Live examples

Watch a swarm work, in each profile

Each example is a real HypSwarm run on a real record: seats pick up work orders from the record, write results and claims through it, get refused when they overclaim, and are reviewed before anything counts. Open one and it replays from the first event — model and tool calls in flight, the graph growing, the board settling. Then explore its pipeline, graph, evidence and report.

Scientific · Facts from the record

Does A beat B

Two analysts and a critic test an evaluation claim. The instrument re-runs every value before any review, and the report is checked against the record.

seats: analyst, stats, critic, lead
run: 14 model calls · 14 tool calls · 12 s · settled
record: 80 events · 29 nodes · no refusals

Watch it run

Software · Change report

Importer keeps quoted commas

The tests count only when the instrument runs them. A tester's “passes” written before that run is refused, reworked and accepted once the lab run lands.

seats: implementer, tester, reviewer, lead
run: 16 model calls · 14 tool calls · 16 s · settled
record: 81 events · 28 nodes · 1 write refused, then rewritten

Watch it run

Trading · Desk note

Mean reversion after a 3 % five-day drop

A backtest on a held-out window, with fees and slippage. A result without costs is refused before it can become a thesis.

seats: quant, risk, lead
run: 14 model calls · 13 tool calls · 16 s · settled
record: 56 events · 21 nodes · 1 write refused, then rewritten

BACKTEST_WITHOUT_COSTS

Watch it run

Knowledge · Sourced brief

Three facts, sourced

Three facts, each a source node and a claim that cites it. A claim with nothing behind it is refused.

seats: researcher, fact_checker, lead
run: 26 model calls · 23 tool calls · 27 s · settled
record: 84 events · 29 nodes · 3 writes refused, then rewritten

EVIDENCE_MISSING

Watch it run

Optimization · Solution report

Assign six customers to three depots at minimum cost

Objective, feasibility and a gap. An objective with no feasibility check is refused, and “optimal” needs a bound that meets it.

seats: modeller, solver, verifier, lead
run: 14 model calls · 13 tool calls · 19 s · settled
record: 59 events · 21 nodes · 1 write refused, then rewritten

OBJECTIVE_WITHOUT_FEASIBILITY

Watch it run

The runs use HypSwarm's scripted models, so they cost nothing and replay the same way every time; run any of them with a real model with python run.py --model <id>. Rebuild these pages with dev/site_examples.py. Each bundle verifies on its own: verify one.