Explainability layer

AI writes the plan.
Guardamos shows what it was checked against.

Guardamos checks AI-generated training and rehab plans against published sports-medicine guidelines, and returns the specific facts and sources behind every verdict. Pass or flagged, you get a record of what was actually verified — not a silent OK.

Before / after

Same request, same context. One of them leaves a record.

A general-purpose AI assistant was asked for a training plan after six months off. It jumped straight into a high-intensity program, with no adjustment for returning after a long layoff. Guardamos returns which rule that breaks, and where it comes from.

chatgpt_6month_layoff.json rejected
"matched_rules": ["L1-RTT-0001", "L1-RTT-0002a", "L1-RTT-0002d"], "explanations": [ "Requires the dedicated 'return from long inactivity' track — a 4-week progressive reduction (Table 9). Not applied here.", "Table 9 (p.16) recommends 1-2 sets per exercise in week 1. Plan's value (4) exceeds it." ] // checked against ACSM 2026 + CSCCa/NSCA 2019
chatgpt_6month_layoff_corrected.json pass
"matched_rules": [], "checked_facts": [ { "rule_id": "L1-RTT-0001", ... }, { "rule_id": "L1-RTT-0002a", ... } ], "layer3_response": "...follows the return-from-long-inactivity track... week 1 parameters are consistent with guidance." // pass still ships with reasoning — not a silent OK
Architecture

Three layers. One of them is the actual moat.

Layer 1

Deterministic rules

Extracted from published guidelines, evaluated in code. No LLM involved in the judgment. This is what's open source.

open source, live
Layer 2

Expert tacit knowledge

Evaluates whether a plan reflects the contextual judgment a human expert would apply — beyond what a literature threshold alone can capture.

in development
Layer 3

Grounded explanation

Once a plan clears Layers 1–2, a short natural-language summary is generated — strictly from the confirmed facts, never from re-reading the plan freely.

live
Grounded in the literature

Every rule cites where it came from.

No rule is treated as reliable until it has been checked, line by line, against the source it claims to represent.

2026
Currier et al. American College of Sports Medicine Position Stand: Resistance Training Prescription for Muscle Function, Hypertrophy, and Physical Performance in Healthy Adults. Med Sci Sports Exerc. 58(4):851–872.
10 rules · verified
2019
Caterisano et al. CSCCa and NSCA Joint Consensus Guidelines for Transition Periods: Safe Return to Training Following Inactivity. Strength Cond J. 41(3):1–23.
17 rules · verified
2013
Meeusen et al. Prevention, Diagnosis, and Treatment of the Overtraining Syndrome. Med Sci Sports Exerc. 45(1):186–205.
2 rules · verified
2019
Fragala et al. Resistance Training for Older Adults: Position Statement From the National Strength and Conditioning Association. J Strength Cond Res. 33(8):2019–2052.
16 rules · verified
Why the record matters

“Nothing was flagged” is not evidence of anything.

Under the EU AI Act, providers of high-risk AI systems are responsible for documenting how their system meets the requirements — and for most categories, that assessment is done internally rather than by an external body. That places the burden of producing credible technical evidence on the provider. An absence of complaints is not that evidence. A record of what was checked, against which published source, is.

Guardamos does not certify anything and is not a notified body. What it does is generate the underlying material: for every plan, the specific facts that were verified and the guideline they were checked against — on passes as well as failures. Whether that material ends up in an internal quality process, a technical file, or a customer’s due-diligence request is up to you.

High-risk obligations under Annex III apply from 2 December 2027; systems embedded in regulated products under Annex I follow on 2 August 2028. Whether a given fitness or rehab product falls in scope depends on its stated intended purpose — general wellness tools generally do not.

Access

Free to self-host. Simple to buy.

Self-hosted

$0
  • Full rule engine, MIT licensed
  • All published rules, same ones running in production
  • Your own OpenAI API key
  • You run the updates, uptime, and maintenance
Clone the repo

Hosted API

$149/month · 14-day free trial
  • One API key, no infrastructure to run
  • Rules kept current as guidelines update
  • Uptime and support included
  • Early access to Layer 2 as it ships
Start free trial