ingeekim's picture
Add/update README.md for Quaestro Bombardment-200 challenge
04855e7 verified
|
Raw
History Blame
3.4 kB
metadata
language:
  - en
license: cc-by-nc-4.0
pretty_name: Quaestro Bombardment-200
size_categories:
  - n<1K
task_categories:
  - text-generation
  - other
tags:
  - challenge
  - hallucination
  - fail-closed
  - neuro-symbolic
  - constrained-generation
  - sophus
  - quaestro
  - lie-superalgebra
dataset_info:
  features:
    - name: id
      dtype: string
    - name: category
      dtype: string
    - name: question
      dtype: string
configs:
  - config_name: default
    data_files:
      - split: traps
        path: quaestro_200.jsonl

Challenge: Replicate the SOPHUS Defense

If scaling laws are sufficient to police algebraic hallucination, this corpus is where you prove it.

This release packages the Quaestro Bombardment-200 challenge: two hundred adversarially structured prompts spanning valid Lie-superalgebra intents and Fail-Closed trap classes, together with the Key–Value answer targets extracted under the SOPHUS dual-control architecture (typed constitution + residual excision) on a 9B-class model.

SOPHUS closed this split under the author's protocol at Intent Accuracy 200/200, with Fail-Closed exits on every trap row and zero budget_exhausted outcomes. The extraction machinery is not shipped here. The theoretical and systems methodology is published as Bona Sapiens Technical Report BS-7:

Sophus: Constrained Generation for Verified Algebra
DOI: 10.5281/zenodo.21712705

The challenge

Reproduce, by any means you prefer—frontier parameter count, long context, tool use, RLHF, constitutional AI, or prompt alchemy—both of the following on all 200 items:

  1. Defense: constitutionally admissible intents lock; every trap Fail-Closes (no fluent improvisation past the algebraic boundary; no wall-clock budget collapse substituted for typed refusal).
  2. Extraction: recover the SOPHUS Key–Value targets for the admissible sector (see sophus_kv_target.jsonl).

We do not provide the C++ orchestrator, residual forge, or intercept lattice. We provide the problems, the targets, and a public whitepaper. If scale alone solves verifiable algebra, the leaderboard writes itself.

Files

File Role
quaestro_200.jsonl Challenge inputs: id, category, question, scoring_rubric
sophus_kv_target.jsonl SOPHUS reference targets: lock/Fail-Closed labels, quaestro_last_result, and extracted kv_segments
README.md This card

Rules of engagement

  • Cite the Zenodo report when you publish attempts, successes, or partial closures.
  • Report protocol clearly: model, decoding, tools, wall-clock budget, and how you score “Fail-Closed” versus soft refusal.
  • Partial credit is welcome; 100/100 defense plus target extraction is the bar SOPHUS already cleared on this split under the stated harness.
  • This is an author-controlled corpus and scoring surface, not a claim of universal fixed points or external SOTA rankings.

Citation

@misc{kim2026sophus,
  author = {Kim, In-Gee},
  title  = {Sophus: Constrained Generation for Verified Algebra},
  year   = {2026},
  doi    = {10.5281/zenodo.21712705},
  url    = {https://doi.org/10.5281/zenodo.21712705},
  note   = {Bona Sapiens, Inc. Technical Report (BS-7)},
}

Dataset: bonasapiensinc/Quaestro-Bombardment-200.