01 · METHOD
Agents hold the watch. Coaches make the call.
The shape of the agent–coach partnership on a Groundforce Intel program — what twenty-four-hour agents watch, what only a human coach decides, what is stored and for how long, and why this beats fully automated coaching apps.
02 · AGENTS IN THE LOOP
What agents actually do.
Twenty-four-hour agents watch the unit. They are a thresholded surveillance layer — not noise — and they surface prioritized alerts, not directives. Every alert they open lands on a coach's queue with rationale attached.
Rolling-window deviation from the athlete’s own baseline — not a population reference. Breakfast-noise filtering on ingest.
Duration × efficiency, tracked across rolling windows. A three-night run below the athlete’s tolerance flag is enough to open an alert.
Manual RPE log + auto-captured session weights against the periodized template, producing a load-to-tolerated ratio per athlete.
Per-day ratio of prescribed + actual load to the athlete’s current tolerated ceiling. The single number that crosses the bar fastest.
What trips an alert — and what we silence on purpose.
Agents do not page a coach on the first bad night. They page when a per-athlete baseline has been breached by a sustained, configurable N-σ drift. False-positive budgets are tuned per unit.
Per-athlete baseline
Each athlete carries their own 30–90 day reference window. No population tables, no static HRV floors.
Standard-deviation breach
Alerts open on a configurable N-σ deviation (typical: 1.5–2.0). Tuned per unit, not shipped as a default.
Drifts-in, not spike-in
A single bad night does not page a coach. A sustained drift does — the kind that arrives 2–4 days before attrition.
Suppressed by design
Signals already in tolerance, signals redundant with another already-flagged metric, and signals below the unit’s noise floor are silenced at the model.
What one intervention-queue entry looks like.
Each entry is a one-line payload — athlete, signal, recommended action, confidence — that a coach approves in one read. Below is the morning's queue the coach would open, drawn from the agent set that ships at launch.
Coach console · Intervention queue
0548Z · 3 alerts
[05:48] agent.hrv-watch
HRV −6.1% vs. 30-day baseline · A-12
queue.open · recommended: deload + zone-2 swap · confidence: 0.78
[05:52] agent.program-tune
Load-to-tolerated 1.12 (3-day trend) · A-31
queue.open · recommended: coach review · confidence: 0.62
[05:55] agent.sleep-coach
Sleep efficiency 71% (3 nights) · A-07
queue.open · recommended: caffeine cutoff 1400 · confidence: 0.71
03 · COACH IN COMMAND
What only the human coach decides.
Agents recommend; coaches decide. NSCA TSAC-certified staff remain the authority on programming. We surface prioritized alerts and recommended interventions so they can lead, not chase data.
Every recommendation is an entry, not a directive.
An override is a positive signal — it means the coach's judgment diverges from what the model would have shipped, and the divergence is loud.
Override trains the model.
What the model recommends in month six looks different from what it recommended in month one — because the unit's voice has been written into it.
Where does the human coach sit in the loop?
In it. Every agent-generated intervention arrives as a recommendation, not a directive — coaches approve, override, or bypass, and the override signal feeds back into the model. NSCA TSAC-certified staff remain the authority on programming; we surface prioritized alerts and recommended interventions so they can lead, not chase data.
04 · DATA OWNERSHIP & RETENTION
What is stored, for how long, and who owns it.
A single, fenced description of the data posture — what flows through Groundforce Intel, what is persisted, where it lives, and how it leaves with the unit on exit.
Ingest stream (wearable + manual), per-athlete rolling baselines, intervention-queue decisions and the timeline they sit on, override-log signals that train the model.
Scoped per assessment, briefed to the contracting officer. The default holds active-season ingest for the season plus a single rollover; configurability is part of the procurement call, not a marketing-page number.
Region-configurable per unit. Air-gapped and on-prem options keep data on sovereign infrastructure for the Agency tier.
The unit does. We never train external models on unit data, period — and the override history travels with the export at exit.
Who owns our athlete data, and where does it live?
You do. Groundforce Intel is a scoring and briefing layer on top of data your athletes consent to share with their program — we do not train external models on unit data, period. Data residency is configurable per region, and an exit clause carries your data with you on the way out.
Specific procurement-environment numbers — retention windows, residency options, and air-gapped/SSE requirements — are briefed at the assessment, where the contracting officer and the unit's data owner sit down with a performance lead.
05 · THE DIFFERENCE
Why this beats fully automated coaching apps.
An agent plus a coach beats either alone. Fully automated coaching apps ship recommendations on a publication calendar; Groundforce Intel briefs a coach against the unit's operational tempo, in queue, one athlete at a time.
vs. content libraries + coaching-hour products
The shelf life of a finished media library is the publication calendar. The shelf life of an at-risk athlete is the duration of a three-night sleep drift. We are built for the second clock.
vs. unit-template S&C SaaS
Templates remain intact. The program still moves as one. At-risk athletes get pulled before they break, not after.
How is this different from SOFLETE, MTNTOUGH, or GPS Human Performance?
Those are content libraries and coaching-hour products — videos, plans, and tuned programming delivered to a unit. Groundforce Intel is an agentic 24/7 watch: ingest, baseline, brief, and queue per-physiology program adjustments that scale to the unit’s operational tempo rather than the publication calendar.
One sentence, the whole posture.
We take the part of a coach's day that is logging, watching panels, and reconciling spreadsheets, and hand them a ranked queue. What they do with it is theirs.
Want the assessment briefed to your contracting officer?
We'll scope the watch, the residency, and the procurement path on one call.