What the v1.0 surface could not answer

§1Hypothesis scoreboard

§2E1 — test-time scaling past the v1.0 cap

21 runs (A3, B3, C1–C5 × 3 seeds) × T ∈ {1,2,4,8,16,24,32,64} on 200 eval items. Every T is computed over the same 200 items: the T ≤ 16 points are re-aggregated read-only from v1.0's per-item rows rather than reused at 400 items, so the curve measures the T range and not a change of subset.

§3E3 — where you spend the steps matters as much as how many

15 C-budget runs × 3 frozen per-level schedules at matched total NFE = 32, full 400 items, 18 000 item-rows. Paired item-level bootstrap (2 000 replicates, RNG 7331); pairing is on the eval item, so item difficulty cancels.

Primary: WER(coarse) − WER(fine) = −0.9289, CI [−0.9595, −0.8999]. Secondary: WER(coarse) − WER(uniform) = −0.0953, CI [−0.1044, −0.0870]. Both supported. At identical inference cost, moving 32 forward passes from the finest codebook level to the coarsest changes WER by 5.2× and cuts the degenerate rate from 39.1% to 0.1%. Level 0 sets the phonetic content every later level conditions on; level 7 carries residual detail that no amount of iteration can use to repair a wrong coarse plan.

Calibration. Uniform at NFE=32 is T=4 per level, and E1 measures WER 0.321 at T=4 vs 0.191 at T=8. Coarse's 0.219 sits between them, so reallocating the same NFE to the coarse level buys roughly what 1.7× more uniformly-spent NFE would buy — for free.

§4E2 — iso-latency allocation, and a corrected interpretation

Zero-GPU analysis over the frozen v1.0 surface with measured c_layer; serial latency L = 8·T·d·c_layer, 40 log-spaced budgets, optimum found two independent ways (fitted M_sep surface, and measured config×T means).

H-E2 is refuted on both grids (5.9% and 37%, against an 80% bar), so the pre-registered verdict is unchanged and now shown to be robust.

§5E5 — robustness of the fitted surface

The known v1.0 defect: M_full's width amplitude A hit its pre-registered upper bound (A = 10.000), leaving α — and therefore ρ = α/β and the H-D1 test — unidentified. The log-amplitude refit makes log A, log B, log C effectively unbounded and leaves exponent bounds exactly as pre-registered.

The headline survives the reparameterisation. Δτ = 0.1262 CI [0.1043, 0.1474] under log amplitudes, against v1.0's 0.1102 CI [0.0921, 0.1312] — same sign, overlapping CIs, both excluding 0. Δρ = −0.0516 CI [−0.4281, 0.3077] still contains 0, so H-D1 remains unsupported here too. Labeled exploratory-sensitivity: this cannot and does not rescue H-D1's pre-registered status.

§5bE4 — does it survive a 4× larger budget?

D1–D5 × seeds {0,1} at 276M non-embedding parameters, trained at the LR the G1-D remedy adopted, then evaluated on the full T grid. Combined with the frozen v1.0 table this gives a 275-row, four-budget surface spanning 20M → 276M.

§5cE6 — is it just undertraining?

{C1, C3, C5} seed 0 retrained to 90k steps (3× v1.0's 30k). Item-level bootstrap, 1000 replicates — item-level because one seed per config leaves nothing to resample at the run level.

§6Gates

§7Defects found and fixed during v1.1

An adversarial review of the extension code before any result was trusted found three real defects. All are recorded in full in LOG-v1.1.md; corrections are appended to RESULTS-FEED.md as new entries, never edits.

§9Why no d*(N) claim appears anywhere

Raw optimal depth by budget reads A 18, B 30, C 36, D 14 — a rise-then-collapse story. It is not one, and the artifact now records why so it cannot be read off naively:

• B and C are censored. Their d* sits at the deepest shape those budgets tested (30 and 36), so they are lower bounds, not interior optima — the true optimum may lie beyond the grid.
• D's minimum is a near-tie. d=14 gives WER 0.0978 against d=38's 0.0990 — a gap of 0.0012, well inside noise, with a bump between them.

Only A (18) and D (14) are genuine interior optima and D's is a coin flip. e4_scale.json carries censored_at_max_depth and margin_to_runner_up per budget. d* was pre-registered as exploratory with no decision rule, and no d* claim goes in the paper.

§8Reproduction

python src/audit_v11.py --smoke          # E-AUDIT: checkpoints + sampler reproducibility
python src/e1_extended.py                # E1: runs_ext.csv, T* values, tau refit, H-E1
python src/iso_latency.py                # E2: H-E2 on the v1.0 grid
python src/iso_latency.py --runs results/artifacts-v1.1/runs_ext.csv \
       --fits-from-e1 --budgets C --out-tag _ext   # E2 sensitivity on the extended grid
python src/e3_nfe.py                     # E3: H-E3, paired item-level bootstrap
python src/logamp.py                     # E5: log-amplitude refit (exploratory-sensitivity)
python src/launch_e4.py --gpus 0,4       # E4: adopt LR-sweep argmin, launch budget-D grid
python src/supervise_v11.py              # drain E4/E6 synth+score onto idle GPUs
python src/e4_analysis.py                # E4: 4-budget surface, H-E4
python src/e6_analysis.py                # E6: 90k control, H-E5 (item-level bootstrap)

Job dispatch uses src/run_v11.py, which claims jobs with atomic O_CREAT|O_EXCL lock files so several dispatchers can share one job file as GPUs free up — the fix for the duplicate-trainer failure recorded in v1.0's log.

← project page · protocol · implementation · v1.0 results