| [ |
| { |
| "id": "autoresearch-cosmic-recreation__note__Users_rs2020_ResearchV3_notes_experiments_autoresearch-cosmic-recreation_EXPERIMENT_README_md", |
| "experiment_id": "autoresearch-cosmic-recreation", |
| "title": "EXPERIMENT_README.md", |
| "filename": "EXPERIMENT_README.md", |
| "relative_path": "/Users/rs2020/ResearchV3/notes/experiments/autoresearch-cosmic-recreation/EXPERIMENT_README.md", |
| "content_md": "# AutoResearch Cosmic Recreation \u2014 Verification Quality + Cost Comparison\n\n## What this actually is (updated after user clarification)\n\nThis is **not** a generic cost-scoping exercise \u2014 it's a controlled comparison experiment.\n\n`wrudman/cosmic-auto-research` is a completed run of **Denario** (a different autonomous-research\nframework) on an ALMA radio-astronomy dataset (project PJ0209, Band 6, lensed target). The user's\nresearch group uses Denario to both **plan** and **execute** real astronomy research on large\ndatasets, then write papers from the results. The user has observed that Denario's output quality\nis poor specifically around **verification between phases** \u2014 models are overconfident, ignore\nfailures, and let bad intermediate results propagate to the next step.\n\nThis is directly visible in the reference run: `results/researcher_pre_results_review.md`\ndocuments that Steps 6, 9, and 11 produced NaN RMS values, a precision-floor table full of NaNs,\nand all-zero moment-0 maps for the only detected transition \u2014 yet the pipeline still advanced\nthrough all 15 steps and several figures were marked \"approved_for_draft\" despite depending on\nbroken intermediate products.\n\n**The hypothesis:** Take the exact same frozen plan (`idea.md` + `methods.md` +\n`data_description.md` + `final_plan.json`) that produced this flawed Denario run, and execute it\nthrough **AutoResearch** instead \u2014 holding the research question, data, and plan constant, and\nchanging only the execution engine. Compare:\n\n1. **Does AutoResearch's independent critic/blind-review architecture catch the same verification\n failures** (NaN propagation, zero moment maps, overconfident approval) that Denario missed?\n2. **What does it cost** to run this comparison?\n3. **What concrete verification-stage design lessons** can we extract for how RACA's own\n experiment pipeline enforces validation between phases \u2014 this directly informs\n `.claude/rules/experiments.md`'s \"No analysis before validation\" gate and the `data-validator`\n agent's criteria.\n\nThis is a legitimate controlled comparison: same idea, same data, same plan \u2014 different execution\nengine \u2014 to isolate whether AutoResearch's architecture (independent multi-model critic + blind\nreview, stateful plan gates) produces better-caught verification failures than Denario's.\n\n## Reference run: what Denario actually cost\n\nThe cloned `cosmic-auto-research` repo includes real `reports/cost/*.json` token logs from the\nPJ0209 run (15 control steps + 1 planning phase, all on a single login-provided dataset \u2014 the 4\nALMA FITS cubes, ~440MB each, already staged on `/scratch` by the original author).\n\n| Metric | Value |\n|---|---|\n| Model used (100% of calls) | `gpt-5-mini-2025-08-07` |\n| Total calls | 95 |\n| Prompt tokens | 6,469,133 |\n| Completion tokens | 492,266 |\n| **Estimated $ cost** (at $0.25/$2.00 per 1M in/out) | **\u2248 $2.60 for the entire project** |\n\nDenario's run was cheap because it used one inexpensive model throughout, with no cross-model\nindependent review.\n\n## AutoResearch: why it won't cost the same for equivalent scope\n\nAutoResearch's default role config assigns **premium models to every key role**, and several roles\nrequire independent (cross-model) review rather than a single model marking its own work:\n\n| Role | Default model | Purpose |\n|---|---|---|\n| `agent` | claude-opus-4.8 | Drives plan \u2192 code \u2192 pilot \u2192 full run \u2192 analysis (this is the bulk of the work, analogous to Denario's 15 control steps) |\n| `code_reviewer` | gemini-3.1-pro | Reviews plan + code |\n| `critic` (+ optional `critic_secondary`) | gpt-5.5 (+ gemini-3.1-pro) | Challenges completion, blind review |\n| `ideator` (\u00d73 distinct models) | claude-opus-4.8, gemini-3.1-pro, gpt-5.5 | **Only used in Idea Generation / Forge \u2014 not used if you bring your own idea, as the Denario reference run did** |\n\nCurrent API pricing (per 1M tokens, verified via web search, Oct 2026 \u2014 **not yet cross-checked\nagainst each provider's own pricing page, flagged in red-team brief**):\n\n| Model | Input | Output |\n|---|---|---|\n| gpt-5-mini | $0.25 | $2.00 |\n| gpt-5.5 | $5.00 | $30.00 |\n| gemini-3.1-pro | $2.00 | $12.00 (\u2264200K ctx) |\n| gemini-3.1-flash-lite | $0.25 | $1.50 |\n| claude-opus (4.5/4.8 proxy \u2014 exact 4.8 pricing unconfirmed) | $5.00 | $25.00 |\n\n### Scenario A \u2014 Execution-only (matches the Denario reference run scope)\n\nThis is the closest apples-to-apples comparison: start from a pre-written idea (like the reference\nrun's `idea.md`), skip Idea Generation/Forge entirely, run only `ar-runtime`'s plan \u2192 code \u2192 pilot\n\u2192 experiment \u2192 critic \u2192 blind review \u2192 close loop.\n\nUsing Denario's observed token volume (6.96M tokens, ~93%/7% prompt/completion split) as a rough\nproxy for the `agent` role's workload, plus smaller estimated volumes for `code_reviewer` and\n`critic`:\n\n| Role | Est. tokens (in/out) | Cost |\n|---|---|---|\n| `agent` (claude-opus-4.8) | 6.47M / 0.49M | \u2248 $44.60 |\n| `code_reviewer` (gemini-3.1-pro) | ~0.9M / ~0.1M | \u2248 $3.00 |\n| `critic` (gpt-5.5) | ~1.3M / ~0.2M | \u2248 $12.50 |\n| **Total (rough order of magnitude)** | | **\u2248 $45 \u2013 $105** |\n\n(Range reflects uncertainty in how much more verbose an agentic Claude Code CLI loop is vs.\nDenario's more direct control steps \u2014 tool calls, file reads, and multi-turn coordination\ntypically add 1.5\u20132x token overhead for equivalent work.)\n\n**That's roughly 17\u201340x more expensive than the Denario reference run for comparable scope**,\ndriven entirely by model choice (premium models + independent review), not by anything vista-side.\n\n### Scenario B \u2014 Full pipeline (adds Idea Generation / Forge)\n\nNot present in the reference run. Cost here scales with seed count \u00d7 knowledge directions \u00d7\nideator seats (3 models, all cross-reviewing every candidate) \u2014 the AutoResearch FAQ explicitly\nwarns this is the most expensive, slowest-scaling stage. **Cannot estimate without knowing how\nmany seeds/directions you intend to run** \u2014 this is an open question below.\n\n## Vista: compute is cheap, storage is the actual constraint\n\nThe LLM orchestration itself (Idea Pipeline + `ar-runtime` coordinator) is lightweight\nPython/Node \u2014 it doesn't need a GPU and barely needs a SLURM job at all. The part that needs\ncluster compute is whatever analysis code the agent generates and runs on the dataset (comparable\nto the reference run's `step2_aperture_extraction.py`, `step9_reconstruction_grid_precision_floor.py`\n\u2014 NumPy/SciPy/Astropy, CPU-only, single-node).\n\n| Resource | Status | Notes |\n|---|---|---|\n| `gg` partition (CPU-only, Grace-Grace) | Matches \"no GPU needed\" | ~0.33 SU/node-hour (Cornell-documented; **not yet cross-checked against TACC's own portal \u2014 flagged in red-team brief**) |\n| `/scratch` | **13PB total, 3.5PB free, no per-user quota** | Right place for \"large amounts of storage\" \u2014 free of charge in SU terms, but has a ~10-day purge policy on untouched files |\n| `/work` | **97.4% of your file quota already used** | Do NOT stage the new dataset here |\n\nAt plausible analysis-job walltimes (single-node CPU jobs, likely 2\u20136 node-hours per project cycle\nbased on the reference run's per-step scripts), vista SU cost per project is **\u22480.7\u20132 SUs** \u2014 against\nyour `AST24021` balance of 80,937 SUs, this is negligible. **The dollar cost is almost entirely\nLLM API spend, not vista compute.**\n\n## Scope: this is Idea Execution only, not Idea Generation\n\nSince we're reusing an already-frozen plan (not discovering a new one), this maps to AutoResearch\nsection 5 (\"Idea Execution\") only. The `ideator`/Forge roles (3-model cross-review for generating\nnew ideas) are **not invoked** \u2014 we write the existing plan directly to `data/ideas/` and run\n`ar-runtime` against it. This keeps Scenario A (~$45\u2013105/project, see above) as the relevant cost\nestimate, not Scenario B.\n\n## Data provisioning\n\nThe original run's data lived at `/scratch/10941/wrudman/alma_pj0209/...` \u2014 **confirmed\ninaccessible to us** (permission denied when checked from vista; likely a different user's\nallocation, possibly on a different TACC system entirely). We will need to re-download the ALMA\narchive data for project **ADS/JAO.ALMA#2024.1.01396.S** ourselves from the ALMA Science Archive.\n\nTwo different storage scales depending on scope (see open question #1 below):\n- **Narrow recreation** (same single target PJ0209, same 4 delivered pipeline-product FITS cubes\n used in the reference run): ~1.8GB. Trivial for `/scratch`.\n- **Full original thesis scope** (`idea.md` calls for **two targets**, raw re-imaging, and\n self-calibration \"if any dataset requires restoration from raw\"): would need the raw ALMA\n delivery package(s), which can run tens to hundreds of GB per target. This is where \"large\n amounts of storage\" actually applies \u2014 but it's a bigger, slower, costlier project than a direct\n Denario-vs-AutoResearch comparison needs.\n\n**Recommendation**: start narrow. Recreate the exact same single-target analysis Denario already\nran (same 4 FITS cubes, same plan) so the comparison is apples-to-apples and cheap, then decide\nwhether to extend to the full two-target thesis scope once we've seen how AutoResearch handles\nverification on the known-problematic case.\n\n## Open questions\n\n1. **Confirm scope**: narrow single-target recreation first (recommended), or jump straight to\n the full two-target thesis scope from `idea.md`?\n2. **Model provider mix \u2014 this directly affects the hypothesis.** You added an OpenAI key, which\n covers `agent`, `code_reviewer`, and `critic` entirely within OpenAI (e.g. gpt-5.5 for\n agent/critic, gpt-5-mini for code_reviewer). But AutoResearch's `critic_secondary` role \u2014 the\n second, independent blind-review pass \u2014 is **skipped automatically when no second provider's\n credential is configured**. Running OpenAI-only means one model family effectively reviews its\n own agent's work for the primary critic role, which risks reproducing the exact \"models are\n overconfident, ignore failures\" problem you're trying to measure improvement on. Do you want to\n add an Anthropic and/or Gemini key too, so the critic/review roles are genuinely independent of\n the executing agent? This seems important for the hypothesis to be a fair test.\n3. **API budget**: any cap you want enforced before the comparison run, given Scenario A's\n $45\u2013105/project estimate (and this adds a second, independent-review-capable provider if you\n go with #2)?\n\n## Status\n\nDesign phase. No red-team review or compute has happened yet. See `red_team_brief.md` for flagged\nrisks (storage location, unverified SU rate, unconfirmed Opus-4.8 pricing, data re-download\nneeded, single-provider verification risk).\n", |
| "created": "", |
| "updated": "" |
| }, |
| { |
| "id": "autoresearch-cosmic-recreation__note__Users_rs2020_ResearchV3_notes_experiments_autoresearch-cosmic-recreation_HUGGINGFACE_REPOS_md", |
| "experiment_id": "autoresearch-cosmic-recreation", |
| "title": "HUGGINGFACE_REPOS.md", |
| "filename": "HUGGINGFACE_REPOS.md", |
| "relative_path": "/Users/rs2020/ResearchV3/notes/experiments/autoresearch-cosmic-recreation/HUGGINGFACE_REPOS.md", |
| "content_md": "# HuggingFace Repos\n\nNo artifacts yet \u2014 this experiment is in the design/cost-scoping phase. No compute has run.\n", |
| "created": "", |
| "updated": "" |
| }, |
| { |
| "id": "autoresearch-cosmic-recreation__note__Users_rs2020_ResearchV3_notes_experiments_autoresearch-cosmic-recreation_red_team_brief_md", |
| "experiment_id": "autoresearch-cosmic-recreation", |
| "title": "red_team_brief.md", |
| "filename": "red_team_brief.md", |
| "relative_path": "/Users/rs2020/ResearchV3/notes/experiments/autoresearch-cosmic-recreation/red_team_brief.md", |
| "content_md": "# Red Team Brief\n\nStatus: **ready for adversarial review**. Scope is settled (narrow single-target recreation,\nexecution-only, local Docker), environment is technically validated end-to-end (live preflight\npasses for all three roles actually used), and data is staged. Items below are marked RESOLVED\nwhere applicable; remaining open items are real residual risk, not blockers to a canary.\n\n## Flagged risks\n\n1. **[RESOLVED] Pricing was third-party-sourced at initial scoping time.** Superseded \u2014 we now\n have live, measured preflight calls against real endpoints. Actual $ cost will come from\n provider usage dashboards after the run, not estimates.\n2. **[RESOLVED] claude-opus-4.8 pricing proxy concern.** Moot \u2014 we switched the `agent` role to\n `claude-sonnet-5` (verified live, cheaper than Opus) rather than guessing at an unverified model.\n3. **[SUPERSEDED] Vista `gg` SU rate.** Moot for this run \u2014 compute moved off vista entirely (see\n below). Still relevant if a future full-thesis-scope run goes back to vista.\n4. **[SUPERSEDED] `/work` quota.** Moot for this run \u2014 no vista storage used. Data lives in\n `private_projects/AutoResearch/local_data/alma_pj0209/` locally and is mounted read-only into\n the container.\n5. **[SUPERSEDED] `/scratch` purge policy.** Moot for this run (see above). The 9.24GB already\n staged on vista's `$SCRATCH/alma_pj0209` remains available for a future full-scope run but isn't\n used by this one \u2014 note it could still be purged after ~10 days of inactivity if untouched.\n6. **[OPEN] Idea Generation cost remains unscoped.** Not relevant to this run (execution-only, no\n Idea Generation/Forge invoked) \u2014 carry forward only if a future run adds that phase.\n7. **[OPEN] Token-volume proxy from Denario may not transfer.** Still true \u2014 the $45\u2013105/project\n estimate is unmeasured. This is precisely what the canary (below) should produce a real number\n for.\n8. **[ADDRESSED, not eliminated] `ar-runtime` requires broad tool permissions**\n (`claude --dangerously-skip-permissions`). Containment plan: runs inside a disposable Docker\n container with only the AutoResearch repo and the read-only data mount \u2014 no host home directory,\n no SSH agent, no host credentials beyond the two API keys explicitly injected as `-e` flags at\n launch (never written to disk). This is a real isolation boundary, not just a config flag, but\n the container still has full network egress and can run arbitrary shell inside itself \u2014 it is\n not a hermetic sandbox.\n9. **[USER DECISION, carried forward]** `agent` is Anthropic (claude-sonnet-5, required \u2014\n ar-runtime's agent role IS the Claude Code CLI itself, not a configurable choice), while\n `code_reviewer` (gpt-5-mini) and `critic` (gpt-5.5) are OpenAI, per user's explicit \"OpenAI-only\n for now\" call. `critic_secondary` is not enabled (would need a third distinct provider). This\n means the run has two genuinely independent model families (Anthropic executing, OpenAI\n reviewing/critiquing), which is better independence than a pure single-provider setup, but not\n the full 3-model independence AutoResearch's defaults recommend. Worth noting explicitly in the\n eventual write-up.\n10. **[RESOLVED] Original data was inaccessible at `/scratch/10941/wrudman/...`.** Re-downloaded\n directly from the ALMA Science Archive (project 2024.1.01396.S, Member OUS\n uid://A001/X3788/X5b8c), byte-verified against `data_description.md`. First download attempt\n silently truncated (1.43GB of 2.28GB) \u2014 caught via exact byte-count check and re-downloaded\n clean. Lesson: always verify downloaded archive sizes against the archive's own reported\n `content_length`, don't trust curl's exit code alone.\n\n11. **[NEW] AutoResearch's template config needed several fixes for current-generation models,\n none of which are specific to our setup.** Documented here since they'd hit any user running\n AutoResearch today: (a) `OPENAI_BASE_URL`/`ANTHROPIC_BASE_URL` must be explicitly set \u2014 the\n endpoint config builds request URLs directly from these env vars with no default fallback to\n the real provider APIs; (b) current OpenAI models reject the legacy `max_tokens` param, need\n `max_completion_tokens`; (c) current OpenAI and Anthropic models reject a non-default\n `temperature` value entirely (`supports_temperature` must be `false`); (d) reasoning models\n (gpt-5-mini) can consume an entire small token budget on hidden reasoning before any visible\n output \u2014 `request_defaults.max_tokens` raised from the template's 8192 to 32768.\n12. **[NEW] Locally patched one line in AutoResearch's own `scripts/preflight.py`.**\n `PROBE_MAX_TOKENS` was hardcoded to 16, which is too small for a reasoning model to reliably\n produce any visible output (it can spend the entire budget on hidden reasoning), causing\n false-negative preflight failures (HTTP 200, but no text) unrelated to actual role health.\n Raised to 512, documented inline in the patch. This only affects the connectivity probe, not\n actual experiment execution request budgets (those come from `request_defaults.max_tokens`).\n13. **[NEW] Pre-existing AutoResearch bug found during environment validation (not Docker-related).**\n `ar-runtime/scripts/tests/test_engine_terminal_receipts.py::\n test_execute_run_prevents_pilot_from_creating_a_main_artifact_directory` fails against\n AutoResearch's own unmodified code. It shows that when a pilot-stage run script writes into the\n sibling `main`-stage artifact directory, `execute_run` does not prevent or clean this up \u2014 a\n containment gap in the \"pilot before scaling\" mechanism the project advertises as a safety\n feature. Relevant to this experiment's own hypothesis about verification robustness: if our\n actual run's pilot stage ever fails or behaves unexpectedly, check whether a stray\n `results/run_artifacts/main/` directory exists afterward before trusting that the main-stage\n run started clean. 954/955 of AutoResearch's own tests pass; this is the one confirmed\n pre-existing defect, independent of our Docker environment (the other 17 initial failures were\n all traced to two fixable Docker gaps: missing `pgrep`/`procps`, and `npm install -g` landing\n in a PATH location one test assumed would be empty \u2014 both fixed in `Dockerfile.pj0209`).\n\n14. **[FOUND DURING CANARY, FIXED] MCP-server-routed roles (`code_reviewer`, `critic`) read `.env`\n fresh from disk via `call_role.py`, completely independent of the container's inherited\n process environment.** Our `-e ANTHROPIC_API_KEY=...`/`-e OPENAI_API_KEY=...` injection worked\n fine for directly-invoked scripts (preflight.py, render_env.py \u2014 they inherit shell env), but\n the long-lived MCP servers (`ar-gemini-review-mcp.ts`, `ar-external-critic-mcp.ts`) are spawned\n by Claude Code itself and only ever see what `call_role.py` parses from the literal `.env` file\n on disk \u2014 which deliberately had no real keys, by design, to avoid persisting secrets to the\n tracked repo. Result: the canary's `plan_gate` review failed 4 times with \"all configured\n models failed for role code_reviewer\" before the coordinator (correctly) diagnosed the actual\n root cause itself and stopped rather than bypassing the gate or fabricating a decision \u2014 good\n behavior, directly relevant to this experiment's verification-robustness hypothesis. **Fix**:\n `run_in_container.py` now writes the merged (template + real keys) `.env` content to a 0600\n temp file outside the repo, bind-mounts just that single file read-only over `/workspace/.env`\n for the container's lifetime, and deletes the temp file immediately after exit (try/finally).\n The tracked `.env` on host disk never contains real keys at any point. (A single-file tmpfs\n mount was tried first and doesn't work in Docker \u2014 tmpfs targets must be directories.)\n\n15. **[OPERATIONAL INCIDENT, SELF-INFLICTED] Ran two `ar-supervisor.sh` instances concurrently\n against the same project directory for ~an hour.** After the first launch attempt appeared to\n fail (root-permission error), a background-task notification reported it as \"failed,\" but the\n container itself never actually exited \u2014 `ar-supervisor.sh`'s own internal attempt-retry loop\n kept it alive and progressing (it had, in fact, gotten past the root issue on a later internal\n retry and was actively diagnosing the `.env`/MCP credential bug). A second, \"resumed\" launch was\n started on top of it without checking `docker ps` first, so both ran against\n `data/projects/canary_smoke/` simultaneously \u2014 duplicating agent spawns and reviewer calls\n (real cost, no corresponding progress) until caught and both containers were stopped via\n `docker stop`. **Impact assessed as recoverable**: `workflow_queue.json`/`workflow_queue.engine.json`\n are lock-protected and verified valid JSON after the kill; `state.md`/`decisions.log` are not\n lock-protected and were clobbered between the two runs' writes, but the final state (plan_gate\n APPROVED on rev=1, next_unit=code_gate) is coherent, not garbled. Stale lock files referencing\n PIDs from the killed container (`.ar-monitor.lock`, `workflow_queue.lock`) were removed before\n resuming, since container PID namespaces reset per-run and a stale PID reference could\n false-positive as \"still alive\" in a fresh container. **Fix going forward**: check `docker ps`\n for any container already running against the same project before launching or resuming \u2014\n never assume a background-task \"failed\" notification means the underlying process actually\n exited, especially for a supervisor script with its own internal retry loop.\n\n16. **[REAL BUG FOUND AND FIXED] `ar-gemini-review-mcp.ts` hardcoded a 4096-token cap that\n silently overrode the configured 32768-token budget on every real `code_reviewer` call.**\n This is what actually blocked the canary for 6 consecutive attempts with \"all configured\n models failed for role code_reviewer\" \u2014 not an intermittent empty-completion issue (an earlier\n hypothesis, addressed with a retry-on-empty patch in `src/llm_client.py:call_role_result`\n which is harmless but was not the real fix). The MCP tool's `max_output_tokens` parameter is\n optional and the real reviewer-agent flow never passed it, so every call fell back to\n `maxTokens: max_output_tokens ?? 4096` \u2014 far too small for `oai-mcp-reviewer` (gpt-5-mini, a\n reasoning model whose hidden reasoning tokens can exhaust a 4096 cap before any visible\n output). This explains why my own direct `call_role.py` test calls kept appearing to work\n (that CLI path uses `request_defaults.max_tokens` directly, bypassing the MCP wrapper's\n hardcoded override entirely) while the real agent-driven reviewer calls failed consistently \u2014\n two different code paths with two different effective budgets, which cost real time and\n confusion before being disambiguated. **Fix**: removed the `?? 4096` fallback so the call\n falls through to `request_defaults.max_tokens` like every other role when the caller doesn't\n explicitly request a smaller budget. Verified the patched TypeScript builds cleanly\n (`bun build --target=bun`). No equivalent hardcoded cap exists in `ar-external-critic-mcp.ts`\n (grepped, confirmed) \u2014 likely why `critic` worked fine throughout while `code_reviewer` did not.\n\n## Canary recommendation\n\nRecommend: run `ar-supervisor.sh` against a trivial synthetic idea (NOT the real pj0209_recreation\nidea) inside the same Docker container, end-to-end, before spending any cost/time on the real\nscientific content. This should:\n- Exercise the full plan -> code -> pilot -> critic -> blind-review -> close loop at least once\n- Produce real artifacts (`workflow_queue.json`, `decisions.log`, `results/run.log`,\n `results/summary.md`, critic/blind-review files) to confirm the trajectory-capture plan actually\n works before relying on it for the real run\n- Produce one real, measured token/cost number from provider usage, replacing every estimate above\n with ground truth\n- Specifically watch for the known pilot/main artifact-directory leak (risk #14) in case it affects\n interpretation of the canary's own pilot stage\n\n## Redteam status: `pending`\n\nReady for adversarial review via a fresh `red-team-reviewer` subagent (files only, no conversation\nhistory).\n", |
| "created": "", |
| "updated": "" |
| }, |
| { |
| "id": "onboarding__note__Users_rs2020_ResearchV3_notes_experiments_onboarding_EXPERIMENT_README_md", |
| "experiment_id": "onboarding", |
| "title": "EXPERIMENT_README.md", |
| "filename": "EXPERIMENT_README.md", |
| "relative_path": "/Users/rs2020/ResearchV3/notes/experiments/onboarding/EXPERIMENT_README.md", |
| "content_md": "# Welcome to RACA\n\nThis is a sample experiment to show you how the dashboard works. You're looking at the **Overview** tab right now \u2014 it displays the experiment's README (this file).\n\nEverything you see here is generated from plain files in `notes/experiments/onboarding/`. You can browse them in your editor anytime.\n\n## How This Dashboard Works\n\nEach experiment has several tabs at the top. Here's what they do:\n\n### Overview (you are here)\n\nDisplays the experiment's README and any notes you've written in the `user/` folder. This is the main landing page for each experiment \u2014 a summary of what the experiment is, what you're investigating, and what you found.\n\n### Red Team Brief\n\nBefore any experiment runs, RACA reviews the design for potential problems \u2014 wrong evaluation metrics, truncated outputs, missing baselines, wasted compute. The brief lives at `red_team_brief.md`. This tab will be empty until you run your first real experiment.\n\n### Timeline\n\nA chronological log of everything that happened: when jobs were submitted, when artifacts were uploaded, when bugs were found and fixed. This is auto-generated from `activity_log.jsonl` \u2014 RACA writes to it as events happen.\n\n### Runs\n\nTracks each job submission \u2014 which model, which cluster, what status (pending, running, completed, failed), and links to the HuggingFace dataset with the results. Empty until you run something.\n\n### Artifacts\n\nLinks to all HuggingFace datasets produced by this experiment \u2014 canary runs, partial results, final data. Each artifact has metadata about what generated it. Empty until artifacts are uploaded.\n\n### Files\n\nAll the markdown and YAML files in the experiment folder. Click any file to read it. This is a quick way to browse the experiment's configuration and notes without leaving the dashboard.\n\n## Folder Structure\n\n```\nnotes/experiments/onboarding/\n EXPERIMENT_README.md \u2190 this file (shows in Overview tab)\n experiment.yaml \u2190 config: hypothesis, models, tasks\n flow_state.json \u2190 current phase (design/running/complete)\n HUGGINGFACE_REPOS.md \u2190 links to all uploaded datasets\n questions.md \u2190 research questions (read-only)\n red_team_brief.md \u2190 created during preflight review\n activity_log.jsonl \u2190 timeline entries (auto-generated)\n user/ \u2190 YOUR notes \u2014 RACA doesn't touch these\n README.md \u2190 your interpretation and observations\n FINDINGS.md \u2190 key results and surprises\n DECISIONS.md \u2190 design decisions and rationale\n summary.md \u2190 one-paragraph summary when done\n```\n\n**Most of this is automated.** RACA creates and updates the experiment files, uploads artifacts, and keeps the timeline current. The only files you write are in `user/` \u2014 that's your space for notes, findings, and decisions.\n\n## What's Next\n\nThis sample experiment hasn't been run yet \u2014 it's just here to show you the structure. When you're ready to run a real experiment, just tell RACA:\n\n> *I want to test whether Qwen3-8B follows complex instructions better than Llama-3.1-8B*\n\nOr try the full guided tutorial:\n\n> */raca:experiment-tutorial*\n", |
| "created": "", |
| "updated": "" |
| }, |
| { |
| "id": "onboarding__note__Users_rs2020_ResearchV3_notes_experiments_onboarding_HUGGINGFACE_REPOS_md", |
| "experiment_id": "onboarding", |
| "title": "HUGGINGFACE_REPOS.md", |
| "filename": "HUGGINGFACE_REPOS.md", |
| "relative_path": "/Users/rs2020/ResearchV3/notes/experiments/onboarding/HUGGINGFACE_REPOS.md", |
| "content_md": "# HuggingFace Repositories\n\n| Dataset | Date | Rows | Purpose |\n|---------|------|------|---------|\n", |
| "created": "", |
| "updated": "" |
| }, |
| { |
| "id": "onboarding__note__Users_rs2020_ResearchV3_notes_experiments_onboarding_questions_md", |
| "experiment_id": "onboarding", |
| "title": "questions.md", |
| "filename": "questions.md", |
| "relative_path": "/Users/rs2020/ResearchV3/notes/experiments/onboarding/questions.md", |
| "content_md": "# Research Questions\n\n1. Can Qwen3-1.7B solve basic Countdown problems (4 numbers, targets < 100)?\n2. What reasoning strategies does the model use (trial-and-error, systematic search, pattern matching)?\n3. Where does the model fail \u2014 wrong arithmetic, giving up, or invalid expressions?\n", |
| "created": "", |
| "updated": "" |
| } |
| ] |
