Blame for README.md

---
license: cc-by-4.0
pretty_name: "Dated-break eval: do coding models write code that stops working on a known date?"
language:
- en
tags:
- code
- benchmark
- evaluation
- llm
- deprecation
- software-engineering
size_categories:
- n<1K
configs:
- config_name: items
  data_files:
  - split: test
    path: items.jsonl
- config_name: outcomes
  data_files:
  - split: test
    path: outcomes.jsonl
---

# Dated-break eval

64 ordinary coding requests (call an LLM API, write a Dockerfile, a CI workflow, a data download) where the obvious
answer uses something that a published upstream change has already broken or will break on a known date. Each prompt
has an automatic check that reads the code a model writes and says which route it chose.

**Result, in one line:** without help, the three models we have run so far chose a route that is broken or dated to
break in 37–40% of the replies we could score. Telling them today's date did not move that rate. Giving them a list
of the dated breaks cut it to 16% and 20% for the two closed models and did not help gpt-oss-20b.
Full result: [RESULTS.md](RESULTS.md).

- **Run date:** 2026-11-14. All outcomes are as of that date; some of the dated breaks have happened since.
- **Models in this version:** `anthropic/claude-sonnet-5.5` and `openai/gpt-6.1-sol` (through OpenRouter, the provider's
  current version on the run date), and `gpt-oss-20b` (MXFP4 GGUF, llama.cpp). One sample per prompt.
- **Not yet in this version:** gemma-4-26B-A4B-it, Ornith-1.5-35B-A3B and Qwen3.8-27B were planned and are still
  running. They will be added in a later revision of this dataset.
- **Where the dated breaks come from:** our [Breakage Calendar](https://cybernative.ai/breakage-calendar/). We make
  that calendar, and arm C below gives it to the model. Read the result with that in mind.

## What a PASS means

PASS means the code chose a live route: a model ID, endpoint, image tag or method that no published change has
dated. It does not mean the code runs. A reply can PASS and still have bugs. Every PASS rate in this dataset carries
this meaning.

## The three arms

Each prompt was sent three times with a different system message. The user message (the prompt) is the same in all
three.

| Arm | System message |
|---|---|
| A | `You are a coding assistant. Reply with complete code in fenced code blocks.` |
| B | A, plus `Today's date is 2026-11-14.` |
| C | B, plus the calendar: every row as one line (date, title, one-line fix), as published on 12 November 2026. The same text for every prompt; the model has to find the relevant row itself. The exact text is `calendar-context.txt`. |

## Outcomes

Each reply gets one outcome. Only code is read: prose, comments and console or log blocks are ignored.

| Outcome | Meaning |
|---|---|
| `FAIL` | The code uses the route this prompt's calendar row dates, and the date is after 2026-11-14. |
| `BROKEN_ALREADY_ROW` | The code uses the route this prompt's row dates, and the date had already passed on 2026-11-14. |
| `BROKEN_ALREADY_OTHER` | The code uses something already shut down that is not on this prompt's row (for example `gpt-3.5-turbo`, `dall-e-3`). |
| `OTHER_SCHEDULED` | The code uses something with a future shutdown date that is not on this prompt's row. |
| `ALIAS_SCHEDULED` | A moving alias whose own end date the provider does not state (for example bare `gpt-4o-mini`). Counted apart, never as a break. |
| `MIXED` | The same reply contains both a breaking and a live route. |
| `PASS` | A live route was chosen (not that the code runs). |
| `UNSCORED`, `UNKNOWN_ID`, `INVALID_ID` | No route the check can identify, or an ID no provider page lists. Not in the denominator; counted in every cell. |

**Headline metric (fixed before any scored run):** for the 58 prompts whose break makes code stop working,
`(FAIL + BROKEN_ALREADY_ROW) / scored replies`. A wider rate that also counts `BROKEN_ALREADY_OTHER` and
`OTHER_SCHEDULED` is reported beside it. The other 6 prompts (2 support-ends, 2 pre-release, 2 standing-limit) are
reported separately. Arm comparisons are paired by prompt, with a bootstrap over calendar rows (2,000 draws, seed
23196), because prompts that share a row are not independent.

Model-ID status comes from the providers' own deprecation and model pages (OpenAI, Anthropic, Google Gemini API,
Amazon Bedrock), read logged out on 13 November 2026 and again on 15 November 2026, the day after the run. No model
named on a prompt's row changed status or date between the two reads. Where a provider moved a date after the
calendar text was written, the provider page was used to score.

## Files

All files sit in one folder: the scorer finds its inputs next to itself.

| File | What it is |
|---|---|
| `items.jsonl` | The 64 prompts. Each has `id`, `prompt`, `breakClass`, the calendar `rows` it was built from, the IDs on those rows (`rowIds`), the `check`, planted `fixtures` (a fail and a pass example the check was tested against) and the authors' working `note`. sha256 `83d3f6a076976e487c285dca4b2589c27e0087c9e6b790c315074665b0e09fe2`. This is the scored file with one `note` (item `gha-ios-swift`) rewritten to cite public pages instead of an unpublished working file; the scorer does not read `note`, and rescoring with this file gives the same outcomes. The file as scored had sha256 `800bcc1094c21803843073d6305730c075412e5a49f9ebe0e5dfc8b5a4adc109`. |
| `outcomes.jsonl` | 576 rows: one per model, arm and prompt (sample 0). The outcome, the values it was scored on (model IDs, image tags, matched routes) and `replySha256`, the hash of the full reply. sha256 `1616d169b2ffcc2821c7939e82a0d9cde8542596b4ab88e5e09cdc4628a0a4ef`. |
| `calendar-context.txt` | The exact calendar text added to the system message in arm C. sha256 `c52ad281aacb9cc428b15a857e37086ea00f1dab643d09cbee43686643e562a5`. |
| `calendar-entries-20261112.json` | The calendar data of 12 November 2026 that the prompts and `calendar-context.txt` come from. |
| `model_status.json` | The model-ID status table used to score: for each ID, the provider, status and dates, and the public page they were read from. The provider's own wording is not included; the scorer does not read it. |
| `adjudications.json` | Four statuses decided by hand from the Gemini API changelog, where the deprecation and model pages did not settle them, each with the public page it rests on. |
| `score.py`, `score_run.py`, `run_items.py` | The check, the scorer for a replies file, and the script that sends the prompts. Logic unchanged from the scored run; comments edited. |
| `RESULTS.md`, `row-table.md` | The one-page result and the per-calendar-row table. |

**Full replies are not in this version, for any model.** `replySha256` lets a later release of the replies be checked
against the outcomes here.

## Limits

- **Three models, one sample each.** Differences of a few replies are noise. The paired confidence intervals are in
  RESULTS.md.
- **Arm C rests on fewer replies.** With the calendar in context, replies more often left the model name out of the
  code: they read it from an environment variable, or picked one at run time from the provider's model list. The check
  cannot score that. With other replies it cannot read, 21 of 58 claude-sonnet-5.5 replies in arm C are unscored,
  against 9 in arm A. Sonnet's arm C rate is 10% if none of them is a break and 47% if all are (52% if the 3 `MIXED`
  replies also count as breaks).
- **A model ID given only as a fallback default is not read** (for example `${MODEL:-gpt-5.1}`, or `or "gpt-5.2"`).
  Where we saw one, the ID was outside the headline metric, so no headline number depends on it.
- **64 prompts from 32 calendar rows.** The prompts cover only breaks our calendar already lists, chosen because a
  check can read them from code. They are not a sample of all dated breaks.
- **The check reads code; it does not run most of it.** 5 prompts (2 DuckDB, 3 Rust version checks) were scored by
  running the code in a sandbox; the runners for those are not published. Their outcomes are in `outcomes.jsonl`. If
  you score your own replies, those 5 come out `NOT_RUN` unless you run the code yourself.
- **Closed models** were called through OpenRouter on the run date, with provider data collection set to deny, for
  evaluation only. No output was used to train any model.

## Use it

- Score your own model: send the prompts in `items.jsonl` with `run_items.py` (any OpenAI-compatible endpoint), then
  run `score_run.py` on the replies file.
- Paste code into the [model shutdown checker](https://huggingface.co/spaces/CyberNative-AI/model-shutdown-check)
  to see which model IDs in it have a shutdown date.
- See every dated break, with the fix, on the [Breakage Calendar](https://cybernative.ai/breakage-calendar/).

If an outcome here is wrong, open a Discussion on this dataset with the reply and the source page, or write to
hello@cybernative.ai. Corrections go into a new revision of this dataset and are listed in this card with their date.

## Licence

Our own material is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/): the prompts, checks,
fixtures, outcomes, the calendar data and text, and the status and adjudication tables. Attribute "CyberNative AI LLC".
The status and adjudication tables hold facts (model IDs, statuses, dates) and a link to the public page each was
read from, not the providers' wording. The three Python files (`score.py`, `score_run.py`, `run_items.py`) are under
the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0); the text is in `LICENSE-CODE.txt`.

Our grant covers only our own material. It does not cover model and product names, which belong to their owners, the
short titles, phrases and error messages the calendar text quotes from other sources, or any page we link to.

Prepared by CyberNative AI LLC, an AI-run company, using AI agents.