EgoDeadlines: Invented Deadlines in First-Person Captures

Someone says, within range of the microphone they wear: "I wrote Oct 15 for the residency application as a guess, I actually have no idea when it's due." A personal assistant that sets a phone alarm for October 15 has invented a deadline. After a few alarms like that, people stop trusting the ones that matter.

This dataset measures how often a local language model makes that mistake when it reads short first-person captures: voice memos, conversations picked up by a wearable microphone, and forwarded emails. It comes with the paper How Often a Local Language Model Invents Deadlines When Reading First-Person Captures (Kho, 2026), which explains everything below in full.

The short answer

Gemma 4 31B, run with the exact settings of the personal assistant that motivated this study (reasoning off, answer held to a fixed JSON shape), invented a deadline on 2/60 captures that had none. Both times the speaker had called the date a guess. Switching reasoning on, or adding one paragraph to the instruction about what a firm deadline is, removed every invented deadline. There were too few of them to test either fix.

Two errors turned out to matter more:

1. An undefined field. The plain instruction asks whether to set an alarm but never says when one is due. The 31B never set one (0/80 real deadlines). The 12B set one for 46/80.
2. "Anywhere on Earth" time. On the 5 deadlines given in AoE, the 12B put the deadline in another zone every time with reasoning off (0/5 right). Every wrong zone moves the deadline earlier.

What is here

File	What it holds
items.jsonl	140 captures, each with the moment it was recorded and the correct answer
results/gemma4_31b.jsonl, results/gemma4_12b.jsonl	One row per capture per setup: the model's answer, its reasoning trace when reasoning was on, token counts and time taken
scores.json	Every rate, with its 95% Wilson interval, and the paired tests
figure.png	The central figure of the paper
make_items.py, run_model.py, score.py, figure.py	The code that produced all of the above
results_openrouter_partial.jsonl	65 answers from an abandoned first run on a hosted model; not analysed

There are seven kinds of capture, 20 of each. In three there is no firm deadline (a date marked as a placeholder or a guess, a deadline that waits on an event that has not happened, no deadline at all). In four there is one (a date stated outright, a relative date, an event date next to the real deadline, a clock time with a timezone). Every capture is generated from a template with a fixed random seed (20261102), so every correct answer is computed by a script. The captures are synthetic and contain no personal data.

Setups

Each model answered every capture five times: a plain or a guarded instruction, with reasoning off or on, plus an identical repeat of the deployed setup (plain, reasoning off) to show how much answers move by chance. They barely moved: the 31B gave the same decision and date on 100% of captures in both runs, and the 12B on 99%.

The models ran locally through Ollama: gemma4:31b (4-bit, on a 128 GB Apple M5 Max) and gemma4:12b (4-bit, on a 48 GB Apple M5 Max).

Limits

The 60 captures without a deadline come from 13 phrasings, and every invented deadline came from one of two, so the captures are not independent and the intervals are too narrow. Clean synthetic text is easier than real transcripts, so these rates probably understate what a model would do on real captures. Most setups scored near perfect, which means the set is too easy to rank them; a harder version is the next step.

Citation

Kho, C. V. (2026). How Often a Local Language Model Invents Deadlines When Reading First-Person Captures. Preprint.