AI never does the math.
Here's the architecture that guarantees it.

Three kinds of truth

Everything Flexbound shows you belongs to one of three truth systems, and each has different rules:

• User truth. What you wrote, edited, and corrected. Your raw journal text is never mutated. Parsing creates separate records underneath it. Ten years from now, the words are still yours.
• Nutrition truth. Source-backed food data. Every estimate shows its source, its serving assumption, and a confidence grade. Your corrections outrank every database.
• Methodology truth. Training and nutrition rules stored as versioned data, never as prompt text. Each rule carries its evidence level and the papers behind it.

Rules are data, not vibes

The methodology engine currently ships 44 approved rules across progression, hypertrophy, strength, volume, recovery, deloading, protein, cardio, split selection, time-constrained training, home exercise selection, and injury caution. Each one cites sources from a library of 25+ peer-reviewed papers, position stands, and professional references. Rules have version numbers, evidence grades such as strong, moderate, and coach opinion. The labels are plain, and the parameters are machine-readable.

When the app suggests adding 5 lb to your bench, that suggestion references the specific rule that produced it. Tap through and you can read the rule and the research it stands on. Nothing is buried in a prompt.

The decision contract

Every authoritative decision in Flexbound, whether it is a calorie target, progression call, or deload suggestion, must identify:

1. The versioned deterministic engine that produced it
2. The user inputs and sources it used
3. The exact outputs allowed to affect app state
4. Its assumptions and guardrails
5. The methodology rule IDs behind prescriptive advice
6. A source trace and a deterministic confidence grade
7. A durable audit record of the whole decision

Incomplete decisions are rejected by the type system before they reach you. That's not a policy we promise to follow; it's a contract the code can't compile without.

What the AI is allowed to do

Four things. Only four:

• Rephrase a weekly review the calculators already finished
• Explain a workout plan without changing a single number in it
• Draft a parse of ambiguous journal language for you to confirm
• Summarize research for human methodology review; it cannot approve rules

And the enforcement has teeth: if AI-written prose introduces a number that isn't in the locked deterministic facts, the app discards the prose and uses its own template instead. Model output never becomes calculation input. Ever.

The planner's hard constraints

Workout plans are where most apps quietly hand the wheel to a model. Flexbound's planner is a pipeline of deterministic services, each citing methodology rules into the plan's source trace:

• Equipment is a filter, never a preference. You declare where you train and what you own; every exercise selection, volume fill, and substitute must satisfy those requirements. An anchored band movement will not appear unless you declared a secure anchor. A plan that fails its own equipment audit is rejected, not shipped.
• Session length is a budget. The declared minutes cap exercises and rest with a fixed trim order: isolation accessories go first, accessory sets second, rest scales third, and the day's priority compounds are protected to the last.
• Split structure is selected, not defaulted. Full-body, upper/lower, or a push/pull/legs hybrid is chosen from your training days, experience, session length, and goal, with the frequency evidence treated as a volume vehicle rather than magic.
• Effort ships with the plan. Every set carries a reps-in-reserve prescription from the effort methodology, and swapping an exercise preserves sets, reps, rest, and RIR exactly.
• Cardio is structured, not sprinkled. Plans carry a weekly easy-aerobic prescription grounded in the public-health minimum and the concurrent-training literature, placed so it does not sabotage the lifting it accompanies.

The engines, in one table

Honest about the edges

Published methodology also means published limits. Estimated 1RM is a display heuristic, most reliable at low rep counts. Nutrition estimates carry confidence grades because serving assumptions are genuinely uncertain. Deload thresholds are gradedcoach opinion, not settled science. The literature gives principles, not exact trigger points. Where evidence is thin, the app lowers confidence or withholds the recommendation rather than pretending.

Read the rules yourself

Each core engine is documented in the science library, including its formula, evidence, and limits:

• Progressive overload & double progression
• Estimated 1RM: the Epley formula and its limits
• Weekly training volume & the recovery guardrail
• Adaptive TDEE from your own logs
• Protein targets that match the evidence
• RPE and RIR: measuring effort you can log
• Deload signals, pain flags & minimum-dose weeks