[FonaTech]

0.3.3a0

Release candidate for the 0.3.3a0 pre-release. The full text is in CHANGELOG.md under [0.3.3a0].

Loop kernel (Torch and MLX). recurrent_order="block", recurrent_later_topk, recurrent_later_groups, lambda_iteration_cluster, lookahead_teacher_scope, recurrent_dt_param="sigmoid" and recommended_dt_init(depth) = min(0.5, 2 / depth). Kernel defaults are unchanged, so existing configs and checkpoints behave as in 0.3.2a3. The planner now recommends block order, sigmoid and the horizon init for a looped plan, and ui/presets_v3.json is regenerated.

Physical mechanism. A shared block applied repeatedly adds nearly parallel vectors, so contributions add constructively (consecutive-call output cosine 0.92 to 0.985 at depth 8). Each iteration multiplies the residual by g = 1 + dt * lambda_eff with lambda_eff learned; depth R raises it to the power R - 1. dt alone is a gauge: g stayed 1.067-1.075 for dt 0.15-0.5 at depth 8 because the block rescales its output. A fixed horizon T = depth * dt = 2 replaces the cubic depth rule, which the depth-8 run refuted.

Fixes. The MLX gradient on dt_min and dt_max now equals Torch's clamp. MLX depth jitter is drawn once per optimizer step under gradient accumulation.

Auto-tune and Web UI. The probe trains on real dataset tokens; recurrent_order and recurrent_dt_init join the search space; new fields and four-language help.

Tokenizer and kernel neutrality. general_32k (32,768 tokens, task-neutral control ids, corpus licences in NOTICE) ships and is the planner default for new plans. The control-token relabel changed its fingerprint (169668b1… to 6d4b3b52…); the old value is kept under derived_from. Kernel text no longer names a translation task, and the translation sections moved to demos/translation.

Verified. Full suite 732 passed, 4 skipped (three need playwright, one needs the v2 reference checkpoint); three further runs under nice -n 19 also passed. ruff check . is clean (local 0.15.11; CI pins 0.17.4, not yet run locally).

Open, and stated in the CHANGELOG.

• All loop measurements are one seed per arm at 15M parameters; the planner's recommendations are extrapolated to every size it plans.
• Lazy serving of a block-order checkpoint is unmeasured.
• The recurrent_later_groups=2 regression is explained only in part: the clamp collapsing the two active layers' dt accounts for most of it, but the sigmoid arm still trails the full block by 0.013-0.055.
• The 4-6% step-time overhead of the new kernel is not attributed.
• One run of the full suite failed once in test_worker_probes_checkpoint_after_validation_and_monitor_shows_it (preexec_fn error spawning the probe subprocess) and passed in 6 later runs and 450 isolated spawns; the cause is not found.

🤖 Generated with Claude Code