BarScript (experimental preview) — C1.3, denominator augmentation, seed 1234

Audio log-mel + a bar grid → a Taiko no Tatsujin chart, decoded under a finite-state grammar over bar-scoped plan, skeleton and realization stages.

BarScript is the release name of this charting model; it is also the name of the bar-scoped multi-stage token encoding it decodes into. It is a preview under evaluation, published so it can be tried, not as a finished model: the champion comparison below is inconclusive, and §2 lists measured defects — one of which fires a pre-registered kill condition. The seven published SoftChart 1.x models are a separate, unaffected line.

This card is written for THIS checkpoint. Every number below was measured on C1.3's own generations under the serving code this package ships (generate.py md5 73b6e49b…, min_onset_gap_sec = 0.020). Nothing is inherited from the C1 card. Where an item on the C1 card has no C1.3 measurement, it is marked not measured on this arm rather than carried over.

Read the KNOWN LIMITATIONS section before using any number above it. This model has real, measured defects, one of which fires a pre-registered kill condition, and its difficulty ladder is roughly 25 coverage points less nested than an authored one.

---

0. Status: the champion comparison is inconclusive

scripts/sc2_gates.py champion, three pairings on one serving code base (experiments/sc2_eval/FINAL_CHAMPION.md §1):

MIN_SEEDS_FOR_VERDICT is 2 in both arms and C1.3 has one seed. The rule fires before any metric is read. No arm has been declared champion.

What the clause ledger says, which is one-directional: C1.3 passes the required §4 clause plan_token_calibration, which C1 fails on both its seeds. No clause is passed by C1 and failed by C1.3. The two a2 span clauses fail on all four arms and belong to the campaign, not to this checkpoint.

The shipping rationale, the two named compromises and the decision rule for choosing a different checkpoint are in experiments/sc2_eval/SHIP_DECISION.md.

Gate items on the 71-chart hard+oni population, this arm against the C1 seed-1234 arm under identical code:

---

1. What it does

• Input — 128-bin log-mel at 22 050 Hz (preprocessor_config.json freezes the exact decode/STFT/mel contract), plus a bar grid with explicit measure edges, exact rational meters and per-bar lattice denominators.
• Conditions — course, authored level, density bucket, and the split-axis knobs big_rate / span_rate / stream / sync.
• Output — a BarScript token stream decoded under BarscriptFSM, exported as TJA on the 96-slot lattice.
• Family decoding — generate_chart_family decodes a song's courses as one nested ladder, hardest first, each easier course biased toward its harder sibling's onsets. Off by default; the campaign's headline population was decoded independently (§2.1).

Capabilities this checkpoint carries (config.json:capabilities): aux, beat_head (hi-res), hierarchical_ctx, axis_knobs, sync_token, stage_emb, span_duration_head, tempo_head, rich_section_stats. It carries no style, sibling, ctx, plan-prefix, slot, dual, align, mask_infill, func_time, global_ctx or complexity conditioning; requests on those axes reach nothing.

What works

Evidence marks: ● n ≥ 71 charts with a measured C1 seed band; ◐ n = 19–71 charts or 8–12 sweep cells, one seed; ○ n < 19 or a CI covering zero.

---

2. KNOWN LIMITATIONS

Nothing in this section is softened. Where a limitation is invisible to the gate suite, that is said.

2.1 The difficulty ladder is 25–32 coverage points less nested than authored

Coverage = the fraction of the easier chart's notes that have a note in the harder chart within tolerance. Two decode modes, two different numbers, and both belong on the record.

Independent-decode figures: FINAL_CHAMPION.md §8.1, n = 13 / 13 / 35 songs, one seed, same serving code as this package (●/◐). Family-decode figures: C13_PREDICTION.md §3.3, n = 13 songs, one seed (◐).

Five things make this worse than the headline:

• nesting_coverage and hand_agreement FAIL the family gate on every arm measured, C1 included. difficulty_monotonicity passes on all of them.
• C1.3 is slightly worse than C1 here. C1's independent-decode coverage on the same population is 0.690 / 0.704 / 0.737; C1.3 is −0.032 / −0.037 / −0.077 against C1 seed bands of 0.010 / 0.021 / 0.025. The move is small against the ~0.30 gap to authored that every arm shares, but it is in the wrong direction.
• Hand agreement at coinciding hits is near chance on hard→oni: C1.3 0.539 against a marginal-chance null of 0.500 and an authored 0.753. When two generated courses agree that a note belongs somewhere, which drum they pick is near-independent across courses.
• The /16 figure overstates the model's nesting. The same checkpoint on a /96 grid drops to 0.649 / 0.730 / 0.752, worse on 11–12 of 13 songs (p = .003 / .022 / .022). A coarse lattice manufactures agreement by leaving few places to disagree (C13_PREDICTION.md §5.2).
• family_bias = 2.0 / family_hand_bias = 1.5 are PROVISIONAL — described in generate.py as logit offsets on a decoder never trained at that setting, and never calibrated against the authored target.

2.2 The note-type channel carries almost no information

71 charts, 22 260 matched slots (FINAL_CHAMPION.md §5.2, ●).

• The model sits below its own majority floor on both endpoints, and below the campaign's corpus baselines (56.3 % / 62.0 %), which are a fixed reference line from a separate 59 961-hit census, not this population.
• The honest caveat, which cuts the other way. The published floors are argmax predictors scored against a temperature-1.0 top-p-0.95 sample. Against the sampler-appropriate i.i.d. floor (Σ p² = 0.4582 4-way, 0.5190 hand) this model is above baseline by +0.0194 and +0.0303 — the largest 4-way excess of any arm in the campaign. Both readings belong on the record.
• MI is flat at ~0.010–0.012 bits on all four arms (seed band 0.0008). The accuracy differences between arms are marginal-matching, not information.
• Accuracies are conditional on coverage 0.6442 (precision side) / 0.5842 (recall side); more than a third of generated hits have no authored partner, and coverage is 9 % lower than C1's, so this population is smaller and differently selected than C1's.
• No inter-charter agreement ceiling exists for this split — no (song, course) carries two independent authored charts — so 100 % is not a legitimate target and is not used as one.
• Greedy re-decoding was not repeated on this arm. The C1-era finding that greedy closes part of the gap at an unacceptable motif_reuse cost is not measured on this arm and must not be quoted for it.

2.3 The note budget is 9 % short, and it is the whole of the apparent timing regression

precision × budget equals onset_recall to four decimals on every arm, by construction. Those three "timing" rows are one quantity, and it is the note budget, not placement: precision is inside the C1 seed band, median offset is 0 ms, exact_slot_lift_over_null is above the band, and significant drift tests fall from 6 / 5 to 2.

The shortfall itself is a real regression and its cause is unknown. It appears on both C1.2 and C1.3, so it belongs to the --cal-axes / --balance-exclude-big pair rather than to the denominator augmentation. It is the largest practical regression in this release.

2.4 Raising density squeezes drumrolls out — this is the compromise a user can hit

Median spans/min over eight oni songs, per requested density bucket (authored 3.885; FINAL_CHAMPION.md §4.3, 8 songs × 5 buckets, ◐):

ρ(density → span_per_min) = −0.759, sign p 0.0078, first→last −2.443 /min. At the top of the density request C1.3 delivers 33 % of the authored span rate; C1 delivers 76–79 %, C1.2 44 %.

• Do not advertise the density knob at its top setting until this is fixed.
• There is a named suspect: cal_span never left 0.1667 on either challenger, so this is plausibly a calibration term that never trained rather than an intrinsic cost of the knob.
• The mirror image is a gain: C1.3's span_rate knob no longer drags hit density, which C1's does.

2.5 Absolute bucket calibration is still wrong on the density axis

The density knob orders the request correctly and lands in the wrong bucket, under-shooting by ~1 bucket on average — more than C1.2 does. Rank control is real; absolute density targeting is not delivered.

Note the cost that comes with the big_rate calibration win: C1.3 has the shortest big_rate ladder of any arm (first→last +0.159 against C1's 0.183 / 0.224), so the top of that request now delivers less accent than C1's did.

2.6 It fails its own span-quality gate and triggers a pre-registered kill condition

71 charts hard+oni (467 closed spans) and 26 charts easy+normal (272):

The pre-registered kill condition (SOFTCHART2_DESIGN.md §4-A2) reads: rate is calibrated but placement/length quality is bad → concede that the bar-scoped commitment mechanism is insufficient and escalate to an independent span-state objective. It fires on all four arms, including C1. Shipping does not discharge it.

placement_fit is a knife-edge, not a finding. C1.2 passes at 0.561 and C1.3 fails at 0.560 with CIs overlapping every other arm across their whole width. One thousandth of AUC decides the verdict. It is reported, not used.

2.7 One span in seven is a micro-span

All generated spans, 97 charts, FINAL_CHAMPION.md §2 (●):

Per course: easy 6.6 %, normal 8.1 %, hard 13.7 %, oni 25.2 % — authored 3.0 / 3.3 / 2.1 / 7.1 %.

• C1.3 does not inherit C1.2's worsening (16.4 %); 14.7 % is exactly at the top of the C1 seed band (13.6–14.7 %). It is not an improvement either.
• Mechanism is known and unfixed on every arm. Duration bucket 0 is open at the bottom; 123 of C1.3's 739 spans sit in it, at a median 0.5 beats against the authored 0.917.
• The shipped floor does not fix it. span_min_sec / span_min_beats fired 263 times (span_close_floor) plus 33 carries across 97 charts and the < 0.25 s share is still 14.7 %.
• The mitigation is unbuilt. span_bucket0_profile is null in this package because scripts/build_span_bucket0_profile.py has never been run on the train split. This is the shortest available fix and it is not in the box.
• The gate cannot see it directly. a2_span_duration compares p50/p90/p99 only and passes; duration_low_tail catches the support violation but not the rate (its rate verdict passes on this arm).
• On a /16 deployment grid the same quantity is 14.7 / 12.4 / 30.6 / 21.8 % — hard is untouched by the augmentation and is the worst cell in the release.

2.8 The deployment grid has a two-role conflict that no checkpoint dissolves

BAR_DENOM is simultaneously an input the grid supplies and a trained output token the model reads as a density announcement. bar_denoms="supplied", the shipped default, teacher-forces it.

13 songs × 4 courses, one seed, --cond authored, C13_PREDICTION.md:

Authored triplet rates for reference: 4.20 / 5.72 / 7.43 / 9.95 %; ultra-fine ~0.00 %.

The recommended deployment configuration is --grid bpm --grid-bpm-denom 16 with the package default bar_denoms="supplied". Its two costs, stated plainly: triplets are exactly 0 and cannot be otherwise, and hard micro-spans stay at 30.6 % against an authored 2.1 %.

Four consequences a deployer must carry:

• Deployment costs 10–46 % of onset F1 relative to the authored grid: easy −46 %, normal −34 %, hard −22 %, oni −10 %. Every --grid authored number in this card is therefore optimistic for a user who supplies only BPM.
• p90 |offset| is 0 ms on the authored grid and 14.6–36.7 ms on the bpm grids. That residual belongs to the bar edges; no denominator choice touches it.
• The model's off-dyadic rate is a property of the model, not the lattice. Given any 3-divisible grid it puts 2–3× the authored share off the dyadic grid; the grid only decides whether that mass gets called "triplets" (/24) or "ultra-fine jitter" (/96). This is a training-side defect and no deployment grid fixes it.
• The gate suite cannot score the deployment path. robustness.py:629-642 raises ContractError on uniform-BPM arms, so every gate timing metric is null for the path the model ships on; the /16 figures above come from timing.py, a second implementation that reproduces the gate on the authored grid.

The deploy_bpm_grid profile in this package is NOT the recommendation. It sets bar_denoms="model", bar_denom_mask="div96", per DEPLOY_GRID_FIX.md §7 — a recommendation made explicitly conditional on C1.3 failing, which it did not. That arm was measured on C1 only: density 1.06 / 1.08 / 0.97 / 1.01 and nesting 0.635 / 0.706 / 0.778. It has never been run on this checkpoint. The profile is kept because it is implemented and because the arm is the obvious next experiment, not because it is advised here.

One contradiction inside this package, stated so nobody has to find it. config.json:serving_profiles.deploy_bpm_grid.basis still quotes DEPLOY_GRID_FIX.md §7 verbatim, including the words "Recommended now", and still carries C1's nesting figures. That string is generated by scripts/sc2_package.py, which is read-only for this release. This section supersedes it. The generated string is left intact rather than silently diverging from the shipped code.

2.9 Playability floor: gaps no human hand can play

Recomputed on this arm's own events, 71 charts hard+oni, generated vs authored on the same songs (producer experiments/sc2_eval/ship_c13/playability.py, records .../records/playability_all_arms.json; ●):

C1 on the same population: 151 gaps (0.395 %), shortest 16.7 ms, peak 22 notes/s. C1.3 is better on every row and still 56× the authored rate.

• min_onset_gap_sec = 0.020 is a degeneracy guard, not a fix: it is set below the fastest thing the series has shipped (29.4118 ms, TAIKO-TONGUE-TWISTER oni, BPM 170, 48th notes) precisely so it cannot refuse a chart the domain writes. It fired 125 times across 97 charts with 0 fail-opens. A mask cannot fix a distribution.
• The generate.py constant comment quotes the generated sub-40 ms rate as ~0.35 % against ~0.045 % authored. That authored figure does not reconcile with the 0.0053 % measured here, and the two have never been put on the same denominator. The excess is 8× on the comment's accounting and 56× on this one.
• Good news that replaces a C1-card claim. The C1 card reported the model placing hits in 19.6 % of sub-0.25 s scaffolding bars. Under today's serving floors this arm places 0 hits in all 112 such bars across 97 charts, exactly as the authored side does; degenerate_bar_hits_blocked fired 12 times with 0 fail-opens.

2.10 Pattern proxies regress against C1 on hard+oni

71 charts, one code version, C1 band = |s1234 − s4321| (FINAL_CHAMPION.md §3):

The ×band column overstates the case — several bands are under 1 % of their own level — so the relative column is the honest one. On motif_reuse the bootstrap CIs are nonetheless disjoint: C1 [−0.392, −0.334] / [−0.388, −0.334] against C1.3 [−0.447, −0.396]. The regression is real and not seed noise.

Four bounds, all measured, none of them a dismissal:

• It does not happen on easy+normal. Every pattern proxy there is inside the C1 seed band, motif_reuse_gap is marginally better than C1's mean, and class_4gram_js is better on both challengers.
• The decomposition puts the loss on rhythm and hand, while the accent layer improves: rhythm 0.1251 → 0.1640, hand 0.1774 → 0.2102, accent 0.0606 → 0.0480.
• ioi_js_per_chart is where the augmentation shows up — it reads the exact rational IOI lattice, and it is C1.3's worst pattern metric relative to C1.2 (+9.7 band against +1.5).
• The proxy reads mostly a channel carrying ~0.010 bits about the authored type (§2.2), and it has never been validated against a listener. It is quoted because §4 names it, not because it is strong.

Realized motif_reuse is 0.3165 against an authored 0.7387 — under half the authored repetition.

2.11 The note-type marginal, and one C1 finding that does NOT reproduce

Big notes are 1.48× the authored rate pooled hard+oni (0.0873 vs 0.0589), worst on the easiest course:

The pre-registered stop window (0.045–0.075 pooled hard+oni) is not met.

The C1 card's §2.8 claim does not reproduce here and is corrected. On C1 the realized class marginal was 13× closer to the training loss weights than to the corpus. On C1.3 it is the other way round: JS(gen ‖ authored) = 0.00588 against JS(gen ‖ class-weight prediction) = 0.00746 (JS(authored ‖ weights) = 0.01452). --balance-exclude-big moved the marginal off the loss weights and toward the corpus. The log-log fit of realized share against class weight collapses from slope 0.378 (ρ 0.436) on C1 to slope 0.078 (ρ 0.156) on C1.3.

2.12 Fine-lattice, tuplet and syncopation accuracy are much worse than average

robustness.py timing strata, 71 charts, exact-slot rate (overall 0.5840):

Every stratum is lower than C1's, which is the note-budget effect of §2.3 acting on a per-stratum recall, but the shape is the finding: tuplet positions and fast lattice steps are 0.24 below the overall rate, and accuracy decays monotonically with syncopation.

2.13 Smaller, but on the record

2.14 The evidence base is thinner than it looks

• One seed. Every number in this card is n = 1 in the training seed. The C1 seed band quoted throughout is C1's, used as a reproducibility scale; it is a range over n = 2 and carries no confidence statement.
• easy and normal are evaluated on 13 songs. The 71-chart population is hard + oni only; the easy/normal population is 26 charts from 13 songs, and its seed band is 4–5× wider than hard+oni's.
• Every difficulty-family and deployment-grid number is 13 songs, one seed (◐). Every knob sweep is 8–12 songs × 5 buckets (◐).
• At these sizes a gate pass carries little information. Worked example from this campaign: the big_share_of_hits gate flips between two C1 training seeds while the failing seed's point estimate is smaller. At n = 36 that gate cannot rank checkpoints, and no pass/fail on it should be quoted as evidence for any arm.
• Nothing here is audio-referenced except a2_span_placement_fit. Every other quantity is chart-vs-chart on the exact rational lattice or on hit order. No timing window and no game judgement parameter appears anywhere in this path.
• best_val is comparable across C1 / C1.2 / C1.3 (one training cache) and not comparable to C1.1 or to any arm on a different cache. C1.3's 0.5769 is 2.9 % worse than C1.2's 0.5605; that was the registered trade and the deployment side won it.
• No plan-neutral arm was generated, so the plan_neutral_fallback §4 clause is na and the ledger is incomplete by one required clause.

2.15 Serving regime: what motif_constraint: auto resolves to, and why it matters less here

motif_constraint: "auto" resolves to OFF for this checkpoint: the training cache's barscript_md5 (634e3dc3…) differs from the serving barscript.py (9afc715a…), so the gold sequences it trained on never satisfied the MOTIF hard constraint. train_serve_matched = true — OFF is the train/serve-matched choice and it is the right default.

Unlike the C1 package, this card needs no correction for it. Every number in this card was measured with the constraint OFF, i.e. under exactly this package's resolved default; the 20 evaluation runs all record serving_fsm.motif_constraint: off, verified: true. In particular motif_ref_marginal_js here is 0.0452 (hard+oni) and 0.0114 (easy+normal), both measured OFF.

Two things that still belong on the record:

• The ON/OFF blast radius was measured on C1 only, where switching the constraint on moved motif_ref_marginal_js 0.0311 → 0.0189 (−39 %). On the C1 package the shipped default therefore makes the correct value 0.0311, not the 0.0189 in the older tables. That contrast has never been measured on C1.3, so no ON-constraint number should be quoted for this checkpoint.
• Serving under the constraint ON would be train/serve mismatched for this checkpoint and is not a supported configuration.

---

3. Serving contract

Every value below is in config.json:serving, with its justification in config.json:serving_basis. Pass them explicitly — the package's serving_kwargs() does — so the recorded contract is the one that reaches the decoder.

_has_sync must be forced on at load. train.py records sync_token=False while the training prefix carries the SYNC slot, so a loader that trusts the recorded flag drops the slot and every sync bucket reaches the decoder as an identical prefix. The package loader does this and records why in config.json:serving_prefix_fix. This is a workaround; the fix belongs in train.py.

Recommended deployment configuration

--grid bpm --grid-bpm-denom 16          # uniform /16 grid
bar_denoms = "supplied"                 # the package default
motif_constraint = auto  -> off         # resolved at package time
family decode: optional; β = 2.0 / 1.5 is provisional

Expected behaviour under it, 13 songs × 4 courses, one seed: notes within 13 / 11 / 7 / 3 % of authored; onset F1 0.333 / 0.374 / 0.466 / 0.580; zero triplets; hard micro-spans ~30 %.

---

4. Weights are bfloat16

Training ran bf16 autocast and CUDA inference runs bf16 autocast, so fp32 storage carried no information the forward pass could use — the argument the v1.5 release made and verified. Measured cast cost: max absolute delta 0.00711 on frontend.2.weight (0.349 % of that tensor's max) over 9 107 291 float elements.

This is an argument about the compute dtype, not a proof that the two checkpoints decode identically. Sampling is chaotic in the logits, so individual charts can differ. The paired fp32-vs-bf16 evaluation has not been repeated for BarScript. The fp32 checkpoint is preserved in the source run directory and is the reference for any bit-level comparison.

---

5. Provenance

Reports: experiments/sc2_eval/SHIP_DECISION.md, experiments/sc2_eval/FINAL_CHAMPION.md, experiments/sc2_eval/C13_PREDICTION.md, experiments/sc2_eval/C12_EVAL.md, experiments/sc2_eval/DEPLOY_GRID_FIX.md.