2026-05-17 Research Project Roots Deep Review

Provenance

Origin: local read-only review requested by the user on 2026-05-17.

Roots inspected:

Commands used included:

git status --short --branch
git remote -v
git ls-files
rg --files
Get-ChildItem -Recurse -File

No files in the four project repositories were modified.

EXTRACTED: All git-tracked paths were classified by category: source, config, test, doc, script, prompt, paper, artifact pointer, or other metadata.
EXTRACTED: Large raw data, generated outputs, logs, .tar.gz, .tgz, and model/checkpoint-like artifacts were inventoried by path pattern and size rather than copied or read line by line.
EXTRACTED: Public wiki pages record source paths, summaries, and safety boundaries only. They do not mirror raw data, logs, private .env contents, model weights, or evidence archive contents.

Project	Branch/status	Remote
analog-agent	`main...origin/main`; clean	`https://github.com/appleweiping/analog-agent.git`
uncertainty	`main...origin/main`; untracked `week8_official_external_qwen3base_multik_comparison.md`	`https://github.com/appleweiping/Pony-Rec.git`
truce-rec	`main...origin/main`; untracked `log/`, `log_tensorboard/`	`https://github.com/appleweiping/TRUCE-Rec.git`
tgl-rec	`codex/phase9e-lora-rerank-eval...origin/codex/phase9e-lora-rerank-eval`; clean	`https://github.com/appleweiping/TGL-Rec.git`

Project	Tracked paths	`rg --files`	Text-like tracked	Main category distribution
analog-agent	376	358	359	source 184; test 82; script 58; config 33; doc 7; template 4; infra 4
uncertainty	1,062	1,063	585	artifact-pointer 452; config 237; source 236; doc 64; test 27; paper 24; script 19
truce-rec	617	622	612	source 151; config 143; test 120; doc 100; script 75; artifact-pointer 19; prompt 7
tgl-rec	588	575	586	source 203; test 149; config 106; doc 60; script 58; prompt 6; artifact-pointer 4

Project	Largest artifact patterns	Public wiki treatment
analog-agent	ignored `archive/legacy/...` and `.artifacts/simulation/.../tran.json` files around 0.2 MB each; no giant raw/model artifact visible in the scan	Record route, truth boundary, and artifact policy. Do not copy archive outputs, simulator logs, PDK/model-card contents, or private paper drafts.
uncertainty	Amazon raw JSONL/GZ files up to about 21.6 GB; processed MovieLens JSONL up to about 4.8 GB; official evidence tarballs up to about 397 MB	Record path patterns, sizes, and status only. Do not copy raw data or evidence archive contents.
truce-rec	Amazon Video Games raw JSONL about 2.6 GB; observation inputs and examples up to about 790 MB; Pony evidence tarballs up to about 379 MB; controlled-baseline score plans about 253-255 MB	Record manifest/evidence status only. Do not copy `.env`, logs, raw outputs, or evidence package contents.
tgl-rec	`TGL-Rec-phase9e-lora-results.tgz` about 2.8 GB; raw Video Games JSONL about 2.6 GB; predictions JSONL about 1.4-1.7 GB each	Record as generated/protocol artifacts. Do not copy full predictions, PDFs, archives, model artifacts, or private configs.

EXTRACTED: The safe default claim is a calibrated, surrogate-guided analog sizing loop under explicit SPICE truth boundaries.
EXTRACTED: Canonical first-read docs are 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, and scripts/run_system_closure_report.py.
EXTRACTED: The project must not claim signoff, full-PDK, layout, PEX, yield, or universal analog-design closure unless evidence is present and reproducible.
EXTRACTED: archive/ is the durable ignored sink for paper drafts, benchmark exports, experiment results, reports, and historical outputs; .artifacts/ is transient simulator/cache output.

EXTRACTED: Defended claim is task-grounded calibrated uncertainty improving controlled candidate ranking/reranking reliability under same-schema evaluation.
EXTRACTED: Canonical first-read docs are README.md, docs/milestones/README.md, docs/paper_claims_and_status.md, docs/top_conference_review_gate.md, and docs/server_runbook.md.
EXTRACTED: Official baselines require pinned official or official-code-level implementations, Qwen3-8B declared adaptation, exact same-candidate score export, and importer gates.

EXTRACTED: Strategic spine is generative recommendation observation → CURE/TRUCE framework → Qwen3-8B-LoRA Ours adapter and ablations → official baseline families → shared same-candidate evaluator → four-domain experiments.
EXTRACTED: Current paper-facing external baseline route is Pony official evidence import, not legacy controlled-adapter reruns.
EXTRACTED: Evidence labels distinguish smoke/mock, diagnostic, controlled adapter pilot, official native controlled, and paper result.

EXTRACTED: Active direction is Phase 10: observation to original framework to full recommender system.
EXTRACTED: Active framework is src/llm4rec; src/tglrec is legacy/compatibility tooling unless explicitly targeted.
EXTRACTED: Active baseline plan reuses Pony official baselines, with promax pending and setrec blocked/replaced in current TGL memory.

INFERRED: Vipin wiki and project-local AGENTS.md files are compatible if treated as layers.
INFERRED: Vipin wiki owns durable routing and partner-language memory; target projects own local edit, claim, baseline, server, and evidence rules.
INFERRED: Opus, Sonnet, DeepSeek/Whale, and Codex parallel selves can be used as read-only or bounded collaborators when the environment permits, but Codex remains responsible for integration, verification, commits, and pushes in Vipin wiki.
INFERRED: Analog-agent’s multi-agent requirement is compatible with the workbench: use explorer/reviewer passes for complex analog tasks, but preserve SPICE/configured-truth gates and target-repo ownership.

This source note is a public-safe metadata record, not a complete raw-file mirror.
Some live artifacts, especially untracked summaries or fresh tarballs, may be newer than canonical docs. Future agents should resolve such drift inside the target repo before changing paper-facing status.
The review did not run project tests or validate scientific metrics; it created a wiki workbench and routing memory.