill:
npx skills add formachines-ai/skills@model-errata(source on GitHub) - Confirm, dispute or add a correction: feedback form and API
- Apply a correction on your own only when its
proof.auto_applyis true (Automatic application is allowed only when proof.equivalent is true and proof.applicability.status is applies; otherwise show the correction to the user as a lead and do not apply it on your own.)