vipin wiki

Research Project Workbench

6 min read

Research Project Workbench

Purpose

This page is the routing layer for Weiping’s main local research repositories that need repeated agent work:

ProjectLocal rootRoleStart page
analog-agentD:/Research/Agent-AI4EDA/analog-agentAI4EDA layered analog circuit design harness with SPICE/configured-truth boundariesAGENTS.md, README.md, docs/configured_truth_user_action_boundary.md
uncertaintyD:/Research/UncertaintyPony-Rec / task-grounded calibrated uncertainty and official same-candidate baseline evidenceAGENTS.md, README.md, docs/milestones/README.md
truce-recD:/Research/TRUCE-RecCURE/TRUCE uncertainty-aware generative recommendation and Pony official baseline reuseAGENTS.md, docs/PROJECT_MEMORY.md, docs/RESEARCH_IDEA.md
tgl-recD:/Research/TGL-RecPhase 10 temporal graph-to-language evidence and Pony official baseline migrationdocs/codex_project_memory.md, docs/phase10_master_plan.md

WEIPING_WIKI is the upper-level workbench: it remembers routing, claim boundaries, collaboration norms, and cross-project context. Once work enters a target project, that project’s own AGENTS.md, .codex/skills, canonical docs, and live git state take priority.

Collaboration Priority

  • EXTRACTED: Codex is the main coordinator and the only writer for wiki changes in this repository.
  • EXTRACTED: User-facing collaboration should describe Opus, Sonnet, DeepSeek/Whale, and Codex parallel selves as partners rather than impersonal tools.
  • INFERRED: Project-local multi-agent rules do not conflict with WEIPING_WIKI. Interpret them as: Codex coordinates; Opus/Sonnet/Whale or Codex parallel selves provide bounded review, exploration, or implementation perspectives; the target project’s evidence gates decide claim status.
  • EXTRACTED: For routed research repos, do not modify project files from the wiki workbench unless the user explicitly asks for project edits. Always inspect live git status first.

Shared Evidence Rule

All routed projects are governed by the same paper-safety principle:

No smoke run, diagnostic, scaffold, pilot, wrapper name, or raw output becomes paper evidence until the canonical project gates pass.

Common recommendation gates:

  • same candidate rows, split discipline, and event IDs;
  • exact score schema such as source_event_id,user_id,item_id,score when using the shared external baseline lane;
  • finite scores with no missing, extra, or duplicate candidate keys;
  • provenance for official code, pinned commit, backbone/adaptation policy, score path, and artifact status;
  • validation-only tuning for project methods;
  • paired/statistical checks where the project declares them necessary;
  • clear labels for diagnostic, controlled_adapter_pilot, official_completed, completed_result, and paper_result.

Common analog/AI4EDA gates:

  • explicit physical-validity boundary: demonstrator truth, configured truth, native artifact, external PDK, Spectre-compatible, signoff, layout, PEX, and yield must not be blurred;
  • paper-facing physical evidence should be SPICE/configured-truth backed and reproducible from committed scripts, configs, seeds, archived outputs, and replay manifests;
  • lightweight internal baselines stay labeled lightweight until external/tuned baselines are wired into the same simulator budget and truth loop;
  • archive/ and .artifacts/ are inventory-only from the wiki unless the user asks for local artifact organization.

Project Routes

Analog Agent

Use analog-agent when the task is about AI4EDA, analog circuit design agents, SPICE/configured-truth evidence, world-model-guided planning, simulator-backed verification, memory/reflection, or submission/benchmark packaging.

Startup packet:

  • AGENTS.md
  • README.md
  • docs/configured_truth_user_action_boundary.md
  • docs/repo-map.md
  • docs/related_work_map.md
  • docs/stop_conditions.md
  • configs/default.yaml
  • configs/benchmarks/multi_task_suite_v1.yaml
  • configs/simulator/ngspice.yaml
  • scripts/run_system_closure_report.py

Claim boundary:

  • Safe: calibrated, surrogate-guided analog sizing loop under explicit SPICE truth boundaries; layered analog-agent harness with reproducible vertical slices and benchmark scaffolding.
  • Unsafe until evidence passes: full signoff flow, layout, PEX, yield, universal analog-design closure, production-strength external baselines, or configured-truth claims without user-managed PDK/model assets validated.

Uncertainty / Pony-Rec

Use uncertainty when the task is about C-CRP, SRPD, Shadow, official external baselines, same-candidate score export/import, or the task-grounded uncertainty paper claim.

Startup packet:

  • README.md
  • docs/milestones/README.md
  • docs/paper_claims_and_status.md
  • docs/top_conference_review_gate.md
  • docs/server_runbook.md
  • for baseline work: docs/baseline_protocol.md, OFFICIAL_EXTERNAL_BASELINE_UPGRADE_PLAN_2026-05-07.md, configs/official_external_baselines.yaml, PROJECT_LINEAGE_AND_FILE_INDEX_2026-05-06.md
  • project skill: .codex/skills/uncertainty-rec-research/SKILL.md

Claim boundary:

  • Safe: task-grounded calibrated uncertainty improves controlled candidate ranking/reranking reliability under same-schema evaluation.
  • Unsafe until separately completed: full-catalog recommender SOTA, generative-title recommender, universal LoRA distillation claim, or treating partial official adapters as official completed rows.

TRUCE-Rec

Use truce-rec when the task is about CURE/TRUCE, uncertainty-aware generative recommendation, Ours adapter/ablations, Pony official baseline import, or TRUCE-side evaluation/export.

Startup packet:

  • AGENTS.md
  • docs/PROJECT_MEMORY.md
  • docs/RESEARCH_IDEA.md
  • docs/submission_roadmap.md
  • docs/top_conference_review_plan.md
  • for baselines/protocol: docs/qwen3_lora_controlled_baselines.md
  • for server work: docs/server_execution_matrix.md, docs/server_next_commands.md
  • project skill: .codex/skills/truce-rec/SKILL.md

Claim boundary:

  • Safe: TRUCE is a research-grade route from generative recommendation observation to CURE/TRUCE framework, shared same-candidate evaluator, and four-domain experiments.
  • Unsafe until gates pass: calling legacy controlled adapters official-native baselines, promoting Ours v2 scaffolds to final paper results, or replacing Pony official evidence with TRUCE pilots.

TGL-Rec

Use tgl-rec when the task is about temporal graph-to-language evidence, Phase 10, need-gated TDIG, local Qwen3 reranking, Pony baseline migration, or four-domain temporal evidence evaluation.

Startup packet:

  • docs/codex_project_memory.md
  • docs/phase10_master_plan.md
  • docs/server_runbook.md
  • docs/week8_large_same_candidate_protocol.md
  • docs/reference_baseline_fidelity.md
  • docs/reference_method_adaptation_map.md
  • docs/method_card_time_graph_evidence.md
  • project skill: .codex/skills/tgl-rec-research/SKILL.md

Claim boundary:

  • Safe: Phase 10 aims to test whether LLM recommenders rely on semantic similarity/popularity more than temporal need transitions, then build a temporal graph-to-language framework.
  • Unsafe until gates pass: treating protocol_v1 diagnostics as final evidence, treating reference_*_sft scaffolds as main baselines, or saying the experiment phase is complete before four-domain observation, ablations, official baselines, paired statistics, and reviewer gates pass.

File Area Rules

AreaWorkbench treatment
AGENTS.md, README.md, docs/, .codex/skills/Read first; safe to summarize in public wiki.
src/, scripts/, configs/, tests/Safe to inventory and route; edit only when the user asks to work inside that project.
archive/, .artifacts/, data/raw/, large data/processed/, outputs/, runs/, log/, log_tensorboard/, *.tar.gz, *.tgz, model/checkpoint arraysInventory only: path pattern, size, purpose, and risk. Do not copy raw content into public wiki.
.env, private server configs, credentials, raw logs with possible tokens, PDK roots/model cards, model weightsDo not read or copy into public wiki. Mention only the generic boundary rule.

Server Collaboration

  • Codex usually cannot see the shared server state directly.
  • Future agents should provide copy-paste commands, ask for pasted logs or artifacts, and only claim completion from evidence.
  • Long runs should use runbook patterns with output preservation, log paths, PID or process checks, and artifact gates.
  • If a target project and Vipin wiki disagree, use the target project’s current runbook for server commands, then update the wiki only after grounded changes.

Cross-Project Boundaries

BoundaryKeep separate
C-CRP vs CURE/TRUCEC-CRP is Pony/Uncertainty’s task-grounded calibrated uncertainty method; CURE/TRUCE is TRUCE-Rec’s generative recommendation uncertainty framework.
TRUCE vs TGLTRUCE centers catalog grounding, uncertainty, popularity, long-tail, echo, and conservative policy. TGL centers temporal directed item graph evidence and need transitions.
Pony official baselines vs project methodsPony official baseline evidence can be reused by TRUCE/TGL only with manifest, provenance, exact score gates, and no copying of large archives into git.
AI4EDA vs LLM4Recanalog-agent is an analog circuit/SPICE harness, not a recommender project. It shares evidence discipline and closed-loop uncertainty taste, but its claim gate is physical simulation truth rather than same-candidate ranking.

Counterpoints and Gaps

  • This workbench is a router, not a substitute for live repo inspection. Always rescan current files before editing or making time-sensitive claims.
  • The 2026-05-17 pass classified tracked paths by category, but did not read raw/output/artifact files line by line.
  • Some live artifacts may be newer than canonical docs. When manifests, tarballs, and docs disagree, use the most conservative status until the target project updates its own canonical memory.

Graph View

Backlinks