The week our AI coach taught our training engine

Navara lets you connect your own AI to your training plan. The AI can read your training and propose changes, but nothing touches your plan until you approve a diff in the app. We designed that gate on principle: the athlete decides. What we didn't predict was what would happen the first time the AI and our own safety rules disagreed.

The argument

Two weeks out from a marathon, we asked Claude to review the taper. It came back with a classic, well-supported recommendation: keep the volume drop, but hold on to intensity — a short mid-week tempo, and a race-week sharpener (3×1 mile at goal pace) to keep the legs snappy. Textbook taper science.

Our plan validator rejected it. We had written hard rules months earlier: taper weeks can only get easier — no added volume, no added quality sessions. Sensible-sounding guardrails against an AI doing something reckless before a race.

Here's the part worth writing down: Claude didn't argue. It quietly redesigned its proposal into something “validator legal” — a watered-down taper with the best sessions stripped out — filed that instead, and only mentioned the compromise in passing. The safety system hadn't prevented a bad recommendation. It had prevented us from ever seeing the good one.

Warnings, not walls

That was the design lesson. We already had the real safety mechanism: a human approving every change with a clear diff. Hard validator blocks on coaching judgment were a second, paternalistic layer that mostly taught the AI to self-censor. So we rebuilt them: structural rules still hard-block (you can't edit a week that already happened), but coaching doctrine — taper conservatism, ramp-rate caps, even injury cautions — became warnings that travel with the proposal. The athlete sees “this adds quality work to a taper week — riskier this close to a race” in amber, right on the approval screen, and decides with eyes open. The AI is told its recommendation carries warnings, so it explains the trade-off instead of hiding it.

Then the engine learned it too

The re-filed proposal — the honest one, tempo and all — made us look hard at our own plan generator. It turned out Navara's tapers were cutting volume and intensity, while the research consensus (Mujika's work, the Bosquet meta-analysis) says an effective taper cuts volume 40–60% and keeps intensity. The AI wasn't being reckless. It was being right.

So the change went deeper than the guardrails: every taper week Navara generates now keeps a small quality touch, and race week gets a goal-pace sharpener by default. A recommendation that started as one AI's proposal for one athlete's race is now how the engine plans tapers for everyone — with a test that locks the doctrine in.

What we took from it

If you're building AI-assisted products: hard blocks on judgment calls don't make the AI safer, they make it quieter — and quiet AI fails in ways you never get to review. Put the human decision where it belongs, make the risks visible at that moment, and let the disagreement happen in the open. Sometimes the AI is wrong and the warning does its job. And sometimes the AI is right, and your product gets better in a way no roadmap predicted.