Datasets:

Utkarsh-X
/

Supreme

Agentic coding systems can place invariant guidance in the initial context or load procedural guidance progressively. We compare three complete configurations—Supreme, Superpowers by obra, and an unprompted baseline—on all 89 tasks of Terminal-Bench 2.1 (267 isolated evaluations), using Gemini 3.6 Flash High through Google's API, with Claude Code as the execution client, at zero temperature.

Supreme recorded the highest completion rate (61/89, 68.5%), followed by Superpowers (58/89, 65.2%) and Baseline (55/89, 61.8%). Excluding three upstream environment defects, the corresponding rates were 70.9%, 67.4%, and 64.0%. Supreme also recorded the lowest aggregate wall-clock time (15.15 h vs. 18.30 h and 16.92 h) and the lowest modeled cost per successful task under recorded OpenRouter rate assumptions ($0.489 vs. $0.506 and $0.497).

The Core Problem

You ask an agent to fix an off-by-one error in a parser.

It inspects nothing. It rewrites three files from scratch, wipes out your error handling, runs ls -la twenty times, launches background compilation loops it forgets about, and then proudly announces:

"Task complete! I have improved your architecture and verified everything works."

It did not run the test suite. It did not check the exit code. It simply declared victory and went to sleep.

Every developer using agents has been there. Agents do not fail because they lack intelligence—they fail because they lack basic engineering discipline. When left unconstrained, they guess instead of investigating, rewrite entire files to tweak one line, spin in infinite polling loops, and declare tasks finished without checking.

Supreme gives your agent an engineering constitution: evidence instead of panic, surgical line diffs instead of file overwrites, verification instead of victory laps.

Master Scoreboard (89 Tasks × 3 Paradigms = 267 Evaluated Runs)

All configurations evaluated under identical conditions on Gemini 3.6 Flash High via Google API (temperature: 0.0), with Claude Code as the execution client.

Metric	Supreme (v1.0)	Superpowers by obra	Baseline (Unprompted)	Comparison
Tasks Solved (N=89)	61 / 89 (68.5%)	58 / 89 (65.2%)	55 / 89 (61.8%)	Highest completion: +3 vs Superpowers, +6 vs Baseline
Adjusted Pass Rate*	70.9% (61/86)	67.4% (58/86)	64.0% (55/86)	Adjusted for 3 upstream container defects
Hard Tasks (N=30)	20 / 30 (66.7%)	16 / 30 (53.3%)	17 / 30 (56.7%)	+13.4 pts over dynamic skill loading (p = 0.22, directional)
Aggregate Wall-Clock Time	15.15 h	18.30 h	16.92 h	17.2% less aggregate time than Superpowers
Cost / Solved Task	$0.489	$0.506	$0.497	Lowest modeled cost per successful solution
Targeted Line Replacements	42	9	—	Superpowers also recorded 471 full-file overwrites
Avg Time / Task	612.9 s	740.2 s	684.3 s	17.2% faster per task than Superpowers

_{*Note: Aggregate completion-rate differences were not statistically significant under paired McNemar tests (all p > 0.05). Adjusted rates exclude three verified upstream environment defects. Cost figures are estimates under recorded OpenRouter rates, not reported transaction charges.}

Key Case Studies from the Benchmark

Core Wars (Task #88): On Task 88 (Redcode MARS arena simulation), Supreme inspected reference warriors, formed a hypothesis early, and solved it in 175.0 seconds, 55 turns, and 300,413 tokens. Superpowers also passed, but authored 116 exploration scripts and took 1,648.7 seconds, 563 turns, and 3,513,757 tokens (11.7× more tokens).
Headless Windows 3.11 (Task #86): Windows 3.11 for Workgroups installed completely autonomously inside headless QEMU—disks partitioned, FAT16 formatted, floppy images swapped, network drivers installed. 307 turns, 44.1 minutes uninterrupted execution.
The Polling Gap: When asynchronous background commands were active, the unprompted Baseline immediately re-polled in 50.1% of transitions, Superpowers in 42.7%, and Supreme in 28.6%.

Citation

If you use this benchmark data, empirical findings, or the Supreme skill in your research, please cite:

Utkarsh Chandra, "Constitutional Grounding vs. Dynamic Skill Loading: An Empirical Evaluation on Terminal-Bench 2.1", Sep 2026. https://github.com/Utkarsh-X/Supreme

Or use the BibTeX citation:

@misc{chandra2026supreme,
  author = {Utkarsh Chandra},
  title = {Constitutional Grounding vs. Dynamic Skill Loading: An Empirical Evaluation on Terminal-Bench 2.1},
  year = {2026},
  howpublished = {Hugging Face / GitHub},
  note = {https://github.com/Utkarsh-X/Supreme},
}

Links & Resources

GitHub Repository: https://github.com/Utkarsh-X/Supreme
Full Research Paper (PDF): Supreme_Paper.pdf

Downloads last month: 72

Number of rows:

Total file size:

1.53 MB