The thesis

A research lab’s method, run with a quant fund’s discipline.

The six platforms exist to make that sentence literally true in software — and to prove it when it isn’t.

Why this needs a machine

Every backtest lies. The architecture exists to catch the lie.

Any one invalidates a result. Most builders stop before checking. This pipeline is built to stop them automatically — and to keep checking after deployment.

The ecosystem

Six Greek-named platforms, one job each.

The lifecycle

An idea’s journey from guess to — maybe — capital.

The through-line

Determinism, from first guess to live order.

This is the spine the harsh questions land on — and the reason a result here means what it says.

Discover εὕρεσις

HEURESIS — the idea engine.

Discovery is industrialised. A grid enumerates a strategy space exhaustively, deterministically and cheapest-first; pre-registered batch screens test a named hypothesis whose registration is frozen — hashed, with the code commit stamped into the artifact — before the run starts. A YAML template sends both down a config-driven path, and a local pre-flight gate rejects weak ideas before they cost a full validation.

Admission runs on staged inference: screen results can disclose but may never bank a claim; only confirmation on data the screen never touched counts. A clean kill is throughput.

Prove αὐτόματα

AUTOMATA — the experiment layer.

Between “a research question” and “a certified spec” sat a gap nothing owned. AUTOMATA is that layer, and its product is not strategies — it is a reproducible evidence graph: immutable experiments, an honest account of how hard the search was, typed provenance on every number, and negative results that stay searchable.

It is a second manufacturer running alongside HEURESIS, not replacing it — one optimised for rate, one for provenance, so the comparison is measurable rather than argued. It cannot certify its own work: a local ladder run is preflight, never green.

Validate ἐπιστήμη

EPISTEME — the gate.

The heart of the system, and the only thing in the ecosystem that can certify. A candidate climbs eleven canonical rungs — data quality, causality, costs, signal selection, walk-forward, significance, path risk, generalization, adversarial stress, Monte Carlo, deployment — plus sub-rungs that bring a default green run to sixteen scored records. Fail a rung and the run halts. Thresholds are committed in advance and cannot be tuned to fit a result.

Time καιρός · optional

KAIROS — knowing when, when it helps.

An edge often lives in one market regime and dies in another. KAIROS is a regime-classifier factory — it builds the classifiers, climbs them up its own thirty-rung ladder, and lends out only the labels that survive. A strategy may opt in to a blessed regime, or ignore regimes entirely; if a spec declares no regime inputs, the whole regime rung is skipped.

It also wears a second hat: KAIROS owns the ecosystem’s measurement instruments — minimum detectable effect, null models, effective sample size. The question “could this sample even detect the effect you are claiming?” has an owner, and it is not the person making the claim.

Execute πρᾶξις

PRAXIS — the disciplined runtime.

PRAXIS runs strategies that cleared the gate. It re-checks each for determinism boot over boot, enforces portfolio and per-strategy risk limits, and isolates every strategy so one failure can never take down the rest. A self-declared green is refused by default; running anything ungraded takes a written, recorded election with a stated reason.

The core runtime has never routed an order to a production exchange endpoint. Nine books run today: eight on a simulated executor, one on a venue sandbox with virtual funds. Reaching real capital is an explicit operator action, never automatic.

Coordinate ἀγορά

AGORA — the commons that owns nothing.

The platforms never call each other. They meet in AGORA — a git repository where candidates, results, validated bundles and a permanent event journal live. AGORA ships no importable code at all: it is files, schemas and history. It is the single place that can answer “what ran, when, and why?”

Where platforms must share a definition, they share a published artifact, not a source tree — canonical metric libraries cross as versioned, hash-pinned wheels published here. An editable install pointing into a sibling’s checkout is exactly the coupling the contract forbids.

The power of the IP

Five months of build, on three years of research.

Plus 6,000+ commits, 5,000+ cross-team hand-off records, and a data universe grown to 50+ instruments across nine asset classes. The six platforms date from April 2026 and rest on three years of prior research. The discipline isn’t a slide — it’s in the repository history, and every figure here was measured from it rather than quoted from a README.

The part most labs skip

The rarest thing we build is a control honest enough to fail.

The same rejection aimed at strategies is aimed at the validators themselves. This is the discipline’s sharpest edge — and the reason to trust the green ones.

“Everything measured was green; everything unmeasured was red. Principles do not hold a line; runnable commands do.”— the ecosystem charter, written after an audit proved the point

• We hunt controls that can’t fail. A gate that passes vacuously is treated as a defect equal to a wrong answer. Every tool must ship a foil — a deliberately broken variant it is proven to catch — or it does not count as a control.
• Unrun is its own state. A check that has never executed does not render as green; it renders as unknown, which is treated as worse than red.
• We retract our own claims. When an internal “moat” claim didn’t survive scrutiny, it was withdrawn on the record rather than defended. Retractions are filed the same day, to the same audience.
• A clean kill is a success. A candidate rejected under pre-registration is throughput, not failure — the archive of what didn’t work is the map of where the edge isn’t.
• Almost nothing clears the bar — by design. The gate is deliberately, rarely met; the system is built to be sceptical of its own output first.

The machine earning its keep

Rigor you can watch work — each caught before any capital saw it.

• Correlation control → a dozen “edges,” one bet. A batch that looked like many independent strategies was shown to be barely more than a single effective bet — the over-diversification trap, caught before it could concentrate risk.
• Generalization floor → strong on one asset, dead on three. A strategy that shone on a single instrument was rejected the moment it had to hold across three.
• Adversarial battery → noise, unmasked. Reverse a strategy’s own signal and shuffle its labels; if “edge” survives the sabotage, it was never edge. Many didn’t.
• A ladder that an RNG could climb. The regime platform tested its own validation ladder against a model that ignores its inputs — and the model passed. It rebuilt the ladder and re-scored every prior result against the stricter bar, rather than keeping the flattering ones.
• A capacity formula that disagreed with itself by three orders of magnitude. Documentation said one thing, the code computed another. The gap fed a flag that blocks promotion. Found and filed before it decided anything.
• An admission gate that could not see its own newest member. The check reported “PASS: 5 siblings” and was structurally incapable of noticing a sixth — because it verified one list against the other in only one direction. It was caught, reported and owned by the very team it exonerated.

None of these were found by an outside auditor. Each was found by the machine, or by the team turning the machine on itself — and written down where it can be checked.

How they compose

Six independent platforms, no shared source tree. That is deliberate.

• No platform can corrupt another. Coordination is files and git history — so a bug in discovery can never reach into execution.
• Shared definitions cross as artifacts, not paths. Where two platforms must agree on what a metric means, they install the same hash-pinned wheel published to the commons. A definition that is hard to import gets reimplemented, and reimplementation is how two teams end up with two answers.
• Every claim is traceable. A live fill points back to a bundle, to a scorecard, to the candidate and the search that produced it.
• Each evolves on its own clock. Independently built, tested, versioned and deployed — coordinated by contract, not by coupling.
• A sixth was admitted in six days. The contract for joining is written down and was exercised in the open this month — including the admission gate discovering, and reporting, that it could not see the platform it had just admitted.

The wedge

No black boxes. No vibes. The discipline is the product.

The market ships autonomous “AI agents” trading capital nobody should have trusted them with. Isard is the deliberate opposite — and the difference is structural, not aspirational.

ML earns its place — never as the trader. It proposes; the rules decide. If the trade can’t be explained, it isn’t taken.