Research
Davinci
ENTON's verifier-driven research model program, and EntonBench, the benchmark it is graded on.
Overview
The platform executes; Davinci learns.
Every fill, reject, journal row and account-state decision the platform makes is journaled, replayed and graded, then turned into verifier-graded benchmark tasks (EntonBench). Those tasks train a model to make stricter research calls and to stop on weak ideas. The model proposes; deterministic verifiers decide. A model never grades itself.
From a decision to a task
Three cadences, per fill, per session and nightly, turn market and broker data into graded tasks and reward labels.
- Ingest
Broker fills, order state, decision journals, the tick archive and account rules become an append-only ledger.
- Replay
The vectorized factory reconstructs the session. Every live trade is paired with its simulated counterpart.
- Attribute
Gaps in P&L are labeled: alpha, entry slip, exit slip, missed signal, sizing, gate mismatch or rule drift.
- Train
Verified outcomes become EntonBench tasks: templates, gates, diagnostics and calibration patches.
Verification ladder
Cheap checks first, expensive truth last. A candidate climbs through deterministic verifiers before out-of-sample, hidden holdout, live shadow or live deployment can earn it reward.
| Step | Check | What it verifies |
|---|---|---|
| 1 | Contract | TypeScript compile, template interface, immutable evaluator. |
| 2 | Feature | No-lookahead: cached features equal what was knowable at signal time. |
| 3 | Factory | Phase 1 sweep, phase 2 out-of-sample, slippage, robustness and duplicate filters. |
| 4 | Replay | The live journal and the factory replay produce the same gate path. |
| 5 | Shadow | Broker fills, rejects and account-state decisions reconcile. |
| 6 | Live | A small-size deployment earns reward only after attributed realized P&L. |
Where things stand
- No strategy is trading live. Live execution is paused until a candidate earns it. The rails stay in place for that step: a mechanism-score gate, a daily-loss governor and an account-level kill switch, with every decision journaled.
- Shadow scoring runs all session. A journal-only scorer evaluates every candidate on both sides, independent of any live module, so evidence accrues even while trading is locked out.
- New templates stay research-only. A candidate needs no-lookahead construction, a factory-native implementation and walk-forward validation before any capital is involved.
Rejected branches
These strategy families failed honest out-of-sample testing and are kept as negative controls:
- Opening-range breakout continuation: 0 out-of-sample survivors.
- Opening-drive pullback continuation: 0 out-of-sample-positive rows.
- Trend-expansion continuation: 0 out-of-sample survivors.
- PM trend continuation: weak or unstable survivors only.
- Short-only factory fade retune: 0 out-of-sample survivors.
The model is trained to stop on weak families, not to run another grid search on branches that have already failed.
Reward
A candidate's reward combines three things, each from a verifier rather than a model's opinion:
reward = replay parity · honest out-of-sample · attributed live P&LEntonBench
EntonBench is the benchmark these tasks form. A model passes a task only when the artifact it produces survives deterministic out-of-sample verification against committed reference data. Rankings use only hidden holdout tasks.
Results, with the run date and access route of every row, are on the EntonBench page. The hidden set is small, so differences between models there are not conclusive.