Tested
Regression-guarded engine
Causal audit engine for coaches
Know which inputs plausibly moved performance — and when the evidence is too thin to say.
AthDash turns athlete exports into within-athlete Driver Cards: effect estimates, confidence intervals, confounder adjustment, placebo checks, and a hard no-claim state when the data cannot support the relationship.
Private beta
Real exports in
MCP claim gate
Audits the exports coaches already have
INTERVALS.ICUWHOOPTRAININGPEAKSOURAGARMINFITAPPLE HEALTHCSV/JSON
01 / The problem
Correlation isn't a coaching decision.
Stack enough athlete signals and something will always line up. Sleep, HRV, load, soreness, weather, travel — one pair will look meaningful by chance. AthDash asks the harder question: after adjustment, uncertainty, and refutation, is there enough evidence to say this driver matters for this athlete?
Everything correlates
The cloud of false signals
Sleep, HRV, load, mood, weather — track a dozen series and dozens of pairs will appear related by chance. Most of it is noise wearing a trend line.
One number, no context
A score that flattens everything
A single readiness figure collapses a dozen mechanisms into one digit. It tells you something moved — never which lever to pull, or by how much.
No sense of enough
It never admits uncertainty
Tools render a verdict on three data points as confidently as on three hundred. So when the sample is thin, you're guessing — and you can't tell that you are.
02 / Effect modifiers
The same input helps — or hurts. Depending on the conditions.
Most tools report one average effect. AthDash tests whether the relationship changes under specific conditions — and only promotes the modifier when the evidence clears the gate.
readout · training load → next benchmark
effect (HRV suppressed)−0.056
effect (HRV recovered)+0.022
modifierHRV vs. baseline
evidenceexploratory · n=19
conditional slope
HRV ↗ rising
−0.056
Load hurts
HRV < threshold
Load helps
HRV ≥ threshold
03 / MCP grounding
Keep your coaching agent honest — by contract.
An LLM gives the same confident answer whether it knows or is guessing. AthDash gives your agent the one thing it cannot generate for itself: a verdict it is allowed to make, the evidence behind it, and a hard stop where there is nothing to say.
› athdash_can_i_claim(
driver="sleep_regularity",
outcome="benchmark_workout",
athlete="A.R.")
evidence: SUPPORTED (n=19, p<.001)
license: ADVISE
interval: [+0.78, +1.93]
› athdash_can_i_claim(
driver="ctl", outcome="ftp",
athlete="A.R.")
evidence: INSUFFICIENT (n=3)
license: DECLINE
driver="sleep_regularity",
outcome="benchmark_workout",
athlete="A.R.")
evidence: SUPPORTED (n=19, p<.001)
license: ADVISE
interval: [+0.78, +1.93]
› athdash_can_i_claim(
driver="ctl", outcome="ftp",
athlete="A.R.")
evidence: INSUFFICIENT (n=3)
license: DECLINE
License ladder · low to high authority
DECLINENot enough of their data to judge yet
BORROWLean on the cohort prior, flagged as borrowed
WITHHOLDSignal is within the noise; do not claim it
FLAG_WEAKWeak signal; mention only if asked
HYPOTHESIZEWorking hypothesis; propose an N-of-1
ADVISEState it, but lead with the uncertainty
ACTEstablished; may base a recommendation on it
The contract
Bound the moment it connects
Your agent does not promise to be careful; it is constrained. The MCP instructions bind it to never claim an insufficient relationship, never invent a number, and offer borrowed evidence only as borrowed.
The gate
It asks before it asserts
Before the agent tells an athlete one thing moved another, it calls athdash_can_i_claim. Back comes a verdict, an interval, and a license — or DECLINE.
The license
Authority, capped at the evidence
Every finding carries a rung, DECLINE to ACT. The agent acts within its license, never above it. That is the API answer, not a prompt suggestion.
The trace
Every claim is on the record
Each thing the agent says traces back to the data, the interval, and the license that allowed it. When an athlete asks why, the answer already exists.
Put your name on what your agent ships.
04 / Claim language
The words an agent is allowed to use.
AthDash does not hand an AI coach a vibe. It hands it a small vocabulary with permissions attached, so every answer can stay inside the evidence.
- grounding block
- The LLM-native version of an athlete's findings: evidence state, effect, interval, caveats, and the plain-language directive the agent must follow.
- claim gate
- The MCP call before an assertion. If athdash_can_i_claim returns WITHHOLD or DECLINE, the agent cannot claim the relationship.
- license
- How far the agent may go, from DECLINE to ACT. The estimate carries its own permission instead of relying on the model to self-police.
- read-only tool
- The agent can query AthDash, but it cannot mutate athlete records, rewrite evidence, or manufacture a new effect through the MCP surface.
- borrowed prior
- A cohort signal for cold-start athletes. It may be offered only as borrowed evidence, never as if it were proven for that athlete.
- decline
- A valid answer, not a failure. When the data is too thin, the safest product behavior is to stop the claim before it reaches the athlete.
05 / One engine
Start with an audit. Expand into a system.
The first surface is a Driver Card. The same gated estimate can later feed a roster console or read-only agent tools without losing the interval, caveats, or license.
SIGNAL FLOW · one estimate, three readouts
Report
Driver Card report
Turn a coach's existing athlete exports into a shareable causal audit: effect, 95% CI, evidence state, caveats, and insufficient when the data is too thin.
Roster
Roster console
Run the same engine across a roster and see which relationships are supported, exploratory, borrowed, or declined.
Agent
Agent/API grounding
Expose the same findings through read-only claim gates, so an AI coach cannot assert what AthDash has not licensed.
06 / The rigor
How the engine earns a claim.
Five checks, in order, every time. A claim only leaves the engine once it has survived all of them.
01
Ingest
sourcesAccept the athlete exports coaches already have — wellness, sessions, FIT files, benchmarks — and align them to one athlete timeline.
02
Model
causal structureDeclare the driver, outcome, lag window, and adjustment set. AthDash tests causal hypotheses; it does not treat the registry as truth.
03
Refute
placebo + uncertaintyStress each estimate with confounder adjustment, small-sample uncertainty, and placebo shifts. Fragile relationships are downgraded or withheld instead of promoted.
04
Gate
honestyReturn SUPPORTED, EXPLORATORY, or INSUFFICIENT, then attach a license from ACT to DECLINE.
05
Deliver
with uncertaintyShip the finding as a Driver Card, console payload, or agent-readable claim gate — with the interval and caveats still attached.
07 / Proof
The discipline, without bluffing.
Exports
File-based ingestion
Intervals
Findings carry uncertainty
Refuted
Placebo-fragile claims downgraded
Gated
Hard no-claim states
Read-only
Claim-gated agent tools
08 / Private beta
Private beta.
We are onboarding coaches manually while the hosted console hardens. Early users get the audit runner, Driver Card reports, roster console payloads, and claim-gated agent tools. Team controls, billing, and managed export sources come later.
Now
Audit runner
For coaches who want to test the causal-audit workflow on real athlete exports.
- Export-based ingestion
- Shareable Driver Card reports
- Manual onboarding support
- No data resale
Early users
Roster reports
For coaches who want the same evidence gate across a small roster.
- Roster console payloads
- Supported, exploratory, borrowed, or declined states
- Claim-gated agent tools
- Priority support
Later
Hosted teams
For squads, federations, and performance departments once the managed product is ready.
- Production billing
- Team permissions
- Managed export-source setup
- Department support
09 / FAQ
Questions, answered honestly.
What does AthDash actually claim?+
AthDash claims only driver-to-outcome relationships that clear its evidence gate. Each finding carries an effect estimate, 95% confidence interval, sample size, placebo check, evidence state, caveats, and a license. If the athlete does not have enough usable outcome events, AthDash returns INSUFFICIENT instead of filling the gap with a guess.
Where does the data come from?+
From exports coaches already have: intervals.icu, WHOOP, TrainingPeaks, Oura, Garmin, FIT files, Apple Health, and tidy CSV/JSON. Some sources are parsed as wellness data, some as sessions, and benchmark outcomes are usually added as a simple CSV.
Is this a medical device?+
No. AthDash is a performance-analysis tool for coaches, not a diagnostic or medical device. It does not diagnose, treat, or prevent any condition, and nothing it returns is medical advice.
How is this different from a readiness score?+
A readiness score is one number with no mechanism behind it. AthDash gives you the drivers of performance, the conditions under which each one holds, and the uncertainty around every estimate — so you know which lever to pull, not just that something moved.
Can I export the numbers?+
Yes. Driver Card reports, roster console payloads, and claim-gated API responses all carry the same estimate, interval, evidence state, and license.
What happens when there isn't enough data?+
Nothing gets claimed. The driver is gated as insufficient and held there until enough evidence accumulates to cross the threshold — at which point it's promoted with its interval attached.
Is AthDash causal or correlational?+
AthDash estimates within-athlete causal effects under an explicit adjustment design. It is stronger than dashboard correlation because it models lagged exposure windows, adjusts for declared confounders, reports uncertainty, and runs placebo checks. It is still an estimate, not proof.
Can it keep my AI coach from hallucinating?+
It gives the agent a claim gate. Before asserting that one driver moved an outcome, the agent calls athdash_can_i_claim. The response includes the license, evidence state, effect, interval, and directive. If the license is WITHHOLD or DECLINE, the agent is not allowed to claim the relationship.
What is an effect modifier?+
A condition that may change a relationship's strength or sign. AthDash tests whether the relationship changes under specific conditions and only promotes the modifier when the evidence clears the gate.
Does it replace a coach?+
No. AthDash is decision support: it tells you what the evidence says and how sure it is. The coaching call — and the relationship with the athlete — stays yours.