Every thesis gets one honest chance.
The scheduler funds the cheapest registered falsifier for every thesis before any survivor receives follow-up capital.
Research portfolio command
Capital should follow disproof.
FLOTILLA gives every thesis the cheapest experiment that could kill it, runs the fair floor first, and concentrates a fixed budget only after evidence earns another round.
5registered theses
12budget-unit mandate
8completed runs
43append-only events
The portfolio problem
Research teams optimize experiments. Few optimize what deserves an experiment.
A leaderboard cannot tell a research director when to stop, what budget was returned, or whether yesterday’s kill should be challenged. FLOTILLA treats those choices as first-class, reviewable artifacts.
Models propose. Predicates dispose.
Kill conditions are allowlisted expressions evaluated against recorded scores. Missing evidence becomes UNDETERMINED, never a guess.
A stop is a decision, not deletion.
Human confirmation is mandatory. REVIVE appends a challenge while preserving the original evidence, confirmer, and lineage.
Bundled Journey 0
One deterministic fixture. The complete portfolio loop.
These are mechanism checks from five paired-measurement theses, not
scientific performance claims. Every number is regenerated by
make demo.
Fair floor completed
All five one-unit falsifiers execute before the first two-unit follow-up.
T-01 and T-03 stop
Each predicate fires, receives operator confirmation, and emits an argued Markdown kill report.
Undetermined in fixture
The default data is complete. The simulator’s thin-evidence scenario exercises the abstention path.
Scope: local standard-library execution, temporary browser simulation, SQLite lineage, and static GitHub Pages. See LIMITS.md for unmeasured launch claims.
Capital visualization
Broad at first contact. Concentrated after survival.
Initial thesis caps do not reserve the whole treasury. Actual spend, earmarks, releases, and reserve remain distinct in the ledger.
Registered allocation sequence
Falsifiers5u
Follow-ups6u
Reserve1u
“The cheapest experiment that could change the decision goes first.”
Open the simulator to change the mandate, earmark reserve, introduce replication headwinds, confirm or overturn kills, and revive a stopped thesis without erasing its history.
Explore capital flowSystem architecture
A decision engine with evidence at every seam.
The static product site is only the visible chart. Under it, the Python implementation enforces plan approval, cost constraints, deterministic predicates, and append-only provenance.
Thesis registry
Prediction, budget cap, deadline, limitations, and safe kill predicate.
Approved plan
Versioned DAG with falsifier, ablation, confound, replicate, and aggregate nodes.
Budget scheduler
Fair-floor dispatch, shared reserve, and explicit local or notebook execution.
Decision gate
Deterministic predicate, UNDETERMINED path, and required human confirmation.
Lineage ledger
SQLite events, runs, scores, code/data identity, kill reports, and reversible revival.
Operating doctrine
Designed for research leaders, legible to reviewers.
FLOTILLA is not a new experiment tracker or autonomous lab. It is the governance layer that makes allocation decisions explicit, executable, contestable, and reproducible.
See the whole mandate.
Compare thesis trajectories and actual capital flow before moving reserve toward surviving work.
Register what would change your mind.
Pair every claim with a cheap falsifier, a deadline, a budget cap, and limitations before dispatch.
Audit or challenge the stop.
Inspect scores, data identity, code commit, confirmer, and the exact predicate before confirming, overturning, or reviving.
Journey 0 · keyless · temporary data