Quant research that stays local.

Every stage, raw file to live order

1. 
   1

   01 Acquire · 02 Clean

   Provenance, then quality.

   Source bytes kept as evidence, catalogs content-hashed. Each session flagged causally against a witness instrument; real regimes are labelled, never dropped.

   TARGETS silent data errors

   • tick-exact binary decode, nothing rounded
   • raw source bytes retained as evidence
   • content-hashed catalogs
   • gap test against a witness instrument
   • robust z-scores on prior sessions only
   • bad-print reversal and median-deviation tests
2. 
   2

   03 Data policies · 04 Calendars

   No look-ahead by construction.

   Development, validation and holdout pinned to exact sessions on the exchange calendar, embargoed, before the first run. Defect sessions excluded.

   TARGETS look-ahead bias

   • chronological dev / validation / holdout
   • embargo sessions between tiers
   • holdout pinned at registration, never slides
   • defect and holed sessions excluded
   • exchange-native session calendar
   • versioned roll rules for continuous contracts, where they apply
3. 
   3

   07 Copilot

   A research partner, bound by quantitative methodology.

   Pre-registration, selection-bias accounting and holdout discipline run on every strategy. The rigour is the default, not a setting.

   TARGETS leakage through the assistant

   • runs on your own AI subscription or a local model, your choice
   • builds, checks and backtests with NERV's own commands
   • designs pre-registered optimizations, not ad-hoc reruns
   • held-out results withheld by default; every read logged
   • can't move the frozen window or spend an OOS look
   • its runs enter the trial register, tagged as copilot
   • validated proposal cards: you Apply, Undo and launch
   • versioned notebook and plan that persist across chats
   • research groups share exact notes between strategies
4. 
   4

   06 Workspace · 08 Hypothesis

   Pre-registration, verified by git.

   Mechanism, kill criteria and the frozen OOS window are committed before the runs they govern. Leakage is counted from commit order.

   TARGETS HARKing and silent re-specification

   • registration committed before the runs
   • kill criteria as metric · operator · value
   • frozen OOS window in the registration
   • registered-after-k and edited-after-m counters
   • every OOS touch logged as a look
5. 
   5

   09 Backtest

   Event-driven and deterministic.

   Tick-level on the Rust core with explicit fill and fee models. Same selection, same fills; every stochastic statistic seeded.

   TARGETS irreproducible results

   • event-driven, tick-level
   • explicit fill model with adverse-tick slippage
   • versioned per-contract fee profiles
   • every stochastic statistic seeded
   • rerun from a sealed code and data snapshot
6. 
   6

   10 Statistics · 13 Canary

   Selection bias, priced in.

   Deflated Sharpe at raw and effective N, BH false-discovery control and PBO via CSCV, all from the actual trial register. Random-entry twins as the null.

   TARGETS selection bias and multiple testing

   • deflated Sharpe at raw, effective and single N
   • effective N by return-correlation clustering
   • Benjamini–Hochberg false-discovery control
   • PBO via CSCV
   • PSR and minimum track-record length
   • stationary block bootstrap
   • Newey–West HAC
   • random-entry null twins, permutation p-value
7. 
   7

   11 Optimize · 12 Gates

   Plateaus over peaks.

   Neighbourhood-median ranking on in-sample results only, walk-forward with embargo. Gates report, nothing auto-rejects, overrides are logged.

   TARGETS overfit parameter peaks

   • plateau ranker: neighbourhood-median objective
   • ranking on in-sample results only
   • rolling or anchored walk-forward, embargoed
   • slippage-stress, one-bar-delay and tick-phase reruns
   • gates report; overrides logged with a reason
8. 
   8

   14 Deployments · 15 Copy trader · 16 Safety

   Research-to-live parity, on every account.

   The validated commit runs live on the same core, flat at every close, and copies across accounts on different broker connections.

   TARGETS research-to-live drift

   • flat at every session close, backtest and live
   • fail-closed reconciliation: missing report ≠ flat
   • the validated commit runs live on the same engine
   • runs on hardware you control
   • every check re-run at every session open
   • built-in copy trader across broker connections
   • per-account size, loss limit, stop and Flatten
   • a refused account is set aside; the rest trade on
   • broker and firm rules stay yours to follow