skycmd
Every skycmd demo in one place, with its verdict
Tiny decision models trained from scratch in Rust that run in the browser: every model with its verdict, its demo and the data it read.
Every skycmd demo in one place, with its verdict
Note Every skycmd demo in one place. Pick a model to open its demo, which says first whether the model passed or failed its gate.
Note Passed its pre-registered gate. 0.480 on the typed-decisions bench, above the majority class (0.398) that its gate asked it to beat, an ECE of 0.0479 (guard 0.08) and no game more than 5% below s3-pico.
Note Passed its pre-registered gate. Second version: 0.553 on the typed-decisions bench (gate 0.45) and ECE 0.0588 (guard 0.08).
Note Passed its pre-registered gate. On its synthetic bench, from our own generator, it beat a gitleaks-style rule scanner: secret 0.9990 against 0.977, service 0.9795 against 0.787.
Note Passed its pre-registered gate. An end-of-turn detector (the last 4 s of audio): 0.8993 on the 31,527 test clips of Smart Turn v3.2 (bar 0.85; the teacher measured 0.9260 on the same clips), false cut-off 0.0847 and END recall 0.8627 at the validation threshold tau85, 2.39 ms per decision in WASM and 631,936 bytes.
Note Passed its pre-registered gate. Passed on fresh slates; the system was defined after v2's test was read, from an exploratory reading of v2's test dumps, with its parameters from validation.
Note Failed its pre-registered gate: published as a result, not for use. It missed its calibration guard: ECE 0.0821 against 0.08 at full depth (0.547 on the typed-decisions bench, where the gate asked for at least 0.45).
Note Failed its pre-registered gate: published as a result, not for use. It missed three of its four gates: end of turn 0.5681 against 0.9263 for Smart Turn v3.2, keywords 0.8885 against 0.90, French zero-shot keywords 0.1330 against 0.20; its control model, trained without the Mimi pretraining, did better on all four sets.
Note Failed its pre-registered gate: published as a result, not for use. It failed E1, E2 and E3: on eot-bench (real human-to-agent speech, 14 languages) read 0.2 s into each pause, its AUC is 0.7197 against v2's 0.7448 on the same cuts (−0.0252 [−0.0336, −0.0167], at least +0.02 with a lower bound above 0 asked), its recall at a false cut-off of 0.10 is 0.0563 below v2's, and outside English it loses 0.0209 [−0.0308, −0.0111] (a lower bound above −0.02 asked); Smart Turn reads 0.7873.
Note Failed its pre-registered gate: published as a result, not for use. On GIFT-Eval it scored nGMASE 0.9471 and nWQL 0.6646, where the gate asked for 0.838 and 0.587.
Note Failed its pre-registered gate: published as a result, not for use. On GIFT-Eval it scored nGMASE 0.8847 and nWQL 0.6190, where its gate G1 asked for 0.838 and 0.587 (v1: 0.9471 and 0.6646; our TinyCast port without flip: 0.7918 and 0.5622); G2 holds, its MASE ratio to the seasonal naive is under 1.00 at each term (short 0.833, medium 0.928, long 0.988).
Note Failed its pre-registered gate: published as a result, not for use. It missed G1 on its every-repository reading: line F1 0.8561 against 0.7332 + 0.15 for the best rule scanner (Betterleaks); the six other gates held.
Note Failed its pre-registered gate: published as a result, not for use. It missed G1 against BM25, the lexical score it reads in its input: 0.8395 against 0.8500 on 2,000 slates of tools never seen in training (above potion-base-2M + 0.03 and both size-matched baselines), and G4: 38.9 ms per request in WASM against 2 ms.
Note Failed its pre-registered gate: published as a result, not for use. It failed G1, G2, G3 and G6: on 2,000 fresh requests about the tools of MCP servers, a family that no training slate holds, the pair reads 0.5610 against BM25 0.5675 (−0.0065 [−0.0235, +0.0120], a lower bound above 0 asked); under masked names it answers none on 0.3345 of the slates (0.10 to 0.30 asked); its ECE is 0.1845 (at most 0.05 asked); with the request hidden its model part is at chance (0.1406) and the pair is not…
Note Failed its pre-registered gate: published as a result, not for use. It failed H1 and H2 and the s6-nano gate: four sleep passes add +0.0030 on the recall test (95% interval −0.0010 to +0.0068, above 0 asked), and it plays two games far worse than the control without memory (flappy 15.87 against 71.05, dodge 5.89 against 7.94); H0 holds: with its memory it is right on 0.4892 of the recall asks against 0.2440 for the control (+0.2452).
Note Failed its gate, fixed before training: published as a result, not for use. It scored 0.343 on the typed-decisions bench, where its gate asked for more than 0.55, and stayed below the majority class (0.398).
Note Published without a gate. A 17M-parameter decision model whose options cannot see each other: reordering the options or renaming their keys cannot change an answer.
Note Published without a gate. A 251 KB ternary decision model that plays four games in your browser, with a looped state block and early exits.
Note Published without a gate. A 545 KB ModernBERT decision model in the System One format, the same file that runs in llama.cpp and wllama.
Note Published without a gate. A 702 KB ternary decision model with a cached question stream and early exits.
Note Published without a gate. A 702 KB ternary model that plays a Doom-like shooter from 14 numbers (Freedoom art).
715 KB model for four games and typed questions
Note Demo: 715 KB model for four games and typed questions.
2.4 MB model for four games and typed questions
Note Demo: 2.4 MB model for four games and typed questions.
250 KB secrets model: a result, not a scanner
Note Demo: 250 KB secrets model: a result, not a scanner.
632 KB end-of-turn detector; passed its gate
Note Demo: 632 KB end-of-turn detector; passed its gate.
3.7 MB tool router; passed on fresh slates
Note Demo: 3.7 MB tool router; passed on fresh slates.
Failed its gate: 2.4 MB s6 model, a result
Note Demo: Failed its gate: 2.4 MB s6 model, a result.
Failed its gates: 899 KB audio model, a result
Note Demo: Failed its gates: 899 KB audio model, a result.
Failed its gate: 632 KB end-of-turn v3, a result
Note Demo: Failed its gate: 632 KB end-of-turn v3, a result.
Failed its gate: 143 KB forecaster, a result
Note Demo: Failed its gate: 143 KB forecaster, a result.
Failed its gate: 149 KB forecaster, a result
Note Demo: Failed its gate: 149 KB forecaster, a result.
Failed its gate: 250 KB secrets verifier, a result
Note Demo: Failed its gate: 250 KB secrets verifier, a result.
Failed its gate: 2.4 MB tool router, a result
Note Demo: Failed its gate: 2.4 MB tool router, a result.
Failed its gate: 3.7 MB tool router v7, a result
Note Demo: Failed its gate: 3.7 MB tool router v7, a result.
Failed its gate: 752 KB memory model, a result
Note Demo: Failed its gate: 752 KB memory model, a result.
Failed its gate: 2.4 MB decision model, a result
Note Demo: Failed its gate: 2.4 MB decision model, a result.
17M decision model, invariant to option order and keys
Note Demo: 17M decision model, invariant to option order and keys.
251 KB model playing four games in the browser
Note Demo: 251 KB model playing four games in the browser.
545 KB ModernBERT decision model playing games
Note Demo: 545 KB ModernBERT decision model playing games.
702 KB ternary decision model with early exits
Note Demo: 702 KB ternary decision model with early exits.
702 KB model playing a Doom-like shooter
Note Demo: 702 KB model playing a Doom-like shooter.
Note Trained or tested on it: s6-nano (passed), s6-micro (passed), s6-micro-v1 (failed), s6-sleep3-nano (failed), s5-lite (no gate), s4-micro (failed), s3-pico (no gate), s1-pico (no gate), s2-nano (no gate), s2-nano-doom (no gate).
Note Trained or tested on it: s6-nano (passed), s6-micro (passed), s6-micro-v1 (failed), s6-sleep3-nano (failed), s3-pico (no gate), s1-pico (no gate), s2-nano (no gate), s2-nano-doom (no gate).
Note Trained or tested on it: s6-nano (passed), s6-micro (passed), s6-micro-v1 (failed), s6-sleep3-nano (failed), s3-pico (no gate), s1-pico (no gate), s2-nano (no gate), s2-nano-doom (no gate).
Note Trained or tested on it: s6-secrets-pico (passed).
Note MASSIVE intents in English and French as command records ({src, lang, user, intent, split}), official splits.
Note Trained or tested on it: s6-nano (passed), s6-micro (passed), s6-micro-v1 (failed), s4-micro (failed).
Note Trained or tested on it: s6-cast-pico (failed), s6-cast2-pico (failed).
Note Trained or tested on it: s6-nano (passed), s6-micro (passed), s6-secrets-pico (passed), s6-micro-v1 (failed), s3-pico (no gate), s1-pico (no gate), s2-nano (no gate), s2-nano-doom (no gate).
Note A fixed sample of the four synthetic series generators that pretrain the s6-cast forecaster (Gaussian processes, pulses, trend-seasonality-irregular, SARIMA), and of their training mix. Generated by skycmd's own Rust code: no third-party data.
Note Complete episodes of the four skycmd games (Flappy Drone, Rescue Run, Obstacle Dodge, a Doom-like shooter) played by four small skycmd decision models and by the exact oracle: at every decision, the state, the model's probabilities, the oracle's soft target, the reward and the events. Generated by our Rust game cores and models: no third-party data.
Note Trained or tested on it: s5-lite (no gate).
Note Trained or tested on it: s6-ear-nano (failed).
Note Trained or tested on it: s6-route-nano (failed), s6-route (passed), s6-route7-nano (failed).
Note Trained or tested on it: s6-route7-nano (failed).
Note Trained or tested on it: s6-nano (passed), s6-micro (passed), s6-micro-v1 (failed), s6-sleep3-nano (failed).
Note The calibration and test questions of the LFM3-1.7B System 1 / System 2 router, which failed its pre-registered gate; no published model reads them.