Fix skill-level scaling: unscaled effect kinds + levelScale/levelConst for derived values by Infernal-Crack-LED · Pull Request #148 · Infernal-Crack-LED/nikke-sim

Infernal-Crack-LED · 2026-08-25T11:26:48Z

What this fixes

A community report compared the Roster Generator against a live Shooting Range fight: nayuta at
skill levels 1/1/1 simulated ~158M against ~77M measured (sim/real ≈ 2.05), and lowering only her
Skill 2 level changed her simulated damage by exactly 0.0%.

src/skills/scale.ts scales an authored override value by matching it against index 9 (max level)
of that unit's blablalink per-level arrays and substituting index L−1. Three independent ways a
value stayed pinned at max level:

No case in scaleEffect's switch — the magnitude was returned untouched and emitted no
warning. weaponSwap.damagePct was the worst (a burst weapon mode's per-shot multiplier is
often the carrier's largest damage term; 11 units carry one) — exactly the reporter's
unexplained "burst level changed damage with no Modeling Note". Also shield.maxHpPct (21
values / 16 units), fillGauge.pct (can shift full-burst counts), stackedNuke, storedHit,
and perResource.mult — the LIVE magnitude of a perResource buff/DoT, covering 8 more carriers.
findArr took the first index-9 match, constant or varying. Slot tables hold constant
arrays (a "for 10 sec" duration is [10,10,…]) beside real magnitudes sharing the same max.
Now prefers a VARYING array (crust, prika).
The authored value is DERIVED, so it matches no entry — a fold (nayuta 530.46 = 150 +
380.46), a time-averaged stack ramp (14.4 from 15.2), a stack-cap product (42 = 1.4 × 30). Fixed
per-value by the new levelScale annotation, with levelConst for verified-invariant
fields.

Measured effect (1/1/1 vs 10/10/10)

unit	before	after
`nayuta`	-15.3%	-44.3%
`mast-romantic-maid`	—	-59.0%
`snow-white-heavy-arms`	-34.4%	-53.0%
`neon-vision-eye`	-49.3%	-48.7%
`little-mermaid`	-5.4%	-22.8%

Census: 0 SILENT + 84 WARNED (was 3 + 124), 47 → 30 units.

Blast radius: ZERO on the graded board

No magnitude changed — only how each behaves below level 10. Every graded comp runs 10/10/10, where
scaleBlocks early-returns, so the regression snapshot is unchanged and was never regenerated.
verify.sh green (5041 tests).

Guards

An anchor that is not a real max-level entry is a structural error, not a silent fallback.
The kind→scalable-fields map is owned and exported by scale.ts and imported by the
validator — never mirrored. A private copy in the census went stale within one session and
reported fixed values as still broken; a test now probes the real scaler against the map in both
directions.
scripts/audit-skill-scaling.ts (census) runs the real scaler and observes what changed
rather than mirroring its switch. --sim <slug> sizes per-slot sensitivity.
scripts/apply-level-scale.ts is the declarative, idempotent applier; every row carries the
note/kit quote it rests on. The census's ← 60 × 12 hint is a brute-force table search that
finds coincidences — eve's 720% is really 240 × 3 per her note, and
snow-white-heavy-arms's 1055.9 is 105.59 × 10 exactly (the hinted 42.24 × 25 = 1056.0).

Review

Cross-family (kimi-code/k3), two rounds — packets and verdicts in
scratchpad/gates/2026-08-25-skill-level-scaling/.

Round 1: FIX-BEFORE-MERGE — 1 BLOCKER (perResource.mult, which also meant this branch's
published mihara-bonding-chain figure was understated), 3 FIX, 3 NOTE. All confirmed and fixed.
Round 2: CLEAN — its 1 FOLLOW-UP + 2 NOTEs were also fixed rather than filed.

Merge with `origin/main` (PR #147)

Both conflicts resolved to upstream, and the changes converge: mihara-bonding-chain's
500.6 dump is now ten paced blocks of a literal 50.06 (native table match) and her burst DoT is
perResource {mirror, 50.05} — covered by the perResource.mult fix. Her two apply-table rows were
removed; she scales with no annotation.

Deliberately NOT enacted (queued)

19 TREASURE units cannot scale at all — a DATA gap. data/skill-levels.json holds the BASE
kit arrays while those overrides model the treasure kit. drake: table 11.85/1254/98.55 vs
override 20.09/3009.6/201.6; boosts are non-uniform (×1.70, ×2.40, one line unboosted), and
blablalink roledata has no favorite-item fields. Annotating off a base-kit anchor would be a
guess about favorite-item scaling. Needs a source or an owner ruling.
Cross-slot anchors — ein skill1 363.24 = 4 × 90.81 with 90.81 in her skill2 table
(also eve, neon-vision-eye, emma-tactical-upgrade, red-hood). May indicate those blocks
are filed under the wrong slot.

⚠ Patch notes NOT included — needs your approval

This changes player-visible numbers, so it wants a Patch Notes entry, but publishing is owner-gated.
Draft is in the PR discussion; say the word and I'll add it to this branch before merge.

🤖 Generated with Claude Code

A live Shooting Range comparison (2026-08-25) reported nayuta at 1/1/1 simulating ~158M against ~77M measured. Two silent scaler failures, both confirmed by the new census tool and reproduced in-sim: 1. SILENT — an effect kind with no case in scaleEffect returned its magnitude untouched AND emitted no warning. Adds weaponSwap (damagePct/chargeMultPct), fillGauge.pct, shield.maxHpPct, stackedNuke.atkPct/hpPct, storedHit.atkPct. weaponSwap.damagePct is the worst of these: a burst weapon mode's per-shot multiplier is often the carrier's largest damage term, and 11 units carry one. damagePct is deliberately NOT scaled on a sameWeapon swap — there the gun is not replaced, so the value is the base weapon's own normalAttackMultiplier (a weapon stat, level-invariant). No unit ships that combination today. 2. WARNED — a DERIVED authored value (a fold of two kit lines, a time-averaged stack ramp, a stack-cap product) matches no level-table entry, so it stayed at max level while emitting "no level table match for N". New optional `levelScale` names the table anchor(s) the number was derived from; the value scales as authored x (sum anchors@L)/(sum anchors@10) — exact for a fold, proportional otherwise. validate-structural rejects an anchor that is not a real max-level entry, so a typo is a build error, not a silent max-level value. nayuta annotated from the derivations documented verbatim in her own override note. Measured in-sim (audit-skill-scaling.ts --sim, liter/crown/nayuta/noise/red-hood): skill1@1 skill2@1 burst@1 1/1/1 before -8.7% 0.0% -7.3% -15.3% after -15.9% -9.6% -24.7% -44.3% skill2 previously moved her damage by exactly 0.0% — all four of its values were derived. Board blast radius is ZERO: every graded comp runs at 10/10/10, where the scaler early-returns, so the regression snapshot is unchanged and verify.sh is green without --update. Tooling: scripts/audit-skill-scaling.ts is the committed census (SILENT/WARNED per unit, --sim to size a unit's per-slot sensitivity). Remaining roster-wide backlog is reported there and left for a batched owner decision. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…, findings doc Follow-ups on the scaler fix: - scale.ts no longer scales weaponSwap.chargeMultPct. All 8 carriers author a round kit constant ("Full Charge Damage: 250% of damage" — 250/300/1750) and not one resolves to a level-table entry, so scaling it only emitted warnings nobody could act on. weaponSwap.damagePct (the real fix) is unaffected. - audit-skill-scaling.ts now RUNS scaleBlocks and observes which fields changed, instead of hand-mirroring scaleEffect's switch. The mirror had already drifted within this session — it still reported all 37 SILENT values as broken after they were fixed. Also anchors the warning match on the full "match for N —" phrase; a bare digit-substring test classified a maxAmmo of 1 as WARNED. - Census after both fixes: 37 -> 3 SILENT, 48 units with derived values remaining. - docs/handoffs/2026-08-25-skill-level-scaling.md: root cause, what landed, and the per-unit backlog sized by measured impact (--sim) rather than warning count. drake is the worst — 6 derived values covering essentially her whole kit, so her 1/1/1 damage reads +0.7% vs 10/10/10 (levels effectively inert). QUEUE.md item added for the batched owner decision. verify.sh green; board artifacts rebuilt (they hash inputs and nayuta.json changed); regression snapshot unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…reasure gap Works the levelScale backlog per unit, each anchor taken from that override's OWN documented derivation rather than a numeric search. Backlog 124 -> 97 values, 47 -> 42 units. Measured level sensitivity at 1/1/1 (was near-flat for several): mast-romantic-maid -59.0% snow-white-heavy-arms -53.0% eve -37.3% guillotine-winter-slayer -34.3% sakura-bloom-in-summer -30.6% ein -29.9% mihara-bonding-chain -23.4% little-mermaid -5.4% -> -22.8% neon-vision-eye -48.7% New `levelConst`: fields VERIFIED level-invariant, so they stop warning. Three shapes — a structural constant (eve's Mk2 "doubles S1" as sequentialMultPct +100), a sentinel (prika's burstCdr -9999 lockout), a non-skill quantity (red-hood's burstCdr 40 = her own burst cooldown). Distinct from having no annotation, which still warns; levelConst is a claim someone checked. scripts/apply-level-scale.ts is the declarative, idempotent applier (--check to dry-run); every row carries the note quote it rests on. That matters twice over: the audit's "<- 60 x 12" hint is a brute-force table search and it finds coincidences. eve's 720% is really 240 x 3 (her note), not 60 x 12; and snow-white-heavy-arms's 1055.9 is 105.59 x 10 exactly, not the hinted 42.24 x 25 (= 1056.0, off by 0.1). TWO FINDINGS, both recorded in the handoff doc: (a) 19 of the 47 backlog units are TREASURE units and are a DIFFERENT bug, blocked on data. skill-levels.json holds the BASE kit arrays; those overrides model the treasure kit. drake: table has 11.85/1254/98.55, override authors 20.09/3009.6/201.6 — boosts are non-uniform (x1.70, x2.40, and maxAmmoPct 72.18 not boosted), so no proportional anchor recovers them, and x2.40 is far too large to be extrapolated skill levels. Blablalink roledata carries no favorite-item fields at all. Annotating these off a base-kit anchor would be a guess about favorite-item scaling — deliberately NOT done. (b) An anchor can live in a DIFFERENT slot's table: ein's skill1 363.24 is 4 x 90.81 but 90.81 is in her SKILL2 table. Not expressible, and would scale off the wrong slot's level anyway. Left WARNING. The structural validator caught it. verify.sh green; board artifacts rebuilt; regression snapshot unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…investigation Second annotation pass, driven from PRIMARY kit text (data/characters.json skills) rather than override notes where the notes were silent. Backlog 97 -> 83 values, 42 -> 30 units; 21 actionable units -> 11 residual. Landed this pass — every row a kit-text or note quote: arcana-fortune-mate 39 = 13 x 3 stacks; 7.47 = 2.49 x 3; 30 -> levelConst asuka-wille 198.6 = 6.62 x 30 Anti A.T. Field stacks cinderella 13659.2 = 1365.92 x 10; 346.8 = 28.9 x 12 Beautiful stacks mast 226 = 4.52 x 50 Sea Breeze stacks maiden-ice-rose 137.28 = 1372.8 x 10% (the kit's Max-HP component) soline-frost-ticket 20 = 10 x 2 tickets (NOT the hinted 10+10 fold) emilia 12.06 = 2.01 x 6 | liberalio 202.5 = 40.5 x 5 | mana 18 <- 0.18s ada 61.3 + rapi-red-hood 20 -> levelConst Two more hint-vs-source catches: soline-frost-ticket's 20 is 10 x 2 tickets per the kit, not the hinted 10+10; and arcana-fortune-mate's 30 = 10 x 3 is levelConst, not levelScale, because its 10 anchor is a CONSTANT array in her skill1 table. The 13 residual non-treasure values are NOT simple derivations and are left warning: six are CROSS-SLOT (anchor lives in another slot's table — ein 363.24 = 4 x 90.81 with 90.81 in skill2; eve, neon-vision-eye x2, emma-tactical-upgrade, red-hood), which may mean those blocks are filed under the wrong slot; seven are datamined or kit-literal values that appear in no level array at all. QUEUE.md: replaced the old batched-decision item with two scoped ones — the treasure investigation (with the drake evidence and the two concrete questions: is there another source, or is there a RULE an owner can state) and the 13 residuals. verify.sh green; board artifacts rebuilt; regression snapshot unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The applier wrote overrides with JSON.stringify(ov, null, 2), which expands every compact one-line object in the file. Prettier's default objectWrap:"preserve" then KEEPS them expanded, so a 1-line annotation landed as ~100 lines of pure formatting churn on protected, concurrently-edited files — mana/mast/sakura-bloom-in-summer were 124/96/117 changed lines for a 1-3 line change each. Rewritten to insert the annotation at the target effect's own byte span, leaving the rest of the file untouched. `prettier --object-wrap collapse` was evaluated and rejected: it is a no-op on mana/mast/sakura but rewrites emilia and nayuta, which are legitimately multi-line in main. Net across all 23 annotated overrides: 442 insertions / 118 deletions -> 59 / 2 (the 2 are objects that genuinely crossed the 80-col wrap once annotated). Verified every touched file is deep-equal to its main version after stripping levelScale/levelConst — 23 files checked, 0 semantic diffs. verify.sh green; census unchanged at 83 warnings; snapshot unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…r, applier Cross-family review returned FIX-BEFORE-MERGE with 1 BLOCKER / 3 FIX / 3 NOTE. All seven confirmed and fixed. BLOCKER — perResource.mult was still SILENT, and the census could not see it. A perResource buff/DoT IGNORES its static value/atkPct at runtime and recomputes the magnitude as resources[name] x mult, so mult is the real level-scaled number. Eight carriers' primary magnitudes stayed pinned at max level with no warning: mihara-bonding-chain 25.08 (her Ensnaring DoT driver), mana 70.4, phantom 25.75/12.86, marciana-marine-study 32.73, e-h 7.5, guillotine 0.96, soda-twinkling-bunny 1.32 — each a max-level entry of a VARYING array. The census was structurally blind to it (nested under perResource; `mult` also sat in TIME_FIELDS), so its own completeness summary was wrong. Consequence I had published: this branch's level-sensitivity figure for mihara-bonding-chain was materially understated — -23.4% -> -35.2% at 1/1/1. Doc corrected. FIX — findArr took the FIRST max-level match, constant or not. A slot's table holds constant arrays (a "for 10 sec" duration) beside real magnitudes sharing the same max; when the constant came first the magnitude silently stayed at max. Now prefers a varying array. Reviewer confirmed the live cases against kit text: crust's burst "Sustained Damage ▲10%" and prika's burst "Charge Damage ▲25%". Census SILENT 3 -> 0. FIX — apply-level-scale.ts claimed text insertion but still re-serialized. The landed overrides had been normalized out-of-band, so the committed tool did not do what its header and prior commit message claimed; anyone adding a row would have reproduced the churn. Byte-span insertion is now actually implemented (verified idempotent: --check reports 0 applied / 41 already present). FIX — levelScale:null crashed the validator (typeof null === 'object' fell through to Object.entries(null)), taking the whole gate down with a stack trace. Guarded. NOTE x3 — the type union under-declared levelScale/levelConst (now on every kind scaleEffect reads: burstCdr, fillGauge, shield, storedHit); the validator did not check that an annotated field is on a scaling path, nor that levelConst names a real numeric field (both now error, with dotted-path support for perResource.mult); the chargeMultPct comment's counts were wrong (~14 instances / ~10 units incl. one flatDamage carrier — substance held, reviewer re-verified none resolves to a table). Tests: +4 scaler (findArr varying-preference, perResource.mult on buff/dot, levelConst on a dotted path), +6 validator (null guard, anchor resolution, non-scalable field, levelConst typo). 4833 pass. verify.sh green; snapshot unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…an scanner Cross-family re-review returned CLEAN (kimi-code/k3, result-r2.json). Its 1 FOLLOW-UP and 2 NOTEs were fixed rather than filed, since all three were defects in code or claims this branch introduced. FOLLOW-UP — SCALABLE_FIELDS was a hand-maintained MIRROR of scaleEffect's switch, and its comment overclaimed: it said skill-level-scale.test.ts "pins the pairing" when nothing in the suite compared the mirror to the switch. That is the exact staleness class that bit the census in round 1 (its private SCALED map went stale within one session and reported fixed values as still broken). scale.ts now OWNS and exports the map; validate-structural.ts imports it; and a new test probes the REAL scaler against the manifest in both directions — every listed kind/field must actually scale, and a non-listed field must stay untouched. NOTE — objectSpans mis-keyed an object FOLLOWING a scalar in a mixed array (it counted container opens, not elements). Unreachable in today's corpus but a mis-keyed span inserts an annotation into the WRONG effect. Fixed by counting element separators, and pinned by a test that round-trips EVERY span in all 189 override files (slice -> parse -> deep-equal, 192 assertions) plus mixed-array and string-escape cases. Second guard added: two rows inserting a duplicate levelScale/levelConst key into one effect now error, because JSON.parse silently keeps the last and the post-write parse cannot see it. Self-inflicted, caught while fixing the above: exporting objectSpans meant importing the module executed its top-level apply loop — under vitest's argv that is WITHOUT --check, so it would have written to every annotated override. CLI is now behind an import.meta.url === argv[1] guard; verified the override diff is unchanged across a test run. NOTE — doc/QUEUE residual counts said 83; the live census says 84. Reconciled from the instrument's own output. verify.sh green (5040 tests, +207). Snapshot unchanged; census 0 SILENT + 84 WARNED. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

# Conflicts: # src/skills/overrides/mihara-bonding-chain.json

PR #147 reworked her while this branch was open. Both merge conflicts resolved to upstream because its rework replaced the exact structures this branch annotated — and the two changes CONVERGE rather than collide: - the 500.6 (10 x 50.06) dump is now ten paced blocks of a literal atkPct 50.06, a direct max-level entry of her skill1 table, so it scales natively; - the static burst DoT atkPct 1001 (20 x 50.05) is now perResource {mirror, 50.05}, and the permanent Ensnaring DoT is perResource {ensnaring, 25.08} — both covered by the perResource.mult case added for the round-1 review BLOCKER. Her values therefore scale after the merge with NO annotation; both apply-table rows were removed (they would hard-error on "no matching effect" — the staleness guard working). Sensitivity restated -35.2% -> -33.9% at 1/1/1, the -35.2% having been measured against the pre-merge model. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

DECISIONS (settled WHY): the three ways a value stayed pinned at max level (no case in scaleEffect, findArr taking a constant array over the varying one, and a DERIVED authored value matching no entry), the levelScale/levelConst primitives, the zero-blast-radius argument, the guards, and the two gaps recorded but deliberately not enacted (19 treasure units blocked on data; cross-slot anchors). STATE (current-state registry): levelScale/levelConst added to the opt-in kit-primitive inventory with their carrier lists. Patch notes drafted for owner approval, NOT committed — publishing is owner-gated. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

scripts/doc-drift.ts --update. Adding two effect-level primitives left engine-modeling-gaps.md's generated census stale, which verify.sh flags. My previous commit chained past that failure and pushed a red tree; this restores it green (108 primitives across 187 overrides). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Infernal-Crack-LED and others added 11 commits August 25, 2026 02:31

Merge remote-tracking branch 'origin/main' into skill-level-scaling

9f71880

# Conflicts: # src/skills/overrides/mihara-bonding-chain.json

Infernal-Crack-LED merged commit 718d679 into main Aug 25, 2026
1 check passed

Infernal-Crack-LED deleted the skill-level-scaling branch August 25, 2026 11:54

Navigation Menu

Repository navigation

Fix skill-level scaling: unscaled effect kinds + levelScale/levelConst for derived values - #148

Fix skill-level scaling: unscaled effect kinds + levelScale/levelConst for derived values#148
Infernal-Crack-LED merged 11 commits into
mainfrom
skill-level-scaling

Infernal-Crack-LED commented Aug 25, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Infernal-Crack-LED commented Aug 25, 2026

What this fixes

Measured effect (1/1/1 vs 10/10/10)

Blast radius: ZERO on the graded board

Guards

Review

Merge with origin/main (PR #147)

Deliberately NOT enacted (queued)

⚠ Patch notes NOT included — needs your approval

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Merge with `origin/main` (PR #147)