Astro Research Program

Owner: Nicolas (PI). Orchestration: OpenClaw. Coding/analysis: Claude Code. Started: 2026-08-09.

Mission

Make a real, credited discovery as an amateur: transient, exoplanet, variable star, or a data-mined anomaly in a public survey. Build toward publishable work.

Strategy (agreed 2026-08-09, Slack #astronomy)

Primary route: data mine the big public surveys. Most discoveries now come from anomalies in big datasets, not telescopes. Machine learning on survey data is where low hanging fruit remains: weird objects, outliers, misclassifications.

Ripest surveys to learn first: Gaia (2 billion stars, underexploited) and TESS/Kepler light curves (exoplanets still found in archival data). ZTF alerts for transients.

Learning loop: read arXiv astro-ph, replicate a recent paper's analysis on public data, then push somewhere the authors didn't.

Backup/fast routes if we want quick wins: SOHO sungrazer comets, supernova hunting vs reference images, AAVSO variable star measurements, Zooniverse projects.

Method guardrails (from second-round critique, 2026-08-09)

Hard requirements for experiment 001, from Nicolas's parallel Claude review: - The gap is documented, not inferred: the TESS Project's own CTOI evaluation says the promoted community sample is exactly FFI targets fainter than the QLP Tmag 10.5 cutoff plus single transits rejected by periodic searches. Our box is the published gap. - Calibration: ~3,400 CTOIs, ~20% convert to TOIs (168 TOIs first identified as CTOIs). Plan for ~1 TOI per 5 well-vetted submissions, not 10. - Detrend from simple-aperture flux ourselves (wotan biweight, ~0.5 to 1 day window). Never build the search on pre-detrended QLP/SPOC products: their filters absorb lone transits upstream of every metric. - Injection-recovery MUST run through the identical detrending path. Injections that bypass the detrender produce a fictional completeness map. Most common self-deception in this exact search. - Dominant false positive in the Tmag 11.5 to 14 box: blended eclipsing binaries (21 arcsec pixels). Centroid vetting carries the load: difference imaging, per-pixel light curves, Gaia DR3 neighbor sweep to ~2.5 pixels with required-depth-per-neighbor computation, plus archival ZTF/ATLAS/ASAS-SN at epoch. No period-based discriminants exist for single events. - Hunt duotransits: for every monotransit candidate, sweep all other sectors for a matching second event. Two events give an alias comb of allowed periods, which is schedulable. That conversion is what makes a candidate worth a TFOP night. - Prefer TESSCut/astrocut server-side cutouts over bulk FFI pulls. - Gaia correction: the consortium-lock claim overshoots (independent groups publish most Gaia papers; VSX is an individual channel). The claim that does the work: discoveries needing spectroscopic confirmation are out of reach. DR4 gets at most one timeboxed week in December, no parallel program.

Active questions

• Experiment 001 (proposed): monotransit/duotransit search in faint TESS FFI light curves (Tmag 11.5-14), extended-mission sectors, 0.5-3% depth, 4-20h duration. Awaiting Nicolas's GO.
• Side quest (approved shape): SOHO sungrazer comet hunting, ~20 min/week, browser only. Honest odds: P(discovery) 0.05 to 0.2/yr at that effort (global volunteer pool, prolific hunters sweep new frames within hours). Real value: trains artifact-vs-real discrimination on fast feedback. Verified 2026-08-09: SOHO still operating, extended to at least 2029; Sungrazer Project also covers STEREO imagery.

Candidate targets / watch list

(Empty.)

Decisions log

• 2026-08-09: Tooling decided. Claude Code for hands-on analysis, OpenClaw for orchestration, memory, scheduling, comms. Skill astro-research created (pending apply).
• 2026-08-09: Discovery route decided after adversarial debate: TESS monotransit hunting via ExoFOP CTOI channel as primary program. Gaia deprioritized (corrected claim: spectroscopic-confirmation science is out of reach for us; DR4 allowed one timeboxed week in December). SOHO kept as side quest only.
• 2026-08-09: Second-round critique integrated (see Method guardrails): documented-gap evidence, CTOI conversion calibration, custom detrending with matched injection-recovery, EB centroid vetting, duotransit stage-two sweep, TESSCut over bulk FFIs, corrected SOHO odds.
• 2026-08-09: Publishing policy set by Nicolas: FULL OPEN. Everything publishes to the site as it happens, including candidate details. Accepted tradeoff: a published candidate could in principle be submitted as a CTOI by someone else first. Mitigation: fast submission turnaround when a real candidate appears, and public timestamped commits as a priority record. (External submissions themselves still require Nicolas's sign-off.)

Next actions

1. Nicolas: GO on experiment 001 brief.
2. Nicolas: authenticate Claude Code on the server. claude setup-token (confirmed present in v2.1.226) mints a long-lived token against his subscription; alternative is ANTHROPIC_API_KEY in env. Offered: run it in tmux and relay the auth URL via Slack.
3. On GO: write experiments/001-tess-monotransits/brief.md per 12-week spine with method guardrails baked in (reproduce known TOI, custom detrend pipeline, single-transit search, injection-recovery through same path, centroid/EB vetting, duotransit sweep, then scale).
4. Optional: arXiv astro-ph digest cron into #astronomy.