Training plans should be controllers, not calendars.

01 · Overview

A traditional plan is written once, on day zero, for an athlete who never misses a run, never gets sick and absorbs every session as intended. Within days reality drifts from it: a missed session is dropped or crammed into the weekend, and the calendar keeps climbing towards a peak the athlete is no longer on course for. Roen closes the loop. The plan is re-solved from what you actually ran.

• What the engine decides. Everything with a number on it: how far each week goes, which day carries the long run and the quality sessions, every pace and heart-rate ceiling, when to cut back, how a missed run or a sore calf changes the week, the race-day pace and the goal readiness score.
• What the AI does. It reads what you write: a debrief note, a request to move a session, a description of your running. It turns the sentence into a change, and the engine then checks that change against the same rules as everything else and may refuse it. If the language model is unavailable, a plain parser reads the sentence instead.
• What the AI does not do. It never chooses a session, a distance, a pace or a score. When it explains a plan, it is handed the engine's figures and told to quote nothing else.
• Same data, same plan. The engine runs on fixed rules with no randomness: the same runs and the same goal always give the same plan. A plan changes only when the data does, so every change can be traced to a run, a debrief or an edit.

Each section below gives the rule in plain words, the number the engine uses and its source, then a worked example. The examples follow one athlete: a marathoner with a 4:16 /km threshold pace, 50 km a week, 6 running days and 16 weeks to race day.

02 · Load and fitness

Every run is scored for load: its duration times how hard it was against your own threshold, so one hour at threshold scores 100. The scores feed a fitness and fatigue model. Fitness builds slowly and fades slowly, fatigue rises and falls within days, and their difference is how fresh you are. The plan leads each day: tomorrow's session is the plan's own, eased to an easy run only when the balance shows real overreach or another hard session sits within 24 hours. It is never made harder.

The rule

Fitness is the long average of daily load, fatigue the short one. A planned session is eased, never raised, when you are deep in fatigue.

The numbers

Fatigue averages over 7 days, fitness over 42. A session is eased below a balance of −30.

Source

Banister 1975; Banister 1991.

A missed day is measured, not made up. Missed easy kilometres go to the easy days that have room, within the week's own total; the week never grows. A skipped long run is not added to other days, and a run cut short is not chased. A move is refused when it would put two hard days back to back or quality on the day before or after the long run. Gaps of up to 4 weeks are understood. After an illness the first 2 runs back are easy, and after 10 days or more the engine offers to rebuild the plan and re-checks the goal.

The rule

Time off is read as time off: fitness decays, and the plan is rebuilt from the new level rather than resumed from the old one.

The numbers

Gaps up to 4 weeks; 2 easy runs after an illness; a rebuild offered after 10 days.

Source

Coyle 1984; Mujika and Padilla 2000.

03 · Perceived effort

A run without heart rate or a usable pace still counts. The debrief asks how hard it felt on a 1 to 10 scale, and the engine turns the rating into an intensity on the same threshold scale as every other run. A rating never replaces a heart-rate or pace score, so a run is never counted twice.

The rule

Session load = minutes × intensity², with the intensity read from the rating you give.

The numbers

A rating of 7 is threshold: one hour scores 100. A rating of 3 is an intensity of 0.7.

Source

Foster 2001.

04 · Hills

Climbing costs more energy than flat running and descending saves some, up to a point. The engine prices every run, every personal best and every race course with one cost curve for running on a slope, so a hilly easy run is not mistaken for a fast one and a hilly 10k counts for what it was worth. Steep descents earn no extra credit: past the curve's cheapest point, braking costs as much as it saves.

The rule

Effort on a slope is priced by the energy cost of running at that gradient, measured against the flat.

The numbers

A 4% climb costs 24% more than the flat, 8% costs 51% more. A descent of 8% saves 35%, and anything steeper is priced as 8%.

Source

Minetti 2002.

05 · Paces and heart rate

Every training pace is a fixed multiple of your threshold pace, calibrated against Daniels' tables, so the paces move together when the threshold does. Easy, marathon-pace and tempo work also carry a heart-rate ceiling from your heart-rate reserve. Most of the week stays easy: the hard work is held to a small share of the week's distance and time.

The rule

Paces come from threshold, heart-rate ceilings from heart-rate reserve, and most of the week is easy.

The numbers

Easy runs stay under 72% of heart-rate reserve, marathon pace under 80%, tempo under 88%. Quality work is at most 18% of the week's distance and hard time at most 10% of its duration.

Source

Daniels and Gilbert 1979; Karvonen 1957; Tanaka 2001; Seiler and Kjerland 2006; Seiler 2010; Helgerud 2007; Haugen 2022; Knopp 2024.

06 · Weekly progression and injury risk

Volume climbs in small weekly steps with a lighter week at regular intervals, and the long run grows on its own, slower ladder. Large studies of runners link injuries to two things the engine caps: a weekly distance that rises too fast, and a single run much longer than any in the previous month. Both are held regardless of the goal.

The rule

Weekly volume climbs gradually with regular cutbacks; the long run never jumps far past the longest recent run.

The numbers

At most 8% more a week (10% in a marathon block); every 4th week a cutback of 25%, and at most 20% back up after it. The long run grows by at most 10% over the longest before it, is at most 33% of the week, and tops out at 180 minutes or 32–34 km.

Source

Nielsen 2014; Damsted 2019; Frandsen 2025; Buist 2008; Ramskov 2018.

Each week also shows a load-ramp indicator, read from the last week's load against the last six weeks'. It is shown as low, moderate or high and never as a percentage, because it is not an injury probability and its coefficients are not fitted to injury data. The ratio comes from team-sport research (Gabbett 2016; Williams 2017), and in runners the evidence for it is weak: large cohorts found no rise in injuries with a higher ratio, or even fewer (Nakaoka 2021; Frandsen 2025), and its predictive value is debated (Impellizzeri 2020). The caps above do the work; the indicator only shows how steep a week is.

07 · Pain monitoring

When a debrief or a message mentions pain, the engine asks how bad it is rather than guessing. A niggle changes nothing; the same site again within 10 days gets a note. Sore or sharp opens a load-shielded window: the next 3 days become short, flat easy runs, and the kilometres removed are not moved elsewhere, since moving them would bring back the load the window exists to remove. Each morning a pain check decides the day, and a failed check turns it into rest.

The rule

Keep running while pain stays low during the run and settles by the next morning; stop when it does not.

The numbers

3 shielded days of at most 35 minutes, at most 50 m of climb and 50 m of descent, no strides. The check: no more than 3/10 during the run, back to baseline by the next morning; 5 × 45 s isometric holds. Still a problem at the end: an active injury, intervals out, 3 more days.

Source

Silbernagel 2007.

08 · Race projection

Your recent races and threshold efforts give a 10k speed. The engine scales it to the race distance with a power law: the longer the race, the more speed you give up. Below the marathon the exponent is fixed. In the marathon it depends on ability, since slower runners lose more over the distance, and a recent long run of 26 km or more earns part of it back, all of it from 34 km.

The rule

Race time = 10k time × (race distance ÷ 10 km) to a fixed power, with the marathon's power set by ability and recent long runs.

The numbers

An exponent of 1.06 below the marathon. In the marathon, 0.06 above 1 for the fastest runners and more for slower ones; a long run credited from 26 km, in full at 34 km.

Source

Riegel 1981.

09 · Durability and goal readiness

Durability is the ability to hold output deep into a long effort. Late in a race, low-volume runners slow more than their short-distance times suggest, so the engine takes a fade off the projected speed that shrinks as weekly volume rises. The distance also sets a minimum peak week and long run, and the goal's readiness score checks both against what the plan can deliver in the weeks available.

The rule

The marathon fade shrinks with volume; the race needs a minimum peak week and long run, reached at a safe rate.

The numbers

A marathon asks for a 48 km peak week and a 29 km long run, reached at no more than 12% a week. The fade is 8.2% of speed at 30 km a week and 3.4% at 70.

Source

Maunder 2021; Fokkema 2020; Doherty 2020; Nielsen 2014.

Goal readiness is two scores on one 0 to 100 scale: Pace, whether the goal time is within reach on race day, and Durability, whether the plan reaches the volume and long runs the distance needs. The score shown is the weaker of the two, so the number, the verdict and the limiting score always agree:

• On track. 70 or more.
• A stretch. From 40 to 70: within reach, not yet on track.
• Out of reach. Below 40.

Goal readiness is a model estimate. It is not a guarantee, and not the probability of finishing in the goal time.

10 · Taper and after the race

Before the race the volume comes down and the intensity stays: cutting the quality as well costs fitness the taper exists to keep. After a marathon the engine does not hand you a new plan the next day. A short return to running leads into an optional maintenance block that holds fitness at lower volume until the next goal.

The rule

Taper by cutting volume, not intensity. After the race, rest, then easy running, then quality, then a lighter maintenance block.

The numbers

A 3-week marathon taper. After a marathon: 2 days of rest, strides from day 10, quality from day 15, the next block from day 21. Maintenance runs 12 weeks at 65% of the block's last four weeks before the taper, the long run at 70% of the block's longest.

Source

Mujika and Padilla 2003; Bosquet 2007; Smyth and Lawlor 2021; Warhol 1985; Martinez-Navarro 2021; Hickson 1981; Hickson 1982; Hickson 1985; Houmard 1990; Spiering 2021.