monarch-exercise-economy-audit

Use when adding, renaming or re-tiering a movement in the Ironvellum (formerly Monarch, D:/Monarch) exercise catalogue, changing an XP constant there, or when a movement scores wrongly: holds or attem

SKILL.md

GitHub

Ironvellum exercise economy audit

Project D:/Monarch (Ironvellum, formerly Monarch). The catalogue is hand-edited data, so it fails silently (generic invariants: hand-edited-catalogue-invariants). This skill holds the app's scoring model and the traps it has hit. The change procedure is score-economy-rework-safety; a whole-economy payout audit is game-economy-ground-truth-audit.

The model (read these five, they are small)

file role
data/Seed.kt allExercises = baseExercises + every Skills.ALL movement not already there + activities. withMetric() stamps HOLD from the map — a hold must be recognised, never given an explicit metric in its row
domain/MovementDifficulty.kt tier: loadPricedTiers → skill tree → catalogueTiers → DEFAULT_TIER = 2. intensity() = 2^(tier-1)
domain/Xp.kt effortUnits = taperedVolume(reps or seconds/5) * intensity * loadMultiplier * modifierFactor
domain/Skills.kt tree data; metric and target are inferred from the standard's prose
domain/StrengthIndex.kt body-scaled; applies XP's taperedVolume and a sqrt-damped MovementDifficulty.strengthWeight, so cheap volume cannot farm the leaderboard's lifetime sum

val ExerciseMetric.isStrength (Models.kt:26-27) is the existing REPS/HOLD vs activity predicate. Reuse it; never write a second one.

Pricing rule for any new movement (settles most tier arguments)

A tier states UNLOADED difficulty only. Xp.loadMultiplier already pays for the kilos, so a tier that prices the load pays for the plate twice (the load double-count trap in score-economy-rework-safety). loadPricedTiers scores loaded names as their unloaded parent: weighted pull-up and weighted dip at 3, and rungs II-V of the Squat, Bench, Press and Deadlift ladders at 1, their line root. Back Squat, Bench Press, Overhead Press and Deadlift are tier I skill-tree rows; isolation is tier 1 (bicep curl, wrist curl).

Leaving a new movement unclassified is not neutral: it falls to DEFAULT_TIER = 2, intensity 2, double a tier-I barbell root. Classify it explicitly; isClassified backs the test that holds the seeded catalogue to full coverage. (verified 2026-09-24)

Audit axes that actually find bugs

  1. Reverse classification direction. The forward test ("every seeded movement is classified") cannot see a key that names nothing. "plank" was priced tier 1 and marked a hold for months while only Weighted Plank was seeded. Walk the public catalogueOnlyKeys accessor back to the catalogue; keep the maps private rather than reflecting or widening them.
  2. Forward test's own blind spot: it filtered metric == REPS.name, so an unclassified HOLD passed. Filter on isStrength.
  3. Prose-inferred metrics. Skills infers metric/target from the standard text. Minute-denominated standards broke it: "Hold 3 minutes" → SECONDS target 3; "10-minute routine" → REPS target 10. Fix the inference, not the strings. minute and metre are one regex slip apart — pin Handstand Walk ("10 metres unbroken") as METRES in the same test.
  4. Activity metrics scored as reps. An ATTEMPTS_GRADE set carries its attempt count in the reps column. SetRecords, ExerciseHistoryCalculator and SessionScreen's block-strength sum all branched only on hold-vs-not, so seven boulder attempts minted a seven-bodyweight-rep PR.
  5. Progression gates. The bypass was HOLD-only, so a DISTANCE_TIME preset entry was told to hit N reps and add 2.5 kg.

The regression that follows fix 4 — do not ship without it

SetRecords.delta returns isRecord = true when no record exists (SetRecords.kt:~120). Excluding activity sets from records therefore makes every climbing set read NEW PR forever. Plumb a scoresStrength flag to the set row and gate both the badge and the block strength line.

Two claims to distrust

  • "Every prerequisite tier is strictly below its dependent." False. A tier is a difficulty band holding ordered chains — Muscle-up (III) requires Pull-up (III); ~12 such pairs. The true invariant is never harder. Acyclicity already stops a band chain closing on itself.
  • "isWeighted contradicts the presets." It is UI-only — the picker's equipment facet and labels (ExercisePicker.kt:295, ExerciseExplorerScreen.kt:196); no scoring code reads it. A +10 kg pull-up is genuinely bodyweight-with-added-load. Not a defect.

Wiring a new row touches

  • Repository.ensureSeeded inserts by exact name every launch, so upgrades pick rows up; a lifter's same-named custom row is kept (merge, not duplicate), and classification is case-insensitive so it still gets the right tier.
  • Progression.weightStepKg is substring/muscleGroup heuristics, no per-name table — new movements inherit sensibly.
  • TitleEngine hard-codes category strings ("Cardio", "Water", "Sport"), never movement names. Adding is safe; renaming breaks skill badge, hold detection and tier, which all key on the display name.

Verification

Mutation-prove every invariant (break the real data once, confirm the matching test fires; commit first and revert as score-economy-rework-safety says). Then on the emulator — never the owner's phone, where logged sets bank XP that is never restated — with a bodyweight logged (without one the badge path is null and the check is vacuous), log an activity set and a lifting set in the same session. Expect the lifting one to show N STR and NEW PR and the activity one to show neither. Without that control, "no badge" only proves badges are broken.

Pin every Gradle run to the emulator from Git Bash: ANDROID_SERIAL=emulator-5554 ./gradlew.bat <task> (see monarch-session-context).