# cognition: method

> how two ai agents researched the cognition history, the evidence rules they followed, and what they could not verify.

## the lanes

the corpus was produced by two agent research lanes — history-people (the company, the founders, the funding arc) and coverage-idea (the discourse, the debunks, the agent-era analysis) — each writing a lane report against one shared scratch source catalog before synthesis. the pattern is the company-research formula proven on roam-research and the earlier dossiers in this series: lane-prefixed source ids frozen upfront, per-claim confidence from the start, and a declared deduplication owner.

## the evidence rules

- every event, quote, and claim carries at least one source id.
- dates keep their precision; approximate dates are labeled, not sharpened.
- confirmed, reported, and inferred stay distinct.
- sources are typed by role: primary, interview, reporting, archive, community.
- sentiment describes the coverage, not the truth of it.

## known gaps

- the founding month is genuinely disputed — august 2023 versus november 2023 versus incorporation around january 2024; the discrepancy is preserved, not reconciled.
- scott wu's birth year splits 1996 (his own telling) versus 1997 (wikipedia); the sources disagree and both are kept.
- every load-bearing metric is company-published — arr and run-rate figures, the 89–95% internal-code claims, swe-1.x performance, the guarantee's estimator — and labeled self-reported wherever it appears.
- no formal cognition rebuttal to the internet of bugs debunk is documented; the company answered with the eval harness and shipped product, not a response.
- steven hao's biography rests on linkedin and speaker bios; walden yan's thiel fellowship and anysphere stint come from community directories, not primary sources.
- windsurf's employee counts, layoff numbers, and the ~$250m deal-price estimate come from press reports and secondary aggregators — no filing exists.
- whether graham moreno transitioned to cognition with the windsurf team is unconfirmed; interim-role coverage stops at the deal.
- the $21m round is labeled both seed and series a across sources; the label question never resolves.
- headcount ranges from ~200 (wikipedia, late 2025) to 300–400 (linkedin, 2026); no canonical figure exists.
- the early-access corpus is partial — much of the waitlist-era record lives in unarchived tweets and deleted posts.

## corrections

corrections and additional primary sources are welcome — cite the claim, the source, and what it changes.

182 cataloged sources. 39 timeline events. 24 writings. 29 coverage items.

AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.

Canonical page: https://hraness.com/cognition/method


Published by @ben in profile revision 1.
Original document: https://hraness.com/cognition/method
Original credited author: Hraness
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.

## Structured source and record data

````json
{"anchors":[{"id":"lanes","originalUrl":"https://hraness.com/cognition/method#lanes","title":"the lanes"},{"id":"rules","originalUrl":"https://hraness.com/cognition/method#rules","title":"the evidence rules"},{"id":"gaps","originalUrl":"https://hraness.com/cognition/method#gaps","title":"known gaps"},{"id":"corrections","originalUrl":"https://hraness.com/cognition/method#corrections","title":"corrections"}],"provenance":{"draftingDisclosure":"AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.","originalAuthor":"Hraness","originalRevision":"d474932300ec1faae89ef601cdd9ae5fd33a0d82485b3119a4d111a105647808","originalUrl":"https://hraness.com/cognition/method","source":"hraness"},"records":[],"sources":[]}
````