This dataset is published for research and evaluation purposes only. Its contents were generated with Anthropic Claude models, so it must not be used to train machine-learning models. By requesting access you agree to use this data for research and evaluation only, and not for model training.

SKPropagate

What this is. A correction-propagation environment built from the record of Seek, an autonomous research agent. Seek keeps thousands of small linked notes. When an audit corrects one in place, every note written on top of it may still carry the old value. Since 2026-09-11 a nightly checker reads each corrected note with each note that links to it and rules: stale (the dependent states the old value as fact, with the line quoted) or clean. A repair pass then fixes the stale ones. Here your agent sits in the checker's chair.

Why it matters. "The knowledge base contains the corrected fact" and "downstream artifacts still behave as though the old fact were true" are different tests. This is the second one, and it is what single-shot research evaluation cannot see. It is the sequel to seekbot/seek-propagation, which measured motion; this measures judgment.

Structure. 447 tasks: 149 stale, 298 clean (2 clean per stale). Per task: instruction.md, environment/ (corrected.md, dependent.md, pair.json), tests/ (gold and grader, verifier-side only), solution/solve.sh (oracle). No network, and no external sources: both files are vault text. Dependents by folder: {'40-entities': 166, '70-drafts': 14, '30-notes': 162, '50-questions': 74, '40-mocs': 24, '80-published': 7}. 100 distinct corrections.

The ruling. stale with the line at fault, or clean. The instruction states the checker's own rule: stale only when the dependent's own text asserts the old value as current fact; clean when it does not carry the detail, already carries the new value, marks the old one as historical, or does not commit.

Scoring. 1 if the verdict matches the checker's, else 0; the headline is the mean of the two class means. The quoted line is compared with the checker's and reported as line_hit, not scored.

Baselines.

strategy	score
always `clean`	0.500
always `stale`	0.500
`stale` if the old wording quoted in the record recurs in the dependent	0.539
oracle	1.000

Beat the string rule before claiming your agent can track a correction.

Strata. Every task carries correction_kind (in tasks_index.jsonl, task.toml and the gold; not in the files the agent sees), so a score can be broken down by what kind of detail was corrected:

kind	what changed	stale	clean
`description`	what the thing is or says: a wrong term, characterisation, mechanism or scope	41	83
`attribution`	who said, wrote, did or signed it; which document or outlet carries it	40	84
`strength`	how strong or how supported: independent, first, only, settled, confirmed	35	58
`number`	a figure, count, date, interval, page or identifier	24	49
`quotation`	wording given as the source's own that the source does not carry as given	9	24

For a stale task the kind is that of the detail the dependent still carries. A clean task inherits the commonest kind among the stale dependents of the same correction. The kinds were assigned by a review pass (claude-fable-5-1, 2026-10-05) reading each correction record with the checker's reason; a correction that fixed several things can sit in more than one kind across its dependents.

Where the gold comes from. 1532 verdicts by Seek's checker (claude-opus-5, one prompt) between 2026-09-11 and 2026-09-29: {'UNCLEAR': 48, 'CLEAN': 1334, 'STALE': 150}. UNCLEAR is left out. Of the 149 stale tasks, 135 had been acted on by the nightly repair pass at export time (confirmed_by_repair in the gold) and 3 had been skipped by it (skipped_by_repair). In version 0.1.1 every skip says the stale wording had already been changed by another pass before the repair ran; none says the verdict was wrong. Repair reports are matched to verdicts by correction and dependent, or by dependent alone where the corrected note was renamed in between (repair_match in the answer key). Clean tasks are sampled by hash, preferring dependents of the same corrections as the stale ones (at most 6 per correction), so most negatives are hard: same correction, different dependent. Skipped: {'dependent already carries repair markers': 18}.

How the pair is frozen. Both files are read from the vault's git history at the last commit before 05:00 on the judgment date, which is when the checker ran, so nothing the repair pass wrote afterwards is shown. Stale evidence lines were required to be present in that text. Lines in the corrected note that mention the repair organ are dropped. Frontmatter is cut to descriptive fields plus audit_status, which is where the correction record lives.

Limits, honestly.

The checker is a model. Agreement means you judge like it. The repair pass is a second model reading the same texts, not an independent ground truth. It has acted on 135 of the 149 stale cases.
Clean verdicts have no second signal yet.
The corrected note is shown with its full correction record, so the agent knows what changed. The task is reading the dependent, not detecting the correction.
Some dependents are question notes and entity hubs with dated logs; the rule that a dated historical entry is not stale is stated in the instruction because the checker applies it.
Single run of a single prompt produced every label; the prompt is reproduced as the instruction's standard.
A spot check before release (claude-fable-5-1, 2026-10-05) read 9 stale and 8 clean tasks against their correction records. None looked wrong outright. Three or four of the stale ones are strict calls: a framing word that survived ("self-treatment", "lineage") where a looser reader would pass the note. Expect some disagreement of that kind to be about strictness, not error.
No agent has been run on this environment yet. The only numbers are the baselines above and an oracle pass in Harbor (447 of 447).

Terms. Research and evaluation ONLY. Notes and verdicts were written by Claude-family models: do not use this environment or its contents to train models. Same gate as every seekbot dataset.

Generated mechanically by seek_export_propagation_env.py (seek-code) from the checker's verdict files, the repair reports, and the vault's git history. No model runs in the exporter.

Downloads last month: 20

Total file size:

11.5 MB