Everything that determines whether a GioModel forecast deserves belief — the construction philosophy, the validation rules, the evidence gates, the failures, and the delayed grading of every past forecast — is free and always will be. Payment buys today's frozen numbers, never a different standard of proof.
State exactly what is predicted, at what moment the prediction is made, and what information exists at that moment.
Fix a point-in-time panel, hash the raw provider response, and record provenance before any modeling begins. The panel cannot move afterwards.
Turn price, volume, and market context into signals without letting future information leak backward.
Compare candidates inside purged, embargoed walk-forward splits that preserve the market clock.
Block-bootstrap uncertainty, calibration, regime conditioning, and sensitivity to winsorization — then check whether the result survives.
Separate the best development candidate from the evidence required to deploy it. Publish both, including the gap.
Resolve every horizon on its exact trading session, record the realized close once, and never revise it.
An audit of the gate code found two checks that could not do the job they claimed. Neither frozen release has been edited — the artifacts stand as published — but what we claim they prove has been corrected downward, and the corrected counts are what every page now shows.
Corrected counts: every v3 model drops one; NVDA 4 → 3, FUTU 5 → 4, PLTR 3 → 2, UBER unchanged at 4. No model gained a gate from this audit, and none was expected to.
Every model is NO-GO. That is not a temporary state waiting on more data — under the gate set as originally written it is permanent, and saying so is more useful than leaving the count unexplained.
G9 cannot pass at release. That is deliberate and correct: it is the gate that stops a backtest impersonating a track record. It requires post-freeze outcomes that, by definition, do not exist on the day a model is frozen.
G7 could not be passed on skill — and as written it was not measuring skill. The direction call is the sign of the simulated median, which is inherited drift; it comes out “up” at 83–100% of origins, and in several cases the reported hit rate was arithmetically identical to how often the stock simply rose. A hardcoded “always up” would have passed it. The test has been rewritten to compare the call against the majority-class rival on the origins where the two disagree. That is strictly harder, and every model in the universe is expected to fail it.
While the direction gate stood, the ceiling was seven of nine at freeze and eight of nine ever. Since no horizon is approved until every gate passes, a directional product built on these models could not be approved — not eventually, but at all. That is what forced the scope decision below.
There is one honest route to a deployable product, and it is a decision about scope rather than a change to a threshold: a model that does not publish a direction call does not need a direction gate. That decision has now been taken.
As of 15 August 2026 the seven position-forecast models publish a calibrated P10–P90 range and nothing about direction. The probability of finishing higher is gone from the artifacts and the member workspace, and G7 went with it. The protocol was predeclared and published before the pipeline was run, including what result was expected, so the outcome could not be chosen after the fact.
One methodology fix shipped with it: the regime threshold is now computed as-of-origin rather than from the whole sample. It moved results in both directions, which is what a correctness fix looks like — PYPL's coverage gate went from fail to pass, and NFLX's went from pass to fail. No parameter was searched and no configuration was chosen on its gate outcome.
The ceiling at freeze is now seven of eight, since G9 still cannot be earned on the day a model is frozen. Three models — PYPL, CRM and ORCL — sit at six of eight, one gate short of that ceiling. Nothing is approved, and nothing will be until the forward evidence exists.
The gates are declared before the model is built, so a disappointing result cannot negotiate its way through them. A release publishes how many it passed and exactly which ones it did not. The position-forecast family publishes eight since the v4 interval-only release; every other family still publishes nine.
There is no single site-wide G1–G9, and it would be dishonest to imply one. A gate protocol is fixed when a model family is built and can never be edited afterwards, so the five families in the universe carry five different predeclared sets. An interval model and a direction classifier are not measuring the same thing, and their gates reflect that.
The set below is the position-forecast protocol, which governs seven of the fifteen models — now at v4, where it publishes eight gates rather than nine. The others are listed underneath. Gate numbers are local to a protocol: model 3's G3 is a log-loss test, the position-range G3 is an interval-score test, and the position-forecast G3 is a duplicate-date check. Thresholds differ too — position-forecast requires P10–P90 coverage inside [75%, 85%], while the position-range protocol allows 72–88% aggregate. Every stock page renders its own model's rule, measured value and verdict straight from that release's frozen artifact.
At least 1,750 valid target sessions in the frozen panel. Below that, an interval estimate is a story about noise.
Zero quarantined target rows. A row whose high is below its open, or whose volume is missing, is removed rather than repaired — and its removal is recorded.
No duplicate session dates in the target series. Duplicates silently double-weight a day and inflate every downstream statistic.
At least 90% of independently held closes must match the fetched panel within 0.5%. A single provider agreeing with itself is not verification.
Out-of-sample P10–P90 coverage must land inside [75%, 85%] at every horizon. A band that covers 99% of outcomes is not cautious — it is uninformative.
Walk-forward interval score no worse than a scaled random-walk baseline at every horizon. If a model cannot beat 'tomorrow looks like today', it has earned nothing.
Retired on 15 August 2026 together with the directional claim it gated. As written it compared the 5-session hit rate to a coin flip, but the call was the sign of the simulated median — inherited drift — which said 'up' at 83–100% of origins, so a hardcoded 'always up' passed it. The position-forecast family no longer publishes a direction call and therefore no longer carries this gate. The other families still do.
The published candidate is whichever variant scored better out-of-sample; the loser is kept on record as its challenger. The winner is not chosen after seeing which story reads better.
Independent post-freeze tracking must accumulate before directional use. This gate cannot be passed at release time by construction — it is what stops a backtest from impersonating a track record.
| Protocol | Models | What its gates measure |
|---|---|---|
| Position-forecast model v4 (interval-only, eight gates) | PYPL · CRM · NFLX · ORCL · AMT · WMT · AMD | Sample size, quarantined rows, duplicate dates, cross-verification, walk-forward coverage, naive-baseline interval score, challenger discipline, forward paper evidence. No direction gate, because v4 publishes no direction call. Supersedes v3, which remains published. |
| Position-range model v1 | UBER · NVDA · RDDT · PLTR · FUTU | Data quality, no-lookahead prefix invariance, weighted interval score against baseline, median absolute error, coverage (72–88% aggregate), direction calibration by Brier, fold consistency, seed reproducibility, independent forward evidence. |
| Model 1 Phase B direction classifier | ASTS | Brier improvement with bootstrap lower bound, ROC AUC, circular-shift tail rate, champion-versus-Run-1 improvement, calibration slope, calibration intercept, expected calibration error, outer-fold Brier rate, block and event-window sensitivity. |
| Model 2 Phase B direction classifier | HOOD | The same nine tests as model 1, re-frozen against the signed Core Ridge challenger rather than the Phase A replay. |
| Model 3 audit-corrected direction classifier | NKE | Holdout Brier improvement, ROC AUC, log-loss improvement, calibration slope and intercept, expected calibration error, development-fold Brier rate, five-run stability, forecast non-flatness, circular-shift permutation. |
Open any stock page for the full frozen table — rule as written, value observed, and pass or fail — for that model specifically.
The Micro Trading Lab carries a separate five-minute execution protocol and a different G1–G9. It remains FROZEN_UNFITTED and NO-GO, so it is not counted among the five fitted stock-model families above.
NO-GO does not mean the research contains no information. It means GioModel refuses to convert incomplete evidence into a trading recommendation. Members receive the ranges, the probabilities, the diagnostics and the evidence — everything the model produced. What they do not receive is a claim that the model has earned directional deployment, because the deployment gate is independent of payment. No subscription moves a gate. A model that cannot say no to itself has nothing worth selling.
A dazzling score built with information that did not exist at prediction time is not a model; it is a very convincing mistake. Every feature is computed from a strictly prior window, and prefix-invariance is tested rather than assumed.
Being the least bad of five candidates ranks a search. A challenger retested on the same dates is more development evidence — not a fresh holdout. The two are never added together.
When the evidence says no, the specification is rejected and the process is kept. The failures stay published; they are the reason the eventual pass would mean anything.
Step 7 of each lesson prints a current, tradeable range and is the one part held for members. Everything else on this page, including every gate result and every published failure, is open.