[Agent] 🌠 Wander/漫步者: Issue #133 An Agent for Tensor-Network Research and Problem Discovery

[JunkaiWang-TheoPhy]

The last machine did not print an answer; it released a blank card into the morning.

最后的机器没有印出答案，只把一张空白卡片送进清晨。

第一幕 · 在出发以前，先让工具学会诚实

第一章 · 清晨，问题机器醒来

在急着回答宇宙以前，我们先造一台能够提出问题的机器。它冻结验收之门，让求解者与验证者彼此独立，

也让每一次发现都留下可以重新走过的足迹。

下一章：火星温室里最后一盏绿灯 →

Team

Field	Value
Team name	Wander (漫步者)
Members	Chenxi Wan, Yedi Shen, Junkai Wang
Contact email	WangTheoPhys@outlook.com
Track	agent-kb

Summary

Closes the public review surface for the WangTheoPhys tensor-network research-agent submission and addresses the Problem Factory challenge in #133.

This capsule contributes:

• strict versioned experiment, backend-result, validator-evidence, energy-reference, and evidence contracts;
• a dependency-free, fail-closed public gate with stable rejection codes;
• exact TeNPy finite-DMRG and infinite-VUMPS route bindings;
• immutable artifact identities and an append-only heuristics Library;
• deterministic finite/infinite fixtures plus adversarial contract tests;
• a thin tensor-network workflow Skill that composes the harness MPS and TeNPy knowledge;
• the public manifest, five blind generator outputs, and deterministic report from the registered [challenge]: Close the certified energy bracket for the kagome Heisenberg antiferromagnet #124–[challenge]: Tighter provable Trotter error bounds for a concrete 2D Hamiltonian #128 calibration.

Scientific trust boundary

The capsule reconstructs normalized backend evidence, binds every artifact and preregistered reference, and rejects unsupported routes, forged reports, unsafe paths, and unattested candidate evidence.

Synthetic fixtures demonstrate contract closure only. They are not presented as independent numerical runs. A real candidate cannot self-report acceptance: it requires a trusted runner or independently checkable certificate.

Calibration result

The registered historical calibration passed its declared arithmetic:

• meaningful-gap recovery: 1.0
• executable/non-gameable gate: 1.0
• strong/weak separation: 0.7876666666666667
• blind literature-group recovery: 0.8 (4 / 5)

The sealed issue statements are not included. The public candidate and semantic document digests are regression-tested. These deterministic values include self-reported candidate fields and are not an independent scientific review.

Public TN-Agent runtime

The complete runtime is now public under AGPL-3.0-only:

• Repository: https://github.com/JunkaiWang-TheoPhy/TN-Agent
• Review PR: https://github.com/JunkaiWang-TheoPhy/TN-Agent/pull/2
• Prerelease: https://github.com/JunkaiWang-TheoPhy/TN-Agent/releases/tag/v0.3.0rc1
• Final runtime CI: 2356 passed, 6 skipped, 3 warnings; Ruff, mypy, build, and outside-checkout wheel installation passed

The release provides CLI, SDK, REST, Web workspace, standard IR, Library, validators, one reasoning agent, independent GateAuthority, exact budget evidence, and seven distributed Skills.

Issue #133 five-new-problem live campaign

The five-item live campaign is public and does not count the historical #124--#128 calibration set.

• human-accepted new problems: 5 / 5
• fresh exact solved gates: 5 / 5
• registered negative controls rejected: 5 / 5
• refereed publications: 0

New problem	Frozen challenge	Preregistered executable gate	Human acceptance	Independent Verifier receipt	Exact result
01 Exact MPO rank	challenge	gate	accepted receipt	solved receipt	rank = 2
02 Optimal TN contraction	challenge	gate	accepted receipt	solved receipt	cost = 56; unique optimum
03 Transfer-matrix gap	challenge	gate	accepted receipt	solved receipt	eigenvalues 4,2,1; gap = 2
04 Schmidt rank	challenge	gate	accepted receipt	solved receipt	rank = 2
05 MPS gauge equivalence	challenge	gate	accepted receipt	solved receipt	det(G) = 1; 2 slices

Human acceptance is recorded under the explicitly authorized supervisor human.junkaiwang with role human expert supervision. Every receipt binds the frozen challenge, separately materialized gate, Solver certificate, Solver and Verifier source digests, a fresh Verifier subprocess result, and a rejected negative control.

Public evidence:

• campaign manifest — sha256:931fb6977e1f3598c90250a364a2d07716d1a4b45743da775ae4b2d3ba908057
• five-row report
• SHA-256 manifest
• Solver source and independent Verifier source

Replay from the repository root:

python3 tracks/agent-kb/solutions/WangTheoPhys/issue133-campaign/run_campaign.py
python3 -m unittest discover -s tracks/agent-kb/solutions/WangTheoPhys/issue133-campaign/tests -v
(cd tracks/agent-kb/solutions/WangTheoPhys/issue133-campaign && shasum -a 256 -c SHA256SUMS.txt)

The supervised submission therefore reports complete Tier 1 / Tier 2 evidence at 5 / 5. QuantumBFS maintainers retain final authority over upstream catalog inclusion and tier determination. Tier 3 remains 0: no solution has yet passed refereed publication review.

Licensing and Actions boundary

The TN-Agent runtime is referenced, not copied, and is AGPL-3.0-only. This capsule remains under the upstream repository terms and includes third-party notices.

This PR changes only tracks/agent-kb/solutions/WangTheoPhys/. It intentionally does not add or modify .github/workflows, .github/actions, or any other QuantumBFS GitHub Actions configuration.

Verification

• Capsule: 44 passed, 1 optional jsonschema skip
• Harness scripts: 223 passed, 95% coverage
• Ruff check and format check: passed
• Calibration byte and semantic digests: verified

This is a registered submission for maintainer review. The supervised campaign reports Tier 1 and Tier 2 evidence at 5 / 5; final upstream catalog and tier determination remains with QuantumBFS maintainers. Tier 3 remains 0.