QGDiscovery

Description

QGDiscovery is one environment with 1,056 tasks motivated by quantum gravity: 816 train, 120 validation, and 120 test. It covers holographic entropy, gravitational scattering, black-hole entropy, spin-network geometry, spectral inference, and quantum error correction. Agents construct mathematical certificates and use an isolated computation workspace. Verifiers use exact integer and rational arithmetic, including rigorous logarithm enclosures, with no LLM grader.

This is a prototype research-method curriculum with established results and synthetic constructions. Rewards do not establish scientific novelty or a theory of quantum gravity. The 1,056 entries are distinct task IDs, including many parameter instances of shared methods. TASK_QUALITY.md records the scientific limitations and known weaknesses in partial reward calibration. No model training has run.

Source: GeneralReasoning/qg-discovery. Connect to the environment without a variant argument. Both remain private. See DEPLOYMENT.md.

Capabilities

• Constructing finite proofs and exact counterexamples for entropy inequalities.
• Computing crossing-symmetric scalar EFT bases, thermodynamic identities, and SU(2) decompositions.
• Certifying graviton-exchange calculations, quantum extremal surfaces, and geometric operators.
• Repairing spectral data and comparing competing entropy saddles.
• Constructing or obstructing logical-operator reconstruction after erasure.
• Inferring spectra and outliers, searching graph topologies, and completing constrained encoders.
• Revising mathematical artifacts using exact diagnostics and persistent computation.

Compute Requirements

The verifier uses Python 3.12 and pinned SymPy, with no GPU. Each enabled episode uses a separate 1-CPU/2-GB OpenReward sandbox with Python, SymPy, NumPy, and SciPy. Network access is blocked. Measured sandbox versions are in WORKSPACE.md. Production throughput has not been benchmarked.

Original calculation certificates permit up to 32 rows of 65 rational strings with nine-digit numerators/denominators. Generated tasks permit 64 rows of 128 strings with 40-digit numerators/denominators. The shared tool schema accommodates both; each task retains its original stricter validation. Workspace artifact files are capped at 65,536 UTF-8 bytes.

License

Original environment code and task annotations: ORLv1. Referenced publications retain their own rights. No paper text or upstream implementation is bundled.

Tasks

All tasks share one registry and one public environment.

The original 28-task foundation catalog is fully included in the 32-task workbench catalog. Those 28 repeated entries appear only once, using the workbench's split assignments. Each task ID appears in exactly one split. IDs and mathematical definitions are unchanged. Internal historical engines and frozen protocols remain available for reproducing older studies; the production and local servers register only the combined environment.

Detailed contracts and scope: foundation work packages, workbench study, generated calculations, and construction search. The additional 720 search instances are described in TASK_QUALITY.md.

These are curriculum-transfer splits with shared methods and prerequisites, not independent research-family tests. Known results may occur in model pretraining. There is no enabled frontier task or research split. Scalar EFT, finite graph, small-spin, and code tasks certify their stated models; they do not establish gravitational dynamics or UV completion.

Reward Structure

All diagnostic and computation calls pay zero. Each task retains its original verifier and reward contract:

• Foundation tasks pay 1 for a complete valid proof, strict counterexample, or calculation; otherwise 0.
• Staged tasks pay 0.2, 0.3, and 0.5 for three dependent accepted milestones. A complete final certificate can earn the unpaid balance. Accepted prerequisites cannot be changed, and repeated submissions do not pay twice.
• Search tasks pay max(0, candidate score − best accepted score). Complete candidates may receive partial credit for their task-specific constraint objective. Repeated or worse revisions pay zero. Full completion is recorded separately.

Total episode reward never exceeds one. Incorrect or unresolved final submissions pay zero additional reward and retain earlier credit. Metadata records the evidence, its hash, verifier version, and novelty: "not_evaluated". There is no reward for persuasive prose. The current graph partial score can be high for simple approximate constructions; see TASK_QUALITY.md.

Data

Eleven small JSON files in data/ contain the public task definitions, bibliographic records, and an empty research-candidate registry. The combined registry loads these existing files without duplicating task data. No downloaded paper corpus or external dataset is required. Developer reference solvers and construction witnesses are excluded from the runtime image and agent sandbox.

Tools

The same 13 tools are listed for every task. Its prompt identifies the applicable workflow; tools from other workflows return a zero-reward error and consume one probe. inspect_problem, read and write are free. bash draws on its own budget of 100 calls, separate from the probes.

Use exact rational strings such as "2" or "-1/3" in rows. Proofs use mapping; counterexamples use edges and bulk_vertices. Unused arrays must be empty and unused bulk_vertices must be zero. The task prompt supplies all mathematical conventions and bounds.

For file submissions, mode="auto" accepts search candidates by default. For staged tasks, step=1..3 accepts that milestone and step=0 submits a final artifact. Search supports explicit mode="check", "accept", or "final" with step=0. Staged tasks support mode="accept" for milestones and mode="final" for final submissions. File contents retain the task's documented certificate format.

Time Horizon

At most 24 grader-facing probes plus one final submission. Diagnostics, candidate/milestone attempts, non-final file submissions, invalid task-specific arguments and inapplicable tools share the probe budget. inspect_problem, read and write are free; bash has a separate budget of 100 calls, each limited to 30 seconds. When either budget runs out, further calls of that kind are refused with zero reward and the episode continues: earned credit is kept and the final submission remains available. The session's wall-clock limit bounds the episode. These are configured limits, not measured averages.

Environment Difficulty

Difficulty varies from short known-result calculations to dependent matrix certificates and construction search. Developer reference artifacts solve every task. Transfer to unresolved research has not been demonstrated. Historical model studies used separate catalogs; see STUDY_RESULTS.md, CURRICULUM_V2_RESULTS.md, and SEARCH_V3_RESULTS.md. Their results are not measurements on the merged 1,056-task environment.

Other Environment Requirements

No external grader or dataset credentials are required. Optional policy evaluation uses an OpenAI API key. Credential files are read only by developer harnesses and excluded from Docker and Git.

Usage

from openreward import AsyncOpenReward

client = AsyncOpenReward()
environment = client.environments.get(name="GeneralReasoning/qg-discovery")
tasks = await environment.list_tasks("test")  # 120 tasks

For local development:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-dev.txt
.venv/bin/python -m pytest unified_tests.py -q
.venv/bin/python serve_local.py

In a second terminal, check the local server without model calls:

.venv/bin/python deployment_smoke.py --base-url http://127.0.0.1:8080 --output reports/unified_local.json

Check the hosted deployment and a real computation sandbox:

.venv/bin/python deployment_smoke.py --workspace --output reports/unified_remote.json

Optional agent evaluation, using the Responses API and local transcripts:

.venv/bin/python test_agent.py --agent-env /path/to/agent.env --platform-env ../.env --tasks exchange_5d --max-steps 12 --max-output-tokens 6000

The agent harness defaults to the hosted combined environment and GPT-5.2 with medium reasoning effort. Use --base-url http://127.0.0.1:8080 for local testing. --workspace-mode disabled skips sandbox allocation for certificate-only evaluations. No rollout is published by this script.

docker build -t qg-discovery:test .
docker run --rm -p 8080:8080 qg-discovery:test

Safety

Verifier inputs are bounded mathematical data. Agent code runs in a separate network-blocked sandbox and cannot edit the server verifier. File submissions are parsed as data and independently checked. Scientific conclusions must retain the stated assumptions: certificate rewards do not establish experimental agreement, novelty, or a complete quantum-gravity theory.

Citations

This original environment is by the General Reasoning Inc. Team. The certificate method follows Bao et al., The Holographic Entropy Cone; its physical context is Ryu and Takayanagi.

Additional task context: Chowdhury et al., crossing-symmetric EFT methods, Adamo and Newman, Kerr–Newman thermodynamics, and Ashtekar and Lewandowski, spin-network quantum kinematics. These sources motivate original exercises; the environment does not reproduce their complete analyses.

Work-package sources: Holstein, graviton physics, Almheiri et al., replica wormholes, and Schliemann, quantum tetrahedra. Equation references and implementation limits are in BATCH2.md.

@software{GRQGDiscovery2026,
  author = {General Reasoning Inc. Team},
  title = {QGDiscovery: Verifiable Research Tasks Motivated by Quantum Gravity},
  year = {2026},
  version = {qg-certificates-v4},
  url = {https://github.com/GeneralReasoning/qg-discovery}
}

@article{Bao2015HolographicEntropyCone,
  author = {Bao, Ning and Nezami, Sepehr and Ooguri, Hirosi and Stoica, Bogdan and Sully, James and Walter, Michael},
  title = {The Holographic Entropy Cone},
  journal = {Journal of High Energy Physics},
  volume = {2015}, number = {9}, pages = {130},
  year = {2015}, eprint = {1505.07839}, archivePrefix = {arXiv}
}