21 modules across 7 families. Each solves a problem nobody else will touch — because the answer requires measurement, not promises.
The fidelity battery as a service — publication cadence floor, two-sided coordination structure, per-side staleness distributions, spread regimes, freeze/rollover behaviour — on the user's own broker feed.
"What does your broker's feed actually do when you measure it instead of trusting it?"
HAR-RV with OOS skill, jump/liquidation shares, cost proxies, tails, session structure — institutional-grade per-pair descriptive analytics retail cannot buy.
Bundled inside P1's subscription. Never standalone.
Statement-import truth about what execution actually cost — slippage distributions, spread-regime context, session timing. Retail TCA effectively does not exist.
Upload-and-check honesty for backtests — PBO across CPCV paths, deflated Sharpe with honest trial counts, effective-N + collapse ratio, dual-denominator FDR.
Customers search for exactly this and find papers, not products.
P2's data-side sibling — the battery pointed at the user's dataset: carry-forward merge artifacts, per-side staleness, gap structure, bar-provenance checks. "Your stats can be clean and your data still lying."
The event-study verdict engine — effect estimates with stationary-block-bootstrap CIs, HAC standard errors, Benjamini–Hochberg FDR masking. Zero-dependency Rust; verdicts are byte-reproducible.
Its keystone self-test plants a known effect and demands recovery — while killing 40 placebos.
The desk-readiness census productized for funded-account candidates — drawdown-halt monitoring that never fabricates equity, sizing grids, risk-of-ruin surfaces, correlation-adjusted heat.
Their #1 failure mode is a rules breach — exactly what fail-closed machinery polices.
"What would 2015-01-15 do to your book today" — historical shock replay (SNB, COVID, carry-unwind, gilts) against current positions with gap-through vacuum-fill honesty and ruin flags.
Pre-trade news gate built on per-pair range-lifts measured per release — published with the artifacts, NOT YET PUBLISHED — and a PIT forward calendar, versus the folklore calendars everyone uses free.
Institutional-grade provenance discipline for raw market data. Fetch → raw PIT-as-published archive → parse → content-hashed manifests → fidelity battery on ingest → completeness ledgers → quarantine.
Vendors sell data. Nobody sells provenance discipline around data.
Registry-as-a-service — pre-registration, the FDR book, walk-forward enforcement, burned doors, kill ledger. Smallest market, deepest moat; the buyers who want it want it badly.
The rails, ack ceremonies, protected paths, claim grades, can-fail proofs, append-only registers — the machinery that caught a forged ratification in production, with receipts. Not a trading product at all.
Biggest addressable market of the five.
Fault doctrine for trading infrastructure — an append-only fault registry where a code cannot claim Active status without a proven emission site, and a crash-honest journaling bus that accounts for its own losses.
"records_written: 0 means the bus was unreachable — never a quiet system."
Dealer-terrain from listed FX options — strike-level OI congregation, risk reversals, skew migration, expiry walls. An established category in equities, near-empty in FX retail.
PIT-certified, prefix-stable, never-repainting regime labels per pair — vol tercile, ADR regime, Hurst roughness, session/liquidity state as an API subscription.
"Never repaints — prefix-stable by construction."
Dependence-aware COT analytics — z-scores, flip detection, and the multivariate joint change-point lane (Hotelling-T²+CUSUM+PELT with attribution). Honest nulls, joint-break attribution nobody else computes.
Event playbooks — per-chain, per-layer response distributions after commodity/macro shocks, with fizzle base rates printed beside resolution rates. Institutional memory as a subscription.
"Resolves or fizzles — always published with n visible."
Multivariate change-point detection — Hotelling-T² with Ledoit–Wolf shrinkage, CUSUM, and PELT — with per-channel attribution: which series moved, decomposed per changepoint. Surrogate-calibrated, pure-stdlib, unit-tested hand-rolled linear algebra.
Not "the regime changed" — which channels changed it, and when.
A working quant cockpit — ~90 panels — where every displayed number carries provenance (git SHA + sha256 of its source artifact) or the panel honestly refuses to render. Born from the incident where fake telemetry was painted into a cockpit: the anti-cockpit.
"It cannot display a number it cannot prove."
Trace every number on a trading UI through component → loader → artifact → publisher → computation, with FABRICATED as a first-class finding status. Executed twice on our own cockpit — it caught real fabrications: hardcoded statistics, random-number telemetry.
We ran it on ourselves first. It found things.
Prove a CI gate can actually fail — plant violations it must catch via self-reverting drills, and grade the evidence with a receipt taxonomy: BEHAVIORAL / STRUCTURAL / DECLARED-GAP. We found one of our own parity gates green over nothing.
"A gate that has only ever passed proves nothing."
The line: sell the instruments, never the alpha.