feat(theory): noise-tolerant loop detection rewrite (task L-30)

detectRepeatingProgression redesigned: consecutive-dup collapse,
weak-period candidate rejection (kills self-overlap ghosts
structurally), lengths 2-8, <=1 substitution/insertion per cycle with
a >=2-exact-occurrences evidence gate, recency-weighted linear
coverage scoring. All 17 C-30 fixtures pass plain (smoke 817/817,
expectedFail markers removed). Critic PASS: 15 novel generalization
probes correct or data-faithful, 0.2-0.7 ms/call, diff confined to the
repeating-progression section, fixture contract untouched.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
vadimwit
2026-07-10 09:10:28 +01:00
parent b51daa1339
commit 4cc0009983
2 changed files with 106 additions and 65 deletions
+8 -40
View File
@@ -12,11 +12,12 @@
// · fixture with expectedFail that PASSES → smoke FAILS: stale marker, flip it
// L-30's definition of done = every expectedFail marker removed, all green.
//
// ─── FAILURE MAP OF THE CURRENT ALGORITHM ────────────────────────────────────
// (theory.js `detectRepeatingProgression`, every fixture RUN against it 2026-07-10;
// the `today:` comment on each expectedFail fixture is the actual observed output)
// ─── FAILURE MAP OF THE PRE-L-30 ALGORITHM (historical — all fixed by L-30) ──
// (theory.js `detectRepeatingProgression` as of C-30, every fixture RUN against
// it 2026-07-10; L-30 redesigned the function and removed every expectedFail
// marker — all 17 fixtures now pass as plain regression guards)
//
// The current matcher: window = last 32 commits; candidate lengths 26; EXACT
// The pre-L-30 matcher: window = last 32 commits; candidate lengths 26; EXACT
// contiguous occurrence counting (scan advances by len on match, by 1 on miss);
// requires reps ≥ 2; score = reps × len²; returns the canonical (lexicographically
// smallest) rotation of the best candidate.
@@ -130,87 +131,54 @@ export const LOOP_FIXTURES = [
expect: null,
},
// ── Expected failures of the current algorithm — the L-30 contract ─────────
// ── Failure modes of the pre-L-30 algorithm — fixed by L-30 ────────────────
{
id: 'vamp-2x4',
description: 'clean 2-chord vamp played 4× — must still report the pair, not a self-overlap',
history: reps(['Am', 'G'], 4),
expect: ['Am', 'G'],
// today: returns [Am,Am,G] — the 3-chord self-overlap [Am,G,Am] scores
// 2 reps × 3² = 18 and beats the true pair at 4 × 2² = 16 (failure map #1).
expectedFail: true,
},
{
id: 'vamp-2-sustained',
description: 'sustained 2-chord vamp filling the whole window (×16) — the everyday two-chord jam',
history: reps(['Am', 'G'], 16),
expect: ['Am', 'G'],
// today: returns [Am,Am,G,Am,G] — a bogus 5-chord self-overlap of the pair
// wins on len² (failure map #1). A musician vamping AmG sees a fake
// 5-chord progression.
expectedFail: true,
},
{
id: 'spurious-substitution',
description: '3-chord loop ×4 with ONE substituted misdetection (Am read as E7 in the third rep) — must still report the 3-loop',
history: ['C', 'Am', 'F', 'C', 'Am', 'F', 'C', 'E7', 'F', 'C', 'Am', 'F'],
expect: ['C', 'Am', 'F'],
// today: returns [Am,F,F,C] — the wrong 4-pattern [F,C,Am,F] straddling the
// noise scores 2 × 4² = 32 and beats the real loop's 3 × 3² = 27
// (failure map #2). This is the user's "doesn't recognize when 3 chords
// return": one bad commit and the display shows a 4-chord ghost.
expectedFail: true,
},
{
id: 'spurious-2of4',
description: '3-chord loop ×4 with an inserted misdetection in two different reps (E7, then Dm7) — realistic sustained noise',
history: ['C', 'Am', 'F', 'C', 'Am', 'E7', 'F', 'C', 'Am', 'F', 'C', 'Dm7', 'Am', 'F'],
expect: ['C', 'Am', 'F'],
// today: returns [Am,F,C,C] — a wrong 4-pattern beats the true 3-loop once
// noise appears in more than one rep (failure map #3).
expectedFail: true,
},
{
id: 'dup-commit',
description: 'consecutive duplicate commit of the same chord inside an otherwise clean 3-loop (C Am F C C Am F …) — dups must collapse',
history: ['C', 'Am', 'F', 'C', 'C', 'Am', 'F', 'C', 'Am', 'F'],
expect: ['C', 'Am', 'F'],
// today: returns [Am,F,C,C] — the duplicate mints a wrong 4-pattern
// [C,Am,F,C] at 2 × 16 = 32 vs the real loop's 27 (failure map #4).
// App.jsx:322-324 currently suppresses adjacent dup commits, so this exact
// history can't arise from today's commit layer — the fixture pins L-30's
// "collapse consecutive duplicates before matching" so the detector is safe
// standalone (and safe if L-31 changes the commit layer).
expectedFail: true,
},
{
id: 'seven-x2',
description: '7-chord loop played 2× — beyond the current length-6 cap',
description: '7-chord loop played 2× — beyond the pre-L-30 length-6 cap',
history: reps(['Em', 'G', 'D', 'A', 'Em', 'C', 'B7'], 2),
expect: ['Em', 'G', 'D', 'A', 'Em', 'C', 'B7'],
// today: returns [A,Em,C,Em,G,D] — a truncated 6-chord slice of the loop,
// because candidate lengths cap at 6 (failure map #5). The user's
// "5 and then 2 others" = a 7-chord form is structurally undetectable.
expectedFail: true,
},
{
id: 'eight-x2',
description: '8-chord loop (extended andalusian form) played 2× — beyond the current cap',
description: '8-chord loop (extended andalusian form) played 2× — beyond the pre-L-30 cap',
history: reps(['Am', 'G', 'F', 'E7', 'Am', 'C', 'Dm', 'E7'], 2),
expect: ['Am', 'G', 'F', 'E7', 'Am', 'C', 'Dm', 'E7'],
// today: returns [Am,C,Am,G,F,E7] — again a wrong 6-chord truncation
// (failure map #5).
expectedFail: true,
},
{
id: 'section-change',
description: 'section change: 4-chord loop A ×3, then 3-chord loop B ×3 — must report B, the loop being played NOW',
history: [...reps(['C', 'G', 'Am', 'F'], 3), ...reps(['Dm7', 'G7', 'Cmaj7'], 3)],
expect: ['Dm7', 'G7', 'Cmaj7'],
// today: returns [Am,F,C,G] — the STALE loop A: no recency weighting, so the
// old section's 3 × 4² = 48 outscores the current section's 3 × 3² = 27
// (failure map #6). The display stays stuck on the previous section.
expectedFail: true,
},
]
+98 -25
View File
@@ -520,18 +520,60 @@ export function toRomanNumeral(chordName, keyRoot, keyMode) {
return isMinorQuality ? rn.toLowerCase() : rn
}
// ─── Repeating progression detection ─────────────────────────────────────────
// ─── Repeating progression detection ─────────────────────────────────────────
// Returns true if arr is made of a shorter repeating unit (e.g. [A,B,A,B] → true)
function isPeriodicPattern(arr) {
for (let p = 1; p <= Math.floor(arr.length / 2); p++) {
if (arr.length % p !== 0) continue
const unit = arr.slice(0, p)
if (arr.every((v, i) => v === unit[i % p])) return true
// True if arr has a "weak period" p < arr.length — i.e. arr[i] === arr[i-p] for
// every i ≥ p, meaning arr is a prefix of some p-periodic infinite sequence.
// This rejects not only exact repetitions ([A,B,A,B], p=2) but also self-overlap
// fragments/rotations of a shorter loop ([A,B,A], p=2; [C,G,Am,F,C], p=4) that
// would otherwise mint ghost candidates out of a short vamp. A genuine loop is
// never weak-periodic: a loop whose tail restates its head would produce an
// adjacent duplicate at the cycle seam, which the window collapse removes.
function hasShorterPeriod(arr) {
for (let p = 1; p < arr.length; p++) {
let periodic = true
for (let i = p; i < arr.length; i++) {
if (arr[i] !== arr[i - p]) { periodic = false; break }
}
if (periodic) return true
}
return false
}
// Match one occurrence of `cand` in `win` anchored at `start` (the first chord
// must match exactly), tolerating at most ONE edit per cycle: a substitution
// (one chord misdetected) or an insertion (one foreign chord slipped between two
// loop chords). The remainder after the edit must match exactly. Returns
// { end, matched, editPos } — `matched` = window indices that matched a loop
// chord, `editPos` = window index of the edit (-1 if the occurrence is exact) —
// or null if no match.
function matchLoopOccurrence(win, start, cand) {
if (win[start] !== cand[0]) return null
const matched = [start]
let i = start + 1
for (let j = 1; j < cand.length; j++) {
if (i >= win.length) return null
if (win[i] === cand[j]) { matched.push(i); i++; continue }
// First mismatch — the single allowed edit. Fork the two readings; each
// requires the rest of the candidate to match exactly from where it lands.
const exactFrom = (wi, cj) => {
const tail = []
for (; cj < cand.length; cj++, wi++) {
if (wi >= win.length || win[wi] !== cand[cj]) return null
tail.push(wi)
}
return { end: wi, tail }
}
const ins = exactFrom(i + 1, j) // win[i] is a foreign inserted chord
const sub = exactFrom(i + 1, j + 1) // win[i] is cand[j] misdetected
const hit = ins ?? sub // insertion keeps one more matched chord
if (!hit) return null
return { end: hit.end, matched: [...matched, ...hit.tail], editPos: i }
}
return { end: i, matched, editPos: -1 }
}
// Returns the lexicographically smallest rotation so the same loop always
// produces the same string regardless of where in the cycle we currently are.
function canonicalize(pattern) {
@@ -544,42 +586,73 @@ function canonicalize(pattern) {
}
/**
* detectRepeatingProgression(history) → chord[] or null
* detectRepeatingProgression(history) → chord[] or null (task L-30)
*
* Tests every unique subsequence of every length (not just the tail) so the
* result is stable regardless of where in the loop the musician currently is.
* Returns the canonical (rotation-normalised) form of the best pattern found.
* Finds the loop the musician is playing NOW in the recent chord history.
* Candidates are contiguous slices (lengths 28) of the last-32 window with
* consecutive duplicate commits collapsed; candidates that are self-overlaps
* of a shorter period are rejected (see hasShorterPeriod). Each candidate is
* scored by recency-weighted COVERAGE: non-overlapping occurrences are counted
* with at most one substitution or insertion per cycle, every matched chord
* adds its recency weight, every edit subtracts the weight at the edit slot.
* Linear coverage (not reps × len²) means a ghost pattern straddling noise can
* never outscore the true loop, and exponential recency decay means the current
* section outscores a longer stale one. Requires ≥2 EXACT occurrences: an
* edit-tolerant occurrence corroborates a loop but cannot establish it — a
* loop means the sequence came back exactly, and a ghost slice that absorbs a
* noise chord into itself occurs exactly only once by construction. Returns
* the canonical (rotation-normalised) best pattern.
*/
export function detectRepeatingProgression(history) {
if (!history || history.length < 6) return null
const win = history.slice(-32)
let best = null, bestScore = 0
// Collapse consecutive duplicate commits — a chord re-committed back-to-back
// is the same loop slot, not two. Non-adjacent repeats (e.g. Em … Em inside a
// 7-chord form) are meaningful and untouched.
const raw = history.slice(-32)
const win = raw.filter((c, i) => i === 0 || c !== raw[i - 1])
const n = win.length
if (n < 4) return null // shortest loop (2 chords) × 2 reps
for (let len = 2; len <= 6; len++) {
if (len * 2 > win.length) break
// Recency weight per window slot: newest chord weighs 1, each step back
// decays by 0.9 (half-life ≈ 6.6 chords).
const RECENCY = 0.9
const weight = Array.from({ length: n }, (_, i) => RECENCY ** (n - 1 - i))
let best = null
let bestScore = 0
for (let len = 2; len <= 8; len++) {
if (len * 2 > n) break
const seen = new Set()
for (let start = 0; start <= win.length - len; start++) {
for (let start = 0; start <= n - len; start++) {
const candidate = win.slice(start, start + len)
const key = candidate.join('\0')
if (seen.has(key)) continue
seen.add(key)
// A pattern that is itself a repetition of something shorter will be
// found at that shorter length — skip it here to avoid inflating scores.
if (len >= 4 && isPeriodicPattern(candidate)) continue
if (hasShorterPeriod(candidate)) continue
let reps = 0, i = 0
while (i <= win.length - len) {
if (candidate.every((c, j) => c === win[i + j])) { reps++; i += len }
else i++
let exactOccurrences = 0
let score = 0
let i = 0
while (i < n) {
const occ = matchLoopOccurrence(win, i, candidate)
// A 2-chord candidate may not take a substitution (1 matched chord is
// no evidence); insertions keep matched === len and stay allowed.
if (occ && occ.matched.length >= 2) {
if (occ.editPos < 0) exactOccurrences++
for (const p of occ.matched) score += weight[p]
if (occ.editPos >= 0) score -= weight[occ.editPos]
i = occ.end
} else {
i++
}
}
if (reps < 2) continue
if (exactOccurrences < 2) continue // implies occurrences ≥ 2
const score = reps * len * len // square length — prevents sub-patterns from beating full loop
if (score > bestScore) {
bestScore = score
best = candidate