Model Docs
What a model can change, what it cannot change, and how to interpret model comparisons without reading more certainty into them than the system supports.
Methodology Reference
Current product links stay exactly the same. Once you are in methodology, use this reference strip to move between related sections without going back through the app UI.
FR means FiteQuant Rating: the internal fighter rating framework used to organize model inputs across the platform.
On This Page
Use this page to understand what model settings can actually influence, how comparisons should be read, and where fitting risk still remains.
A model changes weighting and interaction behavior on top of stored fighter data. It does not rewrite source bouts, odds, or fighter records.
At a high level, a model can change:
Models do not change the underlying fighter profiles, result records, or odds source data already stored by the platform.
They also do not create extra sample size. A model with attractive recent numbers on a small boxing sample should still be treated as a hypothesis, not proof.
Normalization is fixed to fixed_theoretical. The point of that constraint is comparability: models are evaluated on the same scoring frame rather than each model being allowed to redefine the scale.
That makes model-vs-model inspection cleaner, but it does not remove overfitting risk. It only keeps the scoring surface consistent.
Since Created evaluates from the model's initial creation.
Since Last Edited evaluates only the period after the current parameter set was locked in.
For skeptical review, Since Last Edited is usually the cleaner lens because it avoids mixing different parameter regimes into one headline number.
Backtest Lab is a model inspection surface, not the public tracked-results ledger. It is most useful for comparing parameter sets against the same stored fight, result and odds pool.
All model leans treat every usable directional lean as a one-unit strategy bet. Value/advantage modes add a market-difference requirement before a lean becomes a simulated bet.
All three modes protect deterministic history. Win/finish/activity/trend records are reconstructed only from bouts before the sampled fight, with the sampled fight explicitly excluded. Weight-class movement uses verified pre-bout division evidence and becomes neutral when that evidence is unavailable. Height/reach is treated as a stable physical-profile dimension. A broader mode never grants permission to feed the target result or a later objective record backward into a prediction.
Time Safe: On (strict) is the cleanest audit mode. Bout coverage is limited to the strict public-result pool: settled fights that FiteQuant tracked or recovered with enough pre-fight prediction, result and odds evidence to treat them as public ledger candidates. Strict mode uses pre-fight snapshots and disables opponent-rating-derived fields because their upstream rating provenance cannot always meet the strict boundary. Current SSI-derived style or stance data is never substituted for missing historical context.
Time Safe: On (non-strict) uses the strict pool plus explicitly verified non-strict rows. These are settled canonical bouts with closing odds and timestamped pipeline evidence, but without full public tracked-result status. It preserves the same deterministic objective and weight-class protections as strict mode, while allowing the documented opponent-rating-derived exceptions for targeted lab models. Current SSI-derived style or stance data is still not substituted for missing historical context.
Time Safe: Off uses the broad stored canonical fight, result and odds pool inside the selected window. It exists primarily to explore current SSI-derived subjective and style profiles across fights that pre-date FiteQuant tracking. Those SSI-derived inputs are not claimed to be historical snapshots, and opponent-rating-derived fields may include non-strict upstream values. Deterministic objective history and verified weight-class handling remain protected. Off means broader SSI provenance and coverage; it does not mean target-result leakage is allowed.
For serious model comparison, change one lens at a time: strategy, market bucket, confidence tier, time-safety mode, or objective history window. The useful signal is usually the pattern across runs, not one isolated ROI cell.
The AI confidence control is not a claim that the model "knows" when it is right. It is a weighting control that changes how much subjective inputs matter when the underlying subjective profile was extracted with stronger or weaker confidence.
If you want to stress-test whether the subjective layer is helping, compare runs with objective influence increased and subjective confidence tightened rather than taking the confidence label at face value.
You can verify model behavior inside the product by changing one parameter block at a time, then checking how the same settled sample behaves in Backtest and Results.
The system is designed so a user can inspect which settings changed, what window they are judging, and whether the resulting prediction set remained coherent.
The FiteQuant model is the platform baseline. It is meant to be a reference point, not a claim of universal optimality. Non-admin users cannot save model names containing FiteQuant.
Related
