GHCOPILOT: audited candidates and prospective paper laboratory
Agent/model attribution: GHCOPILOT - AI assistant using Copilot SDK in VS Code.
Underlying model identifier is not exposed by this runtime; no model name is invented.
Research / paper only. Not proven, not money-ready, no real orders.
Historical candidates and future paper outcomes are separate. An empty class means no eligible candidate,
not a fabricated winner. Portfolio returns are simulated, with stated limitations.
Proven picks per asset class: none independently established
As of the October 22 independent audit, zero classes have passed all required evidence checks.
This is a failed proof gate, not proof that no trading edge can exist. No individual future pick is guaranteed.
The proposed framework is not yet enforced by this monitor. Its existing PAPER_EVIDENCE_PASS is a preliminary research result, not full certification: familywise confirmation, measured class-specific costs, certified calendars, adequate power and daily mark-to-market portfolio risk still require independent validation. A 50% win-rate test is not an economic proof for unequal payoffs or dependent trades.
Methods include exact-ledger reproduction, source/identity reconciliation, independent hourly and five-minute price paths, costs, one-slot exposure diagnostics, binomial/Wilson, day/week cluster bootstrap, chronological halves, leave-one-month-out, tail sensitivity and multiplicity stress. DSR/PBO/SPA are not invented: complete aligned trial histories are unavailable. Futures rolls/margin, FX sessions/rollover, equity splits/borrow and ETF/bond distributions still require provenance validation.
Snapshot: October 22, 2026, 22:23 UTC. Certified winners: zero.
6,467 recent outcome rows became 3,350 provisionally retained rows across 220 groups.
Only three crypto groups and one equity group meet the modest n20 / 10 entry-day / positive-net-mean / PF>1 research screen.
All four weekly confidence intervals include zero; removing their top 5% returns makes every mean negative.
These are ranked research leads, not five profitable recommendations per class.
Class labels are not certified instrument classifications: some equity groups contain ETFs and some ETF groups contain FX.
COMMODITY/FUTURES overlap must not be double-counted. Missing replay lineage and first-observation times prevent clean-cohort certification.
The forgotten strategy-by-class table contains 100%-win/PF999 entries and an ensemble PF170.8.
They reproduce as status-based aggregates, not independent executable profits.
The shortened mega m084 aggregate PF10.32 contrasts with seven exact-name replay rows at gross PF0.900 before costs.
Alias identity remains unresolved; different populations must not be pooled.
New peer-selected HYPE/PENGU/JTO certificates do not reconcile to the exact-name signal cohorts checked;
that is a provenance gap, not proof that their separate ledgers are false. UNH has only three basic replay observations.
Two sampled momentum TIME_EXIT records conflict with both independent prices and same-symbol database stop paths.
The OPEN lifecycle snapshot contains 258 raw rows: zero known-age signals older than seven days, four unknown-age signals,
and no OPEN rows with closure/PnL contradictions. Do not mistake historical terminal completeness gaps for a stale OPEN backlog.
Only two source systems recorded raw rows in the inspected 24 hours; historical copy-named sources were inactive.
Institutional evidence is more than a high win rate
Require provenance, an ordered daily marked-to-market portfolio net of realistic costs, matched benchmarks,
dependence-aware confidence bounds, familywise selection control and a sealed prospective readout.
Sharpe>2 and Calmar>2 are mandate targets, not guarantees. A confirmatory p<.01 needs a defined null and test family;
"positive DSR" alone is not a proof threshold. CVaR measures tail loss; billion-dollar capacity additionally requires actual
liquidity, impact, participation, funding and borrow evidence. Those inputs are not available here.
This audit does not change the frozen paper candidates, ledger or preliminary evidence rules below.
No new historical shortlist has been retroactively inserted into the experiment.
Advanced entry-condition and risk audit:
the existing RSI50-70 US-session lane has n394, net PF1.019, lower PF bound0.826 and OOS PF0.981.
The attractive confidence/score cell is selected from 574 tested cells and needs valid paths plus a sealed future cohort.
No additional winner is certified by these filters.
Grok critic baseline: all ten outputs required
Use dated out-of-sample daily marked-to-market equity with actual or explicitly simulated sizing, net of costs.
The current monitor reports realized equity only: missing daily-risk metrics are NOT EVALUABLE, not passed.
Trade count, distinct entry/exit days and observation length. At least 100 trades is a floor, not proof of independence or adequate power.
Net expectancy, average win/loss and realistic total-cost stress at 1x and 2x.
Net profit factor with whole-day bootstrap bounds, longer-block sensitivity and explicit no-loss/undefined draws.
Daily annualized Sharpe and Sortino: baseline aims 1.0 and 1.5, with cash return, downside target and annualization convention stated.
Daily maximum drawdown and time underwater, including unrecovered episodes; baseline drawdown ceiling 20%.
Calmar over the same defined window; baseline aim 0.5, with short-history limitations.
Worst complete month and daily worst-5% expected shortfall (CVaR), with tail sample size.
Matched-date benchmark alpha/beta with uncertainty and relevant class-specific factors.
Passive buy-and-hold net return and Sharpe, plus a prespecified risk/exposure-matched comparison.
Time invested, intraday exposure, average/maximum notional, leverage and cash.
These are baseline report requirements, separate from the stronger institutional mandate.
A defensive sleeve need not beat passive raw return if its preregistered portfolio-risk objective is independently demonstrated.
No reviewed candidate currently supplies all ten verified outputs on a sealed OOS daily book.
Realized equity and drawdown exclude open-position mark-to-market and are NOT full daily risk statistics.
No leverage, $10,000 starting book, $1,000 fixed notional per class. At most one active position per class and symbol.
Per-class scoreboard: historical lead versus prospective evidence
Class / status
Frozen strategy
Historical net evidence
Future closed / active
Future net performance
Coin-flip / paired control
What would count as evidence?
At least 100 prospective paired outcomes, 30 entry days and 90 calendar days. Rules and candidate selection stay frozen.
Net PF ≥ 1.5, positive chronological-half means, entry-day-bootstrap mean CI above zero, and BH-adjusted one-sided win-rate p ≤ .05 against 50% across all seven classes.
Paired net excess return over deterministic hash-selected random-direction controls must also have a positive clustered lower confidence bound.
These gates are deliberately stricter than positive win rate alone. They do not correct every serial dependency or prove future profits, fund outperformance or money readiness.
Fixed round-trip cost scenarios: crypto 0.30 percentage points, others 0.10. These are assumptions, not measured execution costs. Futures exposure is normalized unlevered notional, not a broker contract/margin simulation.
Next completed hourly bar open after first observation; 48-hour wall-clock horizon; stop first on ambiguous bars; adverse gap fills at open. Both strategy/control must resolve before slot reuse.
Independent Yahoo hourly prices, not stored DB exit prices. Missing data stays visible; market sessions, funding, dividends, borrow and roll costs require further audit.
Conservative limitation: hourly gaps, including unverified overnight/weekend closures, block resolution rather than assuming no stop was touched. Session calendars are not yet certified; non-crypto paper coverage may therefore remain incomplete.
Prospective observations (latest first)
Class / symbol
Observed / entry
Direction / random control
Status / marks
Strategy result
Control result
Database feed quality - diagnostic, not certification
Class
Raw rows / last 24h
Latest recorded / signal
Missing signal
Closed / missing PnL / zero PnL
|PnL| >30%
Replay coverage / ambiguous
A zero can be a real flat result or a migration placeholder; it is not silently removed or assumed correct.
Ingestion freshness does not certify signal freshness, executable prices, outcome truth or independence.
Rejected signals and screening exclusionsAll screened strategy groups (selection bias disclosed)
Historical screen selects exploratory positive-net candidates; none is automatically promoted. Known historical re-emission strategy and ambiguous historical rows are quarantined. Historical selection remains in-sample.