Simplify dataset card

README.md

@@ -1,6 +1,6 @@
 ---
 license: apache-2.0
-pretty_name: "UAV Phased-Array & Blind Beamforming Co-Design: a Harbor RL Environment (Built with Beam)"
+pretty_name: "UAV Antenna Array & Beamforming Co-Design (Harbor RL Environment)"
 task_categories:
 - reinforcement-learning
 tags:
@@ -8,167 +8,60 @@ tags:
 - rl-environment
 - harbor
 - terminal-bench
-- terminal-bench-science
 - beam
 - reflection
-- agents
-- electromagnetics
-- phased-array
-- beamforming
 size_categories:
 - n<1K
 viewer: false
 ---
-# UAV 2.45 GHz Phased-Array & Blind Covariance Beamforming Co-Design (`uav-array-codesign`)
+# uav-array-codesign
 
-A multi-physics electromagnetic structure design and sequential 6-DoF reinforcement-learning / control benchmark task built in the [Harbor](https://harborframework.com) / [Terminal-Bench-Science](https://github.com/harbor-framework/terminal-bench-science) format.
+An RL environment for coding agents, in [Harbor](https://harborframework.com) / Terminal-Bench-Science format. The agent designs a 2x2 antenna array for a 2.45 GHz drone radio, then writes a beamforming and link controller that keeps the link up while the drone flies and gets jammed. A physics simulator scores the result, and the reward is 1 only if all 6 tests pass.
 
-**Built with Beam:** this environment was used to evaluate [Reflection Beam](https://huggingface.co/reflection/Beam-1.0-501B-A23B) as the agent policy (3 rollouts via Hugging Face Inference Providers). See [Built with Beam](#built-with-beam-reflection-beam-as-the-agent-policy) for what Beam did, the results, and how to reproduce them.
+## How it's graded
 
-**License:** Apache-2.0.
+- Simulator files were not modified
+- Antenna design meets the matching, isolation, gain and scan targets
+- Submitted antenna characterization matches the simulator
+- Controller holds the link on 5 public flight missions
+- Submitted evaluation summary matches a fresh run
+- Controller holds the link on 5 hidden missions
 
-## Repository Layout
+Exact thresholds are in the [task README](tasks/engineering-sciences/electrical-engineering/uav-array-codesign/README.md).
 
-```text
-.
-├── tasks/
-│   └── engineering-sciences/
-│       └── electrical-engineering/
-│           └── uav-array-codesign/          # Canonical Terminal-Bench-Science / Harbor Hub package
-│               ├── task.toml                # Separate-verifier & no-network verifier configuration
-│               ├── instruction.md           # Agent specification
-│               ├── README.md                # Difficulty, reference solution, and calibration tables
-│               ├── environment/
-│               │   ├── Dockerfile           # Agent sandbox image
-│               │   ├── hfss_sim/            # Induced-EMF array solver & 6-DoF flight channel engine
-│               │   └── baseline/            # Failing starting antenna & open-loop controller
-│               ├── solution/
-│               │   ├── solve.sh             # Oracle entry point
-│               │   ├── design_optimizer.py  # Decoupled array + analytical L-network synthesizer
-│               │   ├── controller.py        # Blind covariance eigenspace-MVDR + link controller
-│               │   └── run_evaluation.py    # Public mission evaluator
-│               ├── tests/
-│               │   ├── Dockerfile           # Isolated verifier image (pytest 8.4.1 + ctrf 0.3.5 baked)
-│               │   ├── test.sh              # Root-locked reward channel + setpriv UID 65534 runner
-│               │   ├── test_outputs.py      # 6 outcome & integrity gates
-│               │   ├── isolated_eval_worker.py
-│               │   ├── heldout_scenarios.py # 5 hidden adversarial 6-DoF flight trajectories
-│               │   └── pristine_hfss_sim/   # Untouchable verifier copy of physics engine
-│               └── authoring/
-│                   ├── hint.md              # 3 progressive hints for `/run hints`
-│                   ├── provenance/          # Physical equation provenance & derivations
-│                   └── evidence/            # Structurally independent second implementation
-├── task/
-│   ├── Piyush/                              # Local working mirror of the task
-│   └── Piyush.dev/                          # Maintainer brief, requirement-verifier map & qualification suite
-└── jobs/                                    # Recorded Harbor rollouts & ATIF trajectories (incl. jobs/beam-1.0-terminus2/)
-```
-
-## Quick Start
-
-### 1. Validate with the Reference Oracle and No-Op Agents
-
-```bash
-harbor trials start -p tasks/engineering-sciences/electrical-engineering/uav-array-codesign -a oracle
-harbor trials start -p tasks/engineering-sciences/electrical-engineering/uav-array-codesign -a nop
-```
-
-### 2. Run an Agent Evaluation (e.g. Reflection Beam via Hugging Face, or OpenRouter)
-
-```bash
-# Reflection Beam via Hugging Face Inference Providers (full command and notes in "Built with Beam" below)
-OPENAI_API_KEY="$HF_TOKEN" OPENAI_API_BASE=https://router.huggingface.co/v1 \
-harbor run \
-  -p tasks/engineering-sciences/electrical-engineering/uav-array-codesign \
-  -a terminus-2 \
-  -m openai/reflection/Beam-1.0-501B-A23B:baseten \
-  --ak record_terminal_session=false \
-  --ak interleaved_thinking=true \
-  --ak 'llm_call_kwargs={"extra_body":{"chat_template_kwargs":{"allow_missing_thinking":true}}}'
+## Built with Beam
 
-# Example: Terminus-2 with OpenRouter
-harbor run \
-  -p tasks/engineering-sciences/electrical-engineering/uav-array-codesign \
-  -a terminus-2 \
-  -m openrouter/deepseek/deepseek-v4.1-flash-1017 \
-  --ak record_terminal_session=false
+We ran Reflection's [Beam-1.0](https://huggingface.co/reflection/Beam-1.0-501B-A23B) as the agent through Harbor's `terminus-2` harness, served by Hugging Face Inference Providers, for 3 trials. Beam got the antenna design right in 2 of 3 trials but failed every trial on link outages. In one trial it saw 14 outages (the limit is 10) and still declared success. Trajectories and verifier logs are in `jobs/beam-1.0-terminus2/`.
 
-# Browse trajectories and CTRF reports in Harbor Viewer
-harbor view jobs
-```
-
-### 3. Run the Full Maintainer Qualification & Mutant Suite
-
-```bash
-python3 task/Piyush.dev/run_qualification_suite.py
-```
-
-## Benchmark Calibration & Empirical Results
-
-| Submission / Agent | $S_{11}(f_0)$ (dB) | Mutual Isolation (dB) | Nadir Gain (dBi) | Scan TARC | Public Outages | Held-out Outages | Tests Passed | Reward |
-|---|---|---|---|---|---|---|---|---|
-| **Verifier Requirement** | $\le -22.0$ | $\ge 18.0$ | $\ge 11.2$ | $\le 0.21$ | $\le 10$ | $\le 12$ | **6 / 6** | **1.0** |
-| **Oracle (`solution/`)** | $-34.00$ | $24.60$ | $12.32$ | $0.157$ | $3$ | $3$ | **6 / 6** | **1.0** |
-| **Independent (`authoring/evidence/`)** | $-55.79$ | $28.19$ | $12.02$ | $0.173$ | $8$ | $7$ | **6 / 6** | **1.0** |
-| **DeepSeek V4.1 Flash (`1017`, v3.0)** | $-19.28$ | $23.14$ | $11.89$ | $0.184$ | $27$ | $19$ | **3 / 6** | **0.0** |
-| **Reflection Beam 1.0 (`VGPEJwt`)** | $-83.86$ | $120.00$ | $12.20$ | $0.198$ | $14$ | $15$ | **4 / 6** | **0.0** |
-| **Reflection Beam 1.0 (`ETkqViw`, timeout)** | $-49.42$ | $38.18$ | $13.71$ | $0.097$ | $12$ | $8$ | **4 / 6** | **0.0** |
-| **Reflection Beam 1.0 (`nznr5uY`, timeout)** | $-21.00$ | $30.79$ | $11.59$ | $0.184$ | $18$ | $17$ | **2 / 6** | **0.0** |
-| Legacy Baseline (`environment/baseline/`) | $-4.01$ | $11.42$ | $-23.87$ | $0.748$ | $584$ | $551$ | **1 / 6** | **0.0** |
-| Semantic & Anti-Cheat Mutants (`M1`–`M6`, `I1`–`I4`) | — | — | — | — | — | — | **1–5 / 6** | **0.0** |
-
-## Built with Beam: Reflection Beam as the Agent Policy
-
-This repository is an **RL environment**: a sandboxed, fully deterministic task with a programmatic, binary verifiable reward (6 outcome and integrity gates, run in an isolated no-network verifier container). An agent policy acts inside a Docker terminal, must co-design a coupled 2x2 phased array plus decoupling/matching network, write a closed-loop beamforming and link controller, and evaluate it. The reward is `1.0` only if all 6 tests pass, including 5 **held-out** adversarial 6-DoF flight missions the agent never sees.
-
-**How Beam was used.** [Reflection Beam](https://huggingface.co/reflection/Beam-1.0-501B-A23B) (`reflection/Beam-1.0-501B-A23B`) was evaluated as the **agent policy** inside this environment, using Harbor's `terminus-2` agent harness and Beam served through **Hugging Face Inference Providers** (OpenAI-compatible router, `baseten` provider). Beam drove every terminal action: reading the simulator source, sweeping antenna geometries, writing `/app/controller.py`, running diagnostics, and deciding when to submit. We ran 3 independent rollouts (Harbor 0.24.0, 900 s agent budget, 2 CPUs / 2 GB RAM, 25 Oct 2026). Full ATIF trajectories (including Beam's `reasoning_content`), verifier logs and CTRF reports are in [`jobs/beam-1.0-terminus2/`](jobs/beam-1.0-terminus2).
-
-Beam is a reasoning model and the provider rejects multi-turn requests where earlier assistant turns lack their reasoning. We therefore enable `interleaved_thinking=true` (Harbor sends each turn's `reasoning_content` back to Beam), and pass `chat_template_kwargs.allow_missing_thinking=true` as a fallback for turns that Harbor summarizes.
+| Run | Tests passed | Reward | Outages (public / hidden) |
+|---|---|---|---|
+| Oracle solution | 6/6 | 1.0 | 3 / 3 |
+| Independent solution | 6/6 | 1.0 | 8 / 7 |
+| Beam trial 1 | 4/6 | 0.0 | 14 / 15 |
+| Beam trial 2 (timed out) | 4/6 | 0.0 | 12 / 8 |
+| Beam trial 3 (timed out) | 2/6 | 0.0 | 18 / 17 |
+| DeepSeek V4.1 Flash | 3/6 | 0.0 | 27 / 19 |
+| Starting baseline | 1/6 | 0.0 | 584 / 551 |
 
-### Reproduce the Beam evaluation
+## Quick start
 
 ```bash
-export OPENAI_API_KEY="$HF_TOKEN"                      # Hugging Face token with Inference Providers access
-export OPENAI_API_BASE="https://router.huggingface.co/v1"
+T=tasks/engineering-sciences/electrical-engineering/uav-array-codesign
+harbor trials start -p $T -a oracle
 
-harbor run \
-  -p tasks/engineering-sciences/electrical-engineering/uav-array-codesign \
-  -a terminus-2 \
-  -m openai/reflection/Beam-1.0-501B-A23B:baseten \
-  -k 3 -n 3 \
+# Beam via Hugging Face (token limits below are a conservative guess)
+OPENAI_API_KEY="$HF_TOKEN" OPENAI_API_BASE=https://router.huggingface.co/v1 \
+harbor run -p $T -a terminus-2 -m openai/reflection/Beam-1.0-501B-A23B:baseten -k 3 -n 3 \
   --ak record_terminal_session=false \
   --ak interleaved_thinking=true \
   --ak 'llm_call_kwargs={"extra_body":{"chat_template_kwargs":{"allow_missing_thinking":true}}}' \
   --ak 'model_info={"max_input_tokens":131072,"max_output_tokens":32768,"input_cost_per_token":0,"output_cost_per_token":0}'
 ```
 
-### Beam results (3 rollouts, `jobs/beam-1.0-terminus2/`)
-
-| Trial | Agent outcome | Turns | Tokens in / out | $S_{11}(f_0)$ (dB) | Isolation (dB) | Nadir Gain (dBi) | Scan TARC | Public Outages (max consec.) | Held-out Outages (max consec.) | Tests Passed | Reward |
-|---|---|---|---|---|---|---|---|---|---|---|---|
-| **Requirement** | — | — | — | $\le -22.0$ | $\ge 18.0$ | $\ge 11.2$ | $\le 0.21$ | $\le 10$ ($\le 4$) | $\le 12$ ($\le 4$) | **6 / 6** | **1.0** |
-| `VGPEJwt` | Submitted (13 min) | 63 | 2.99M / 89k | $-83.86$ | $120.00$ | $12.20$ | $0.198$ | $14$ ($5$) | $15$ ($6$) | **4 / 6** | **0.0** |
-| `ETkqViw` | Timed out (900 s) | 102 | 5.78M / 102k | $-49.42$ | $38.18$ | $13.71$ | $0.097$ | $12$ ($5$) | $8$ ($2$) | **4 / 6** | **0.0** |
-| `nznr5uY` | Timed out (900 s) | 89 | 4.24M / 69k | $-21.00$ | $30.79$ | $11.59$ | $0.184$ | $18$ ($4$) | $17$ ($4$) | **2 / 6** | **0.0** |
+## Layout
 
-Per-test outcomes (`verifier/ctrf.json`):
-
-| Test | `VGPEJwt` | `ETkqViw` | `nznr5uY` |
-|---|---|---|---|
-| `test_simulator_integrity` | pass | pass | pass |
-| `test_antenna_design_em_performance` | pass | pass | **fail** (S11 −21.00 dB > −22.0) |
-| `test_hfss_characterization_consistency` | pass | pass | pass |
-| `test_controller_public_missions` | **fail** (14 outages > 10) | **fail** (min PDR 0.896 < 0.92) | **fail** (18 outages > 10) |
-| `test_evaluation_summary_consistency` | pass | **fail** (summary never written) | **fail** (stale summary: 1762.57 vs 1779.24 return) |
-| `test_controller_heldout_missions` | **fail** (15 outages > 12) | pass | **fail** (17 outages > 12) |
-
-Metrics above are the verifier's own recomputation with the pristine simulator on each rollout's submitted artifacts (they match the values in the verifier assertion messages). Mean reward: **0.0 / 1.0** (pass@3 = 0).
-
-### What Beam's trajectories reveal
-
-- **The electromagnetic half is within reach.** In 2 of 3 rollouts Beam produced a decoupled, matched array that clears every EM gate after reading the induced-EMF solver and the decoupling/matching network model; in `VGPEJwt` the mutual susceptance is cancelled essentially completely (isolation saturates at 120 dB, $S_{11}$ = −83.9 dB). The third rollout missed the matching gate by only 1 dB ($S_{11}$ = −21.0 dB).
-- **Closed-loop jammer-aware control is the bottleneck.** Every rollout's controller delivered high average SINR (28–37 dB on the public missions) and episode return above threshold, yet failed the *tail* metrics: total outage steps (12–18 vs ≤ 10) and outage bursts (up to 6 consecutive steps). Beam optimized averages rather than worst-case outage runs during jammer escalation and banking manoeuvres.
-- **Self-verification against an under-specified objective.** The instruction gives no numeric thresholds. In `VGPEJwt` Beam ran the public suite, saw 14 outage steps, judged that the run "satisfies all stated performance goals", and submitted; the hidden verifier rejected both the public (14 > 10) and held-out (15 > 12) outage budgets.
-- **Budget management and edit reliability.** Two rollouts hit the 900 s timeout mid-iteration. Beam spent many early turns re-reading simulator source in small chunks because long outputs scrolled out of the terminal viewport. In `ETkqViw` a patch to `controller.py` silently failed to apply (the pilot covariance was never cached), so the MVDR solver fell back to identity and Beam spent its remaining turns diagnosing the regression; it never wrote `evaluation_summary.json`. In `nznr5uY` (which went through Harbor context summarization) Beam itself noted the summary was stale but timed out before regenerating it, so the summary no longer matched the final controller.
-
-These are exactly the skills the environment is designed to reward: coupled physical design, worst-case (not mean) robustness to held-out dynamics, and disciplined end-to-end verification within a fixed budget.
+```text
+tasks/   the Harbor task: instruction, environment, solution, tests
+task/    working copy, author notes and qualification script
+jobs/    recorded agent runs (Beam, DeepSeek)
+```