Meta

This site is built and run by a swarm of agents

Not a metaphor — : below is the flight log of the origin research, the numbers of the current build, and the verdict receipt the portal issued to itself.

The origin research: one night of the swarm

The portal's origin is the AI 2026: 15 web-research topics, an adversarial fact-checker behind every dossier, 12 parallel writers, and a panel of 8 simulated experts. The core phases completed in roughly 40 minutes; in total the swarm processed about 4 million across 1,060+ .

  • ~40 minutes of core phases
  • peak: 10 researchers in parallel
  • ~4M tokens
  • 1,060+ tool calls
  • 34 of 147 claims corrected by fact-checkers

— timeline.json, the flight log of the July 12, 2026 run. Each bar is one ; the phases deliberately overlap: verification starts as dossiers become ready.

This site build

live telemetry arrives after launch

28 agents

repository context audit

8 agents

product recon

~3.5M

tokens for research and audit

Claude Fable 5

orchestrator — plus one human

The numbers are honest data from the current build session, recorded by the .

Quality mechanisms

The 's rests not on trusting an 's self-report but on three mechanisms: a panel settles the contested calls, a scanner checks the styling, and the verifies every PR in person.

The archetype panel as a method

Contested decisions go to a simulated editorial panel of six archetypes with a critic and a buyer's advocate. That is how the F.1 ↔ F/ONE brand system was settled — glyph and wordmark separated, the F/1 hybrid removed, the F1 key honestly opening help — and how the C-Suite intro layer was shaped: an answer-first standfirst plus a "how to read" block.

Semantic visual lint

npm run audit:visual — a scanner that catches a mismatch between styling and role: prose wearing a badge's chrome, affordance without interaction, raw hex outside the . The heuristics are deliberately conservative; the triage regulation is docs/visual-qa-loop.md.

Personal verification as the gate

The checks every PR in person: own test runs, a live browser, an isolated e2e; merge conflicts are modelled in advance. In a single night shift this closed a registry of 30 issues across six PRs, and the test suite reached ~975 unit and 33 end-to-end.

Genesis Receipt

Before building the portal, we ran it through our own engine — . Here is the receipt.

F/ONE VERDICT

genesis receipt · check #001 — this portal

Task profile

site build
shared context, shared files
origin research
15 independent topics

Verdict

site build
single-with-subagents
research
workflow (swarm)

Rationale

reads
in parallel
writes
single-threaded
extra agents
into verification

TOTAL: the method applied to itself

fedoseev.one/verdict

Swarm or single agent — for your task?

9 questions — and you know the verdict and the cost of being wrong.

Run your own task