acuity. release notes

production acuity arc: the demo now runs the current mavis/acuity adaptive pipeline: deterministic routing into projection pursuit, learned depth, and ensemble paths instead of the older site-local ladder. regression uploads use the production adaptive arc; classification uploads use the acuity multiclass depth path.

csv/arff support expanded

• upload cap raised to 384 MB; live-demo caps are now 75,000 rows, 5,000 prediction rows, 4,096 raw feature columns, and 8,192 encoded features.
• uploads now accept CSV and dense/sparse ARFF, including OpenML-style attribute declarations and sparse data rows.
• target detection now accepts target, class, label, y, or outcome, with last-column fallback for benchmark tables.
• numeric and low-cardinality categorical feature columns are accepted; categorical features are deterministically encoded and common missing markers are handled.
• the dossier now includes OpenML-style RMSE for regression and ranked AUC for classification when class scores are available.
• string class labels now display safely in the prediction table.

classification: acuity now learns label targets, auto-detected: if your target column holds 2–12 distinct integer values, the dossier reports held-out accuracy against two baselines at once: the majority class and the shuffled-label control. blank-row predictions return the label.

measured before shipping

• the classic xor-of-halfspaces (not linearly separable): linear methods pinned at 0.59 accuracy; acuity reaches 0.907 in 443 ms, shuffled control at the majority baseline.
• 4-class variant: 0.920 accuracy vs a 0.307 majority baseline.
• regression behavior is unchanged; calibrated class SETS (conformal for labels) are on the roadmap.

the cannonball challenge: the first dataset you can verify with knowledge you already own. 1,400 projectile shots (launch speed + angle → distance), 16 prediction rows at round numbers, and the textbook law printed right on the page so you can check every answer with any calculator.

measured before shipping

• held-out r² = 1.000, learned in 137 ms server-side; the ladder stopped one rung early (linear 0.845 → focused 1.000).
• across all 16 checkable shots: maximum error 13 cm, mean 5.5 cm; every textbook answer inside the ±0.23 m interval.

auto-scaling: acuity now spends only what your data needs. the engine climbs a ladder of increasingly powerful configurations and stops the moment the held-out grade stops improving; the dossier shows the climb ("auto-scaled: linear 0.029 → focused 0.777 → full 0.823").

• linear relationships now answer in ~0 ms at the first rung; noise exits cheaply without ever reaching the heavy machinery.
• the shuffled control climbs the same ladder and stays cheap, because garbage never earns a climb.
• measured easy-to-hard cost ratio: roughly 1 : 11.

the dossier learns to describe the shape of your data: a per-dimension signal profile ("linear 12% · kink 18% · curve 18% · wave 36%") and detection of cross-feature effects, pairs of columns that act together rather than independently.

• the signal profile is read from the data, not assumed; on the bundled sample it correctly reports the hidden wave and kink components.
• cross-feature admission is gated against a shuffled noise floor with margin, so the count is evidence, not decoration.

initial public demo: upload a csv, get a fresh model learned live with a held-out grade, a shuffled-target control on every run, 90% split-conformal intervals on blank-row predictions, a strict 1 MB / numeric-csv contract, and the in-memory, nothing-retained privacy posture. the honesty rails shipped first, on purpose.