vipin wiki

2026-05-17 Research Project Workbench Audit

5 min read

2026-05-17 Research Project Workbench Audit

Question

How should Vipin wiki remember D:/Research/Agent-AI4EDA/analog-agent, D:/Research/Uncertainty, D:/Research/TRUCE-Rec, and D:/Research/TGL-Rec so future agents can route work smoothly without weakening project-local evidence rules?

Short Answer

Use Vipin wiki as an upper-level workbench and keep each project repo authoritative for its own claim, server/simulator, baseline, and edit rules. The routed set now has one AI4EDA/SPICE project plus three same-candidate LLM4Rec projects:

  • analog-agent: layered analog circuit design harness with SPICE/configured-truth evidence boundaries.
  • uncertainty: C-CRP/task-grounded calibrated uncertainty and official baseline evidence.
  • truce-rec: CURE/TRUCE uncertainty-aware generative recommendation and catalog-grounding/fallback policy.
  • tgl-rec: temporal graph-to-language evidence and need-transition modeling.

Coverage

This audit classified every git-tracked path in the four repos by file role, then inventoried large raw/generated artifacts by metadata only.

ProjectTracked pathsText-like trackedPrimary readable surfaceInventory-only surface
analog-agent376359AGENTS.md, README.md, docs/, configs/, libs/, apps/, scripts/, tests/, templates/, infra/archive/, .artifacts/, local PDK/model-card paths, simulator logs, generated paper/benchmark outputs
uncertainty1,062585AGENTS.md, README.md, docs/, configs/, src/, scripts/, tests/, prompts/, Paper/data/raw/, large data/processed/, outputs/, evidence tarballs
truce-rec617612AGENTS.md, README.md, docs/, configs/, src/, scripts/, tests/, .codex/skills/data/raw/, large outputs/, log/, log_tensorboard/, evidence packages
tgl-rec588586AGENTS.md, README.md, docs/, configs/, src/llm4rec, scripts/, tests/, .codex/skills/outputs/, artifacts/, runs/, root *.tgz, raw data, PDFs/zips

Project Scale And Shape

Analog Agent

  • Tracked shape: source 184; tests 82; scripts 58; configs 33; docs 7; templates 4; infra 4.
  • Core modules: apps/api_server, apps/orchestrator, worker adapters, libs/schema, libs/interaction, libs/tasking, libs/world_model, libs/planner, libs/simulation, libs/memory, libs/eval, and libs/vertical_slices.
  • Current claim gate: do not claim configured truth, external PDK, Spectre, signoff, layout, PEX, yield, or universal analog-design closure without reproducible evidence and the closeout/configured-truth checks.
  • Main risk: confusing demonstrator-truth submission readiness with configured-truth or signoff-grade readiness.

Uncertainty / Pony-Rec

  • Tracked shape: artifact-pointer 452; config 237; source 236; docs 64; tests 27; paper 24; scripts 19.
  • Core modules: src/uncertainty, src/shadow, src/methods, src/training, src/baselines/official_runner, src/eval.
  • Current claim gate: do not expand beyond controlled same-schema candidate ranking/reranking reliability unless separate protocols complete.
  • Main risk: live evidence tarballs and untracked summaries may tempt overclaiming before canonical manifests/docs agree.

TRUCE-Rec

  • Tracked shape: source 151; config 143; tests 120; docs 100; scripts 75; artifact-pointer 19.
  • Core modules: active src/llm4rec; older src/storyflow still exists and should be inspected before changes.
  • Current claim gate: Pony official baseline reuse is paper-facing only through manifest, evidence package, and exact gate status; TRUCE controlled adapters remain diagnostic unless reopened.
  • Main risk: treating Ours v2 or legacy controlled-adapter pilots as finished paper results.

TGL-Rec

  • Tracked shape: source 203; tests 149; config 106; docs 60; scripts 58.
  • Core modules: active src/llm4rec; legacy src/tglrec is not the default edit target.
  • Current claim gate: Phase 10 is not complete until large observation, Pony baseline migration, TGL ablations, paired statistics, leakage/reproducibility checks, and reviewer gate pass.
  • Main risk: branch and historical Phase 9E files can mislead future agents unless Phase 10 docs remain first.

Artifact Safety

Do not copy these into public wiki:

  • raw Amazon/MovieLens datasets;
  • generated predictions, raw LLM outputs, and logs;
  • .env or private server configs;
  • model weights, checkpoints, embeddings, arrays, and large evidence archives;
  • full PDF or zip contents from reference folders.
  • local PDK roots, external model cards, raw simulator logs, transient .artifacts/, or private analog paper drafts.

Public-safe summaries may include:

  • path pattern;
  • approximate size;
  • purpose;
  • tracked/ignored/untracked status when known;
  • reportability status from canonical manifests;
  • risks and next command source.

Collaboration Conflict Judgment

No hard conflict was found between Vipin wiki and the four project-local agent rules.

Interpretation:

  • Vipin wiki controls durable routing, cross-project memory, public/private safety, and partner-language collaboration norms.
  • The target project controls what to read, what to edit, which tests to run, which server commands to issue, and what evidence qualifies as paper-facing.
  • Opus/Sonnet/Whale or Codex parallel selves can provide review or exploration when appropriate, but their output remains advisory until Codex verifies it against live files.
  • For analog-agent, target-project rules specifically require explorer/reviewer passes for complex tasks and keep claims inside SPICE/configured-truth boundaries.

Maintenance Recommendations

  • Keep research-project-workbench as the first route page for these four projects.
  • Update the entity pages whenever a canonical project doc changes baseline status, current phase, truth boundary, or paper claim.
  • When downstream projects reuse Pony official evidence, record only manifest state, score-gate status, paths, sizes, and hashes; do not copy evidence archives.
  • When analog-agent configured-truth or external-PDK status changes, update analog-agent and research-project-workbench together so future agents do not confuse demonstrator-truth and configured-truth claims.
  • If ProMax status changes in Pony/Uncertainty, update uncertainty, truce-rec, tgl-rec, and the workbench together to avoid stale cross-project memory.
  • Future deep code reviews should happen inside the target project repo after reading its local AGENTS.md and skill.

Counterpoints and Gaps

  • This audit did not run scientific tests, reproduce metrics, or validate output files.
  • It did not read raw datasets, logs, PDFs, or generated prediction JSONL files fully.
  • Artifact counts can drift quickly because experiments generate large ignored/untracked files. Future agents should rerun live inventory before acting on size or status details.

Graph View

Backlinks