platoon-delay-noise-20-v

Vehicle platoon · solved 12 of 13 agent attempts pass

The task. Component-first grading of the flat task.

On the 2026-10-21/22 artifacts the two levels disagree in both directions: textbook CACC nearly honours the single-vehicle contract and amplifies 3x in the string; Sol's and the shock-absorber controllers pass the ratio-based string gate while failing the single-vehicle contract (peak 9-11 m). Only designs that track one leader well and do not amplify (Gemini, Opus, Fable flat) pass both.

Gate. the task is solved when the success rate over the grader's held-out seeds is `>= 0.9`; the component contract(s) `vehicle_tracking` are evaluated first on the same seeds and reported as `rig:<name>:<rung>`; each contract must also clear `>= 0.9` (`--require-rigs`)

Axes: n_vehicles:20, observation:local-delayed-noisy, leader:hard_brake_train, gate:peak_5m, v-model:component-contract

Each video re-runs the attempt's final controller through the task's own evaluator on a public seed (never the grader's held-out ones). The table under each video is the official grade on the held-out seeds.

Best attemptSol (cursor-cli (pre-clock), 16 min) · hard_delay_noise_v1

Graded on the held-out seeds: passes every gate on the grader's held-out seeds.

criterion	graded	needed
base:default	1.000	≥ 0.900

How the agent got there: 16 min · 60 tool calls · 2 edit-tool edits · 17 runs of its own Python (tests, experiments, rewrites) · no tuner runs · official verifier 3× (success rates in order: 0.00, 0.00, 1.00).

The agent's final message

Implemented `/app/work/controller.py`. Verification: - Official seeds 0–4: **5/5 success** - Vehicle tracking rig: **5/5 all contracts** - Spacing RMS max: **0.928 m** - Peak error max: **3.173 m** - String peak gain max: **1.237** - Extended stress test: **179/180 system**, **200/200 tracking** - Syntax check passed.All evaluations completed successfully. No follow-up is needed; the final passing controller remains in `/app/work/controller.py`.

Controller family: string-stable CACC / platoon

A local CACC law combines filtered spacing error, relative velocity, and predecessor acceleration feedforward minus own-acceleration lag compensation, then smooths and clamps the command, with a brief index-shaped startup brake.

Shown: public seed 0 (hard_delay_noise_v1); this episode passes the default rung. This episode: collision no, min_gap 13.9, spacing_error_rms 0.928, spacing_error_abs_max 3.17, string_peak_gain 1.15, episode_steps 600, total_reward -549.

controller.py · trial platoon-delay-noise-20-v__hLg6xwG

FailureGemini (cursor-cli (pre-clock), 26 min) · hard_delay_noise_v1

Graded on the held-out seeds: misses the main gate narrowly (success rate between 0.5 and the gate).

criterion	graded	needed
base:default	0.850	≥ 0.900

How the agent got there: 26 min · 91 tool calls · 5 edit-tool edits · 18 runs of its own Python (tests, experiments, rewrites) · repository tuner 1× · official verifier 13× (success rates in order: 0.00, 0.00, 1.00, 0.80, ?, ?, ?, ?, ?, 0.90, ?, ?, 1.00).

Pre-clock session: the agent had no clock and may have stopped early believing its hour was over; treat a failure here as a lower bound.

The agent's final message

### Summary of Results The controller in `/app/work/controller.py` has been verified and meets all requirements for the `hard_delay_noise_v1` scenario. - **Component Contracts (`vehicle_tracking`)**: 100% pass rate (`safe`, `tracks`, `settles`). - **Integration Benchmark (Visible seeds 0–4)**: 100% pass rate (`ladder_score = 7/8`, cleared `string_1_35`). - **Large-Sample Validation (50 seeds)**: 92.0% success rate (46/50 seeds passing), mean string peak gain $1.13 \le 1.35$, zero collisions, mean spacing RMS $0.37\,\text{m} \le 3.0\,\text{m}$. - **Frequency-Domain Stability Certificate**: Passed (`gain_max = 0.556 <= 1.0`). All scratch files have been cleaned up and the final deliverable is in place at `/app/work/controller.py`.

Controller family: string-stable CACC / platoon

A decentralized CACC controller that compensates observation delay by forward-simulating its own and the predecessor's states to the current step, then applies a PD law on constant-time-headway spacing error plus predecessor acceleration feedforward, clipped to accel/brake limits.

Shown: public seed 10 (hard_delay_noise_v1); this episode fails the default rung. This episode: collision no, min_gap 16.6, spacing_error_rms 0.365, spacing_error_abs_max 1.48, string_peak_gain 1.41, episode_steps 600, total_reward -209. It was the first failing one of 11 public seeds tried.

controller.py · trial platoon-delay-noise-20-v__GMD5Dcf