prediction system

recommendation guardrails

Updated Oct 4, 2026

A safety layer that rounds targets to real gym increments, clamps them to recent anchors, and widens ranges when confidence is low.

Why guardrails exist

The predictor does not send its raw output straight into the logger. logit adds a second layer whose job is not to be clever, but to keep recommendations within a believable working range for the next session.

That distinction matters. The predictor estimates where the anchor set should land. The guardrails decide how aggressively that estimate is allowed to move once it is translated into a real plate-loaded recommendation for today.

1. Gym-increment rounding

Every weighted recommendation is snapped to the increment the user can actually load in the gym. In logit that increment is five pounds in pound mode and 2.5 kilograms in kilogram mode.

g = \begin{cases}5 & \text{lb mode}\\2.5 & \text{kg mode}\end{cases}

Interpretation. g is the available gym increment in display units. The final constrained load is rounded to the nearest loadable step.

Internally, stored loads remain pound-based. The product converts to the active display unit, rounds there, and then converts back to stored pounds so the displayed target and persisted value stay aligned.

2. Anchor clamp around recent reality

The learned forecast is limited to a 10% strength change, then converted to load at the last anchor's rep count. The load clamp uses the rounded previous anchor. Upward movement is limited to one increment. Downward movement allows two increments, or three after more than 28 days away. Positive loads have a one-increment floor.

\Delta_+ = g

Interpretation. d is days since the last matching exposure. The final anchor load ŵ cannot jump more than one step above the recent anchor, and can only fall by two steps unless the layoff is longer than twenty-eight days.

3. Later-set shape constraints

Later visible sets use the latest five matching sessions' median backoff ratios and rep deltas. Missing offsets use the fallback values below.

\hat{w}_j = \min\!\left(\hat{w}_{\mathrm{anchor}}, \hat{w}_{\mathrm{anchor}} \rho_j\right)

Interpretation. Each later set inherits the anchor estimate and is adjusted by the user's historical weight ratio ρ and rep delta δ. The later-set load is never allowed to exceed the anchor load.

Later-set fallback	Weight ratio	Rep delta
Set 2	0.97	0
Set 3	0.94	-1
Set 4	0.92	-2
Set 5+	Steps down to a floor of 0.88	Subtracts one more rep per set

4. Confidence-linked rep guidance

The API includes a rep range whose width depends on the history-quality label. Low confidence gets a wider range. These ranges are not calibrated prediction intervals, and the labels are not probabilities.

s = \begin{cases}2 & \text{low confidence}\\1 & \text{medium or high confidence}\end{cases}

Interpretation. This is a product rule, not an uncertainty estimate learned by the regression.

5. Conservative ceilings for sparse and bodyweight history

Some limits sit above the numeric score itself. A prediction based on only one matching session is always labeled low confidence, even if the raw score would have landed higher. Bodyweight-only predictions also cannot rise above medium confidence.

Sparse history repeats the last anchor. A learned forecast requires at least 14 usable sessions and must beat repeat-last on four chronological checks.