How Grinding Reps Corrupt Your Load-Velocity Curve and Skew 1RM Estimates

The Rep That Looked Fine on Video

A lifter I worked with last spring came in convinced his squat 1RM had stalled at 165 kg for two months. His load-velocity profile agreed with him — the regression put his estimated max right around 163 kg, session after session. Then he hit a genuine 172 kg single in a meet, no belt panic, no grind, clean depth. The profile had been lying to him by nearly 5%, and the cause wasn't his training. It was one rep at 70% that he'd ground out three weeks earlier while distracted, mid-conversation with a training partner, and that single data point had been sitting in his regression ever since.

This is the failure mode nobody warns you about when they hand you a load-velocity profiling protocol. Everyone talks about load selection, warm-up structure, and reps-per-load. Almost nobody talks about what happens when one of those reps wasn't actually maximal — because the athlete eased into it, hesitated at the sticking point, or simply wasn't locked in that set. The velocity reading from that rep is real. It just doesn't belong in the same dataset as reps performed with genuine maximal concentric intent, and mixing them produces an estimate that looks precise and is quietly wrong.

What a Grinding Rep Actually Does to the Regression

A load-velocity profile assumes every plotted point represents the fastest concentric velocity the athlete can produce at that load. The regression line — and the 1RM estimate you pull from its x-intercept at the minimal velocity threshold (MVT) — is only as good as that assumption holds across every point.

A grinding rep breaks the assumption in one direction only: it produces a velocity that is too slow for the load, because the athlete decelerated early, paused unintentionally at a sticking region, or simply didn't drive with full intent from the first millimeter of the concentric phase. Sanchez-Medina and Gonzalez-Badillo (2011) demonstrated that velocity at a fixed sub-maximal load is exquisitely sensitive to effort — their fatigue protocol showed that as little as a 10% within-set velocity loss already reflects measurable neuromuscular and metabolic changes, meaning the sensor is picking up real physiological signal, not noise. The problem is that a distracted or half-committed rep produces the same kind of velocity suppression as fatigue does, and the regression algorithm has no way to tell the two apart. It just sees a slow rep at a given load and treats it as ground truth.

Because a linear regression minimizes squared error across all points, a single anomalously slow rep at a mid-range load (50-70% 1RM, where most profiling points sit) pulls the entire line downward and rotates its slope. The x-intercept — your predicted 1RM — shifts left. In practice, one contaminated rep out of five to six loading points is enough to shift the estimated 1RM by 4-7%, which on a 150 kg squat is 6-10.5 kg of prescribed load sitting in the wrong place for every session that references the profile.

What the Research Shows About Intent and Velocity

The load-velocity literature is built almost entirely on protocols that explicitly instruct athletes to move the bar "as fast as possible" and that discard sets where this instruction clearly wasn't followed — which is itself evidence that researchers know intent contamination is a live threat to data quality, not a theoretical one.

Gonzalez-Badillo and Sanchez-Medina (2010) established the foundational squat load-velocity relationship using strict rep inclusion criteria: only reps performed with maximal intended velocity were retained in the dataset used to build population norms (r = 0.97-0.99 for the load-velocity relationship when intent was controlled). Their protocol required real-time velocity feedback specifically so lifters could self-correct in the moment rather than coast through a rep — an implicit acknowledgment that without that feedback loop, effort drifts.

Separately, Pareja-Blanco et al. (2017) compared velocity-loss-based autoregulation (stopping sets at 20% velocity loss) against fixed-rep training and found the velocity-loss group achieved equivalent strength gains with roughly 40% less total volume over a 6-week squat mesocycle. That result depends entirely on the velocity-loss cutoff being trustworthy — if reps within a set contained inconsistent intent, the 20% threshold would trigger at the wrong moment, either ending sets too early on inflated fatigue readings or letting genuinely fatigued athletes grind out extra reps because one earlier rep in the set was itself a slow, low-intent outlier that reset the baseline comparison point. Both papers converge on the same underlying requirement: velocity data is only informative when intent is standardized, and neither the regression math nor the autoregulation cutoff has a built-in mechanism to detect when it isn't.

How to Spot a Grinding Rep Without a Force Plate

You don't need lab equipment to catch most contamination before it enters your dataset. Three signals, in order of reliability:

Within-load variance

At any given load in a profiling session, two or three reps performed with genuine maximal intent should cluster within about 0.03-0.05 m/s of each other. A rep that deviates more than 0.05 m/s from its same-load partners is the single most reliable red flag — and it's the exact threshold the original Gonzalez-Badillo protocol used for exclusion. Don't average a slow outlier in with the good reps; drop it and keep the best two.

The velocity-time shape of the rep

A maximal-intent rep accelerates hard out of the bottom and holds or gradually decelerates through lockout. A grinding rep shows a visible mid-rep velocity dip — the bar slows measurably at the sticking point before recovering, rather than following one continuous deceleration curve. If your device shows a velocity-time trace, that dip shape is unmistakable even at loads well below true 1RM.

Self-report timing

Ask the athlete immediately after the rep, not at the end of the session: "Was that full effort from the first inch?" Delayed self-report is unreliable — people rationalize a slow rep as intentional after the fact. Asked immediately, most athletes will honestly flag a rep where they eased in, got distracted, or hesitated, especially once they understand it's corrupting their own 1RM number.

• Discard any rep more than 0.05 m/s slower than its same-load partners rather than averaging it in.
• Treat a visible mid-rep velocity dip as disqualifying regardless of the final velocity reading — the shape matters more than the endpoint.
• Re-test the load immediately rather than substituting a rep from a different session; conditions drift too much to splice sessions together.

A Data-Cleaning Protocol for LV Profiling Sessions

Build the habit of screening reps as you go rather than cleaning the dataset after the fact, when memory of which rep felt off has already faded. This is the sequence that keeps a five- or six-load profile trustworthy from the first session:

1. Brief before every load, not just once at the start: Restate "full speed, every rep" before each new load, not just at the top of the session. Focus drifts by the fourth or fifth load, which is exactly where mid-range contamination tends to appear.
2. Take three reps per load, keep two: Never build a profiling point from a single rep. Three reps let you identify and discard one outlier while still keeping two clean, matching data points at that load.
3. Check the 0.05 m/s spread rule live: If your device shows velocity per rep in real time, glance at the spread before moving to the next load. A spread over 0.05 m/s means one of those reps needs to be repeated now, while the setup and warm-up state are still identical — not reconstructed from memory later.
4. Log conditions, not just numbers: Note anything that could explain an outlier — a missed cue, a training partner interrupting the set, unusual foot positioning. This turns a mystery outlier into an explained one instead of a silent contaminant in next month's re-test.
5. Re-run the regression with and without suspected outliers: If removing one point shifts the estimated 1RM by more than 3%, that point was carrying disproportionate leverage — a strong sign it was contaminated rather than genuine biological variance. Trust the cleaner regression.
6. Validate against a recent true 1RM whenever one exists: If your cleaned profile still misses a known recent max by more than 5%, don't assume the profile is fine and the max was a fluke — assume there's a contamination source you haven't caught yet, most often at the mid-range loads where reps are hardest to keep sub-maximal in load but maximal in intent.

PoinT GO Integration

PoinT GO's 800 Hz IMU captures the full velocity-time trace of every rep, not just a single mean-velocity number, which makes mid-rep deceleration dips visible in the app rather than invisible inside an averaged figure. During a profiling session, reps that fall outside the expected 0.05 m/s spread for a given load are flagged automatically so you can repeat them on the spot instead of discovering the contamination weeks later when the predicted 1RM stops tracking reality. The app also timestamps flagged reps against session notes, so a pattern of contamination at the same load or the same point in a session becomes visible across weeks of training data. See also: How to Calculate 1RM from Velocity Data for the underlying regression math this protocol protects.