Methodology
Cexi estimates training readiness, perceived soreness, and next-workout prescriptions from the information you log. This page explains the inputs, model families, confidence signals, fallbacks, and limitations behind those estimates.
Scope
Cexi has separate systems for recovery context and training progression. They are not one universal “AI score,” and they are not always active on every route. The app records which prediction source or policy path produced a result so that a fallback can be distinguished from a learned path.
The short version
- Recovery estimates are based on recent training, optional health signals, and personal history.
- Progression proposes complete, equipment-realizable sets, reps, loads, and effort targets.
- Missing data, sparse history, unavailable artifacts, and route configuration can select a heuristic or deterministic fallback.
- Confidence describes evidence and uncertainty, not the probability that an answer is correct.
Recovery estimates
Readiness and functional capacity
A two-compartment state-space model represents fast training fatigue and delayed soreness, with a separate estimate of the chance that a muscle can be trained normally. When a compatible model artifact is unavailable, Cexi uses its tested heuristic estimator instead.
Perceived soreness
The soreness v4 path uses direct numeric soreness feedback, per-muscle workout events, recent exposure context, and optional event-time sleep, HRV, and resting-heart-rate features. It is activated only when a validated production artifact is available; otherwise a fallback is used.
How the state changes
Workout stimulus is converted into a non-negative dose and passed through time-decay kernels. A residual state captures unexplained soreness, while a personal adapter learns sensitivity, recovery speed, delayed timing, volume, effort, frequency, repeated-bout, and biometric effects.
What is estimated
Depending on the active path, Cexi can estimate readiness from 0–100, perceived soreness from 0–100, acute fatigue, training priming, hours to a readiness threshold, hours to low soreness, and the main drivers behind the estimate.
Direct numeric soreness labels update the soreness model. Pain, injury, and “not sure” responses are not treated as soreness labels; they are separate safety or feedback signals. Readiness feedback updates the readiness system rather than training the soreness target.
Apple Health access is optional. Without it, Cexi can use logged workouts and manual inputs, but the estimate may have less evidence for sleep- or biometrics-related changes.
Progression decisions
Progression is route-dependent. In the template prescription path, Cexi generates complete candidates rather than choosing only “add weight” or “hold.” Candidates can vary load, reps, working sets, target reps-in-reserve, recovery, deloading, and re-entry after a training gap.
1. Build candidates
Recent loads, reps, RIR, failures, grinders, exercise identity, equipment increments, rep ranges, set schemes, and training gaps define the feasible action space.
2. Apply hard constraints
Non-finite values, invalid ranges, equipment-inexpressible loads, pain or injury restrictions, clinician ceilings, excessive jumps, and excessive set counts are rejected where the active policy supports those gates.
3. Estimate outcomes
When the learned outcome path is available, it estimates completion, missed reps, RIR, capacity change, recovery cost, pain or soreness, adherence, and related outcomes. Its transition component remains heuristic.
4. Select and materialize
A utility objective balances productive progression, completion, effort accuracy, adherence, future performance, recovery cost, failure risk, pain or soreness, abandonment, load jumps, and stagnation. A Bayesian policy can explore among the surviving candidates.
Other routes can use deterministic progression policies such as double progression, linear progression, top-set/backoff, RIR autoregulation, or deload rules. The Python prescription v2 package is a research design and is not described here as a shipped serving system.
Confidence and uncertainty
Cexi treats confidence as evidence strength and uncertainty around a model output. It is not a calibrated probability that a recommendation is correct, safe, or optimal.
Higher evidence
Repeated, recent workout history; direct feedback for the target; required inputs present; and a validated model or policy path active.
Moderate evidence
Enough history to form a personal estimate, but some optional signals are missing or the estimate remains partly anchored to a population prior.
Lower or unknown evidence
Sparse history, missing inputs, a fallback path, or a horizon that cannot be reached reliably. An unknown time-to-threshold is not treated as zero hours.
Prescription uncertainty
Selection uncertainty and selection probability describe posterior exploration behavior. They do not provide a clinical confidence interval or a verified chance of success.
The current iPhone runtime exposes some uncertainty and provenance fields, but it does not expose every interval or evidence-band field used by the research reference implementation. We therefore do not claim public, calibrated 80% intervals for soreness or prescription correctness.
Known limitations
- Model activation depends on feature flags, compatible artifacts, route, and runtime availability. A learned label does not prove that a learned model generated it.
- The template prescription path and ordinary recommended-session path do not currently share exactly the same policy provider.
- The live outcome model can be paired with a heuristic transition model and a minimal feature/history representation.
- Personal estimates are weaker after a break, with sparse logging, inconsistent effort labels, or missing optional biometrics.
- Recovery and wearable signals do not directly personalize every learned prescription decision; some learned decisions use neutralized or limited wellness inputs.
- The system can be wrong, especially when a user's circumstances differ from their logged history. You can override a recommendation.
- Cexi is not a medical device, does not diagnose or clear injuries, and cannot replace a clinician or qualified trainer.
Updates
We update this page when a production model family, policy path, data source, or material limitation changes. Product releases and corrections are listed in the public changelog. For questions about a recommendation, contact support@cexi.app.
Home · Privacy · Privacy choices · Terms · Support