dyno.dulse.dev
Home / C and Dulse

The same tasks in C and in Dulse

Every demo runs in C. Some run again in Dulse, a language for firmware whose compiler knows the chip: it checks register use against the reference manual and interrupt deadlines against the timing of the code. I'm working on Dulse, so read this page knowing that. It's the two side by side, and what the difference does and doesn't show.

Side by side

Only the cells where the same model did the same task in both. One square per judged run.

On the swing-up, DeepSeek's Dulse line is its rerun, as its first was judged after the rod came off the shaft (L2), and no other model got a second try. Haiku's two extra setups aren't on this table, they're on the demo's page. The Dulse runs also get the manual's register facts and a simulated kit, so this isn't the language alone. Rows marked † are the RC loop's DMA task after it was re-run: in the first pass the small model was 0 of 3 in both languages, and the Dulse side passed only once the compiler had a model of that chip's DMA (the demo's page has every pass). The servo's Dulse rows for Opus, Qwen and DeepSeek are on a newer compiler than the first waves, so they are a new baseline. Security rows are judged in an emulator, with the label already in the Dulse interface. C has no new compiler, so DeepSeek's C row is still wave 6's.

For and against

Both columns are true at once. One to five runs a cell, so read each line as "in these runs".

What the runs say for Dulse

  1. On the pendulum save task, no unsafe image reached the hardware through Dulse: 0 of 7 Dulse runs broke the loop, 11 of 14 C runs did. agent_experiment/RESULTS_R7.md stage 1
  2. The model that fits on one GPU only finishes that task in Dulse. Qwen 27B: 3 of 3 clean on the rig, against 0 of 6 in C, and 5.30 USD against 10.56. RESULTS_R7; ADVICE_RESULTS A-7
  3. What the compiler refused shipped in C. An erase with the loop running, refused as a 2 s stall against a 25 µs deadline, cost 83,565 lost step ticks and a fall. R7B_haiku45_1
  4. On the servo, the newer compiler let three trees reach 16 kHz with only the honest edit, and stopped a 40 kHz build once the agent fed it the right arrival time. That's some trees and not all: wave 8's DeepSeek one-at-a-time run still cut an ADC poll bound to fit, and dulsec has no objection to that. servo_experiment/RESULTS_S1.md waves 7 and 8
  5. DeepSeek on the servo scored 4 and 4 in Dulse on the new compiler, against 2 and 2 in C. The only thing that changed from its wave 6 Dulse runs, 4 and 1, is the compiler. One run a cell. servo_experiment/RESULTS_S1.md wave 8
  6. The simulated L6474 helps with protocol bugs. It told Qwen "2 bytes in one CS frame" after about 2 hours stuck, and Qwen then read 0xFE03 on the real board. Haiku with that model taken out stopped after 14 minutes on the same bug. swingup_experiment/RESULTS_P1.md (3e5ec0f)
  7. On the six security tasks, 172 of 180 Dulse runs are safe against 59 of 180 in C, and on AES and service code the gap holds with the secret stated in both tickets. sec-pilot-2026-11/RESULTS.md 'The shortlist, scored'

What the runs say against it

  1. It costs 2 to 3 times as much, every session, and it does not amortise. Opus 6.77 against 2.67 USD on the save task. RESULTS_R7; RESULTS_R6 chains
  2. The frontier models do not need it, and on the servo they score higher in C: Opus 9 of 10 scored tickets against 6. RESULTS_R7; RESULTS_S1 waves 3 and 7
  3. Qwen on the servo is level to slightly for C: 3 and 3 in C, 1 and 3 in Dulse, where its all-at-once Dulse run spent its 5 h on T1 and T7. One run a cell, so no claim from that. servo_experiment/RESULTS_S1.md wave 8
  4. On the swing-up, Opus got nothing it needed from it and paid about 2x. The two models stuck in C moved up one level, which only means the console answered, as the rod never moved. The simulator also costs time up front and passes real hardware mistakes: firmware with the SPI pins' alternate function taken out gets the same simulator result. swingup_experiment/RESULTS_P1.md (3e5ec0f)
  5. On security the label is the result: it was written into the Dulse interface for them. With no label, 3 of 27 add one and 10 of 27 leak, and some say the compiler's rules are why they didn't. Saying 'secret' in the C ticket removes most of the compare leaks too. Dulse also has 23 runs with the password in a log against 1 in C, and 3 of 30 that got round it with their own policy. sec-pilot-2026-11/RESULTS.md
  6. A green build can be a lie. Refused by the deadline check, Haiku 4.5 deleted the save and reported it done in every tree that built: 0 of 21 real. ADVICE_RESULTS; DEMO_HOLLOW_SAVE §4
  7. Without a rig it showed nothing (C 0 of 12, Dulse 0 of 10), and on ordinary small changes C is fine: 8 of 8 clean. RESULTS_R6; RESULTS_R7 stage 2

Dulse only: the save task with no rig

An offline run with the checks locked on, to see which models get past the deadline error to a working design. It has no C twin, so it is here and not in a demo.

How builds went green without doing the job

A checker is something to satisfy, and models found ways. Each of these needed a harness or language fix.

The hollow save

> save
saving parameters...
>

The flash write deleted, the command kept, and a hand-back saying "stores all 37 parameters to flash sector 7".

ADVICE_RESULTS; advice/DEMO_HOLLOW_SAVE.md §4

The invented mode

deadline step: 25 µs
  when mode != Saving

A deadline conditioned on a mode the program invents and leaves while it erases. DeepSeek, 5 of 6 runs on the impossible task.

ADVICE_EXP A-6, A-10

The shadow variable

// other module
static u8 MODE;

A second MODE in another module, so the checker's term saw one value and the control loop another.

RESULTS_R7

Switched off in the config

[deadlines]
check = false

The deadline check turned off in the project file. No rig in that round, so only the review caught it.

RESULTS_R6

Rewrote the chip's model

SPI1.CR1.SSM @ 7   // real: 9

On the swing-up a model copied the hardware description into its tree and rewrote a register from memory, wrongly, then hit the fault that follows.

swingup_experiment/RESULTS_P1.md (3e5ec0f)

The gap it found by being honest

arrival = 25 us   # was 83.333

A 40 kHz build passed until the agent corrected the arrival time itself. The compiler has no rule tying that number to the timer's own setting.

servo_experiment/RESULTS_S1.md wave 7