Designed to resist self-deception.
The methodology treats data access, candidate selection and deployment authority as separate irreversible gates. A result is not valid because it looks good; it is valid only if it survives the preregistered process.
The research contract.
01 · Causality
Observations may contain only information economically available at that time. Higher-timeframe information is aligned only after availability; action is followed by next-bar execution semantics.
02 · Frozen design
Treatments, thresholds, candidates, windows, eligibility and ranking rules are fixed before Fresh performance reveal. Post-hoc parameter repair closes the lineage rather than rewriting it.
03 · Economic realism
Fees, half-spread, slippage, historical funding, mark-to-market accounting, leverage, sizing and risk ordering remain part of the environment rather than being removed from headline performance.
04 · Failure visibility
Ineligible candidates and closed studies remain in the record. “Least negative”, “lowest turnover” or “safest” is not promoted when frozen eligibility fails.
TRAIN ≠ Fresh ≠ TEST ≠ deployment.
Each evidence class answers a different question and has a different authorization boundary.
Model formation
Training and checkpoint creation only. TRAIN performance is not used to choose among the ten frozen checkpoints.
Selection evidence
All 360 episodes must complete before performance reveal. The complete 120,960-row source is required first.
One-champion confirmation
Locked unless Fresh yields exactly one eligible champion and a separate future TEST authorization chain is frozen.
Separate protocol
Paper, sandbox or live deployment is not implied by research success. Live trading remains unauthorized.
Partial results have zero selection authority.
Partial performance printing, arm comparison, early stopping because an arm looks good or bad, treatment adaptation, favorable-window retries and selection from a failed partial run.
Every checkpoint SHA256 reverified, model spaces checked, arm threshold verified, candidate/window identity checked and deterministic predictions enforced.
Four conditions. All must pass.
No descriptive fallback.
Within each arm
Eligible candidates only: maximize median economic return → exact tie maximize mean economic return → exact tie minimize mean maximum drawdown → exact tie = AMBIGUOUS, no winner.
Across arm winners
Only eligible unambiguous arm winners enter study-level ranking using the same order. If none exist: NO_ELIGIBLE_ENTRY_CONFIRMATION. CONTROL is allowed to win.
Explain results, never override rules.
| Category | Examples recorded per Fresh episode |
|---|---|
| Economics | Net economic return, fees, spread cost, slippage cost, funding cashflow |
| Risk | Maximum drawdown, risk termination, exact termination reason |
| Activity | Turnover, turnover ratio, legs, entries, exits, reversals, exposure minutes |
| Policy | Requested FLAT/LONG/SHORT and effective action accounting |
| Treatment | Suppressed requests, confirmation completions, pending replacements and resets |
| Accounting | Reward-equity sum, TRAIN-only turnover regularizer, macro shaped reward, service and agent minutes |