seekbot/seek-propagation

Correction propagation v0 — does the fix travel?

Context first: what system is this from?

Seek is an autonomous research agent that has run nightly since June 2026 on a Mac mini. She follows her own curiosity across the open web and writes what she finds into a knowledge base (a "vault") of small Markdown notes: claim notes (one falsifiable claim each, with a source quote and DOI/URL), observations, entity pages (hubs about a person or thing), questions, and essays. Notes reference each other with [[wikilinks]], so the vault is a linked graph. Crucially, her output becomes her future input: tomorrow's research agenda is generated from today's knowledge state. The whole vault is public and browsable at talk-about.ai; her decision records are published in seektraces-decisions.

Seek's claims are audited by a second model (a "cross-model audit"), and sometimes an audit finds a claim wrong or overstated and corrects it, leaving a dated correction narrative in the note. That raises the question this dataset measures — posed by an external reviewer of Seek's public traces:

When a long-running agent corrects a claim, how far and how fast does the correction propagate through everything built on top of it? "The knowledge base contains the corrected fact" and "downstream artifacts still behave as though the old fact were true" are different tests.

For intuition: a newspaper prints a page-one correction, but last month's op-eds citing the wrong number are still on coffee tables — and tomorrow's columnist might quote the op-ed. For a system that writes tomorrow's research from its own back-issues, this is the core maintenance problem — and it is what ordinary single-shot "deep research" evaluation cannot see.

A worked example (real records from this dataset)

On 2026-07-11, an audit corrected the claim claim-forsythe-coined-computer-science-and-founded-stanford-cs (the popular attribution is a myth — per Knuth, Forsythe did not coin the term). This dataset then asks: what happened to every note that links to the corrected one?

dependent artifact	what happened	class
claim-purdue-established-first-us-cs-department-1962	modified 1 day after the correction	touched-after, lag_days 1
observation-numerical-analysts-founded-first-us-cs-departments	modified 2 days after	touched-after, lag_days 2
claim-forsythe-did-not-coin-computer-science-per-knuth-1972	created 2026-08-07 — the corrected knowledge later crystallized into its own claim	created-after-correction
entity-george-forsythe	entity hub created after the correction	created-after-correction

Whether each touch actually incorporated the correction (vs. edited for unrelated reasons) is the human-judgment layer — see "the open task."

The data

25 corrections (dated audit-correction records, June–July 2026) joined to their dependent artifacts via the vault's wikilink graph (breadth-first, depth ≤ 3, across claims, observations, maps, entities, questions, drafts, essays) = 2,122 (correction, dependent) records. Dates come from the vault's git history. Fields per record:

• correction, correction_date, correction_type — the repaired claim
• dependent — a note that links (directly or via ≤3 hops) to it
• depth — link distance (1 = links directly to the corrected note)
• created, last_modified, first_touch_after_correction, lag_days
• class — mechanical status:
  • touched-after: dependent existed before the correction and was modified afterward (fast lag suggests, but does not prove, repair)
  • untouched-since: existed before, never modified after — the prime stale candidates
  • created-after-correction: dependent didn't exist yet — new work built atop a corrected node; it inherited either the fix or the ghost (stale-action exposure candidates)
• judgment_needed: always true — see below
• dependency_proxy: how "dependent" was operationalized

Mechanical baseline: 453 touched-after (median lag 1–3 days, reaching depth 3), 45 untouched-since, 1,624 created-after (the corrections are early-era and the vault tripled around them afterward).

The open task (why judgment_needed is on every record)

The mechanical pass measures motion, not correctness. Distinguishing updated / intentionally historical / stale / not applicable requires reading the dependent against the correction — every source text needed is public (the vault, plus full audit narratives in seektraces-decisions). That makes these 2,122 records a natural benchmark for the capability the reviewer named: can a model (or a human, or the agent herself) correctly classify downstream state after an upstream repair? Beat the baseline; publish your labels.

Honest limits

Wikilink linkage is a proxy for semantic dependency (expect not-applicable cases). Depth-3 closure over a dense graph inflates counts (filter depth == 1 for the strict view). The system evolved while running — treat eras carefully; these corrections predate several machinery changes. Analysis is model-free (string/graph/git mechanics).

Provenance and permitted use

Vault text was written by Claude-family models inside the Seek scaffolding; per Anthropic's Consumer Terms this dataset is for research, evaluation, and reading — not training competing models. Siblings: skbench · seek-trajectories · seek-sessions · seektraces · seektraces-decisions.