EXP-0012 · CONTROL
Immediate admission
N=1
A single LONG or SHORT request is admitted immediately under the frozen parent semantics.
CONTROL10 candidates
systematic RL research · phase 5B
ZeroAI Research Lab
A preregistered trading-research program built to make causality, frozen lineage, execution economics and out-of-sample discipline visible — before any performance claim is allowed.
Historical TEST locked Live trading disabled Candidate performance sealed Verified 18 Sep 2026
FROZEN30 × 12360 EPISODES
EXP-0012CONTROL · N=1
EXP-0013TREATMENT · N=2
EXP-0014TREATMENT · N=3
OBSERVATION474 × float32
Current study
ENTRY
ENTRY-CONFIRMATION-STUDY-0001
Experimental arms
3
N = 1 · 2 · 3
Frozen candidates
30
10 checkpoints per arm
Fresh matrix
360
30 candidates × 12 windows
project status
Prepared. Frozen. Waiting for complete data.
The scientific toolchain is ready, but the Fresh Internal Validation reserve is still accruing. No partial candidate performance is allowed to leak into the decision process.
Current gate — complete Fresh source required
2026-08-24 00:00 UTC → 2026-11-15 23:59 UTC · exactly 120,960 one-minute rows before performance reveal.
Scientific branchFROZEN
Frozen HEAD964e514…
Fresh candidate performanceSEALED
Historical TESTLOCKED
SelectionNOT EXECUTED
Live tradingNO
01
Protocol + treatmentEntry-confirmation state machine frozen
DONE02
Three-arm training30 candidates frozen by SHA256
DONE03
Execution guardValidator + adversarial red-team
PASS04
Fresh source accrualFuture reserve must become complete
CURRENT05
Full reveal + selectionOnly after all 360 episodes complete
BLOCKEDcurrent study
One hypothesis. Three symmetric arms.
The only treatment variable is how many consecutive identical directional requests are required before admitting a new position. Economics, risk, PPO architecture and decision cadence remain frozen.
EXP-0013 · TREATMENT
Two-step confirmation
N=2
A directional entry needs two consecutive identical requests before exposure is admitted.
TREATMENT10 candidates
EXP-0014 · TREATMENT
Three-step confirmation
N=3
The strictest preregistered admission gate requires three consecutive matching requests.
TREATMENT10 candidates
fresh validation
A sealed reserve, not a dashboard metric.
The page may show calendar progress and structural readiness. It deliberately does not show candidate returns, ranks, winners, or partial validation statistics.
Fresh reserve calendar
PERFORMANCE SEALEDEnds 2026-11-15 23:59 UTC
—days left
—hours
—minutes
—calendar accrued
Important: calendar progress is not a claim that the source has passed structural validation. Exact row count, timestamps, schema, causality and SHA256 must still pass the release gate.
Evaluation matrix
STRUCTURE 360/36030 candidates × 12 windows
Every cell = one complete candidate/window episode◆ SEALED
independent CI evidence
Checks before claims.
The frozen execution-guard toolchain is compiled and independently reviewed on the technical CI branch without touching Fresh performance or Historical TEST.
33/33
Independent validatorExecution-guard contract, boundaries, invariants and release conditions.
49/49
Adversarial red-teamFail-closed source review against leakage, TEST access and unsafe execution paths.
360/360
Matrix structureThirty frozen candidates mapped across twelve disjoint Fresh windows.
scientific lineage
Every gate leaves an audit trail.
The current study inherits economics and causality from the frozen parent stack, but candidate selection and model initialization are not inherited from the closed predecessor.
FOUNDATION
Causal MarketGenomeMulti-timeframe market data, normalization and deterministic observation contracts.FUTURES STACK
Execution economicsFees, spread, slippage, funding, leverage, margin and risk semantics.PREDECESSOR
TURNOVER CONTROLClosed with 0/30 eligible candidates. No champion promoted.CURRENT
ENTRY CONFIRMATIONThree preregistered arms; 30 frozen candidates; Fresh reserve sealed.HEAD 964e5140ae7d7dfa108ead9cbae1784432ab3358NEXT GATE
Full Fresh revealOnly after complete source, release checks and all 360 episodes.technical contract
Small treatment. Frozen environment.
The study modifies entry admission only. Directional symmetry, execution friction, risk priority, reward semantics and PPO configuration remain fixed.
Observation & clocks
Observation(474,) float32
Parent features472
Added state2
Policy cadence30 min
Execution clock1 min
ActionsFLAT/LONG/SHORT
Training freeze
AlgorithmSB3 PPO
PolicyMlpPolicy
Network[256,256]
ActivationReLU
Outer transitions33,340 / arm
Candidates10 / arm
Fresh discipline
Source rows120,960
Windows12
Action min/window10,000
Predictionsdeterministic
Partial revealFORBIDDEN
Historical TESTLOCKED
research atlas
Explore the full scientific record.
The public site now separates experiment history, provenance, methodology and the current preregistered protocol while keeping one machine-readable status snapshot.
Experiment ledger →
HistoryEXP-0003 → ENTRY
Closed predecessor0 / 30 eligible
Current armsEXP-0012/13/14
Scientific lineage →
Scientific HEAD964e514…
CI HEADfc0c6b42…
Latest CISUCCESS
Methodology →
Reveal360 / 360 first
Fallback winnersFORBIDDEN
DeploymentSEPARATE GATE
360-episode map →
Geometry3 × 10 × 12
Filtersarm / candidate / window
PerformanceSEALED
Audit console →
Latest CISUCCESS
Independent33 / 33
Red-team49 / 49
Current protocol →
Observation(474,) float32
Decision clock30m
Fresh source120,960 rows
status.json →
Formatmachine-readable
Snapshot2026-09-18
Public schemav1.0
Next scientific gate
Fresh end15 Nov 2026
First complete check16 Nov 2026
Performance nowSEALED
research principles
Built to resist self-deception.
The objective is not to manufacture an attractive backtest. The objective is to maintain a research process where assumptions, failures and authorization boundaries remain visible.
[01]
Causality first
Information enters observations only after it is economically available.
[02]
Frozen decisions
Treatments, windows, candidates and selection rules are specified before reveal.
[03]
Real friction
Fees, spread, slippage, funding and risk mechanics stay inside the economics.
[04]
Failures remain visible
Closed studies and ineligible candidates remain part of the scientific record.