# Paper 1 finish plan

Updated October 2, 2026. The immediate objective is to close a bounded set of claims and produce the manuscript, not extend every development investigation. This plan supplements the detailed task board and preserves existing owners and accepted decisions. No messages or new assignments were dispatched to other tasks by creating this record.

## The two active research workstreams

| Workstream | Current task | Required handoff |
|---|---|---|
| Account for removed neighbours | Trace PS1 star removal records | Versioned Gaia/PS1/image-source associations; exact removal masks and finite-pixel changes; full/partial survival; seam ownership; explicit unmatched/unknown cases; product hashes and a statement of whether template pixels changed |
| Test neighbour-aware ePSF fitting | Summarize ePSF atlas findings | Matched baseline and augmented scenes; fixed evaluation stars/pixels; free-flux and calibrated soft-prior comparisons; role flags; held-out profile/astrometry/flux results; convergence and residual images; adopt/reject recommendation |

The ledger implementation record is `/astro/armin/koji/syndiff/dev_runs/ps1_ledger_fix_20261002/README.md`. Its approved scope preserves production pixels while instrumenting and validating bookkeeping. Do not assume the ledger repair changes the removal algorithm.

The active fitting task is checking the existing `/astro/armin/koji/syndiff/dev_runs/gaia_neighbour_joint_fit_20261002/` harness before launching duplicate work. That harness came from **Summarize ePSF atlas findings (2)**, whose recent turns were interrupted. Reuse its verified pieces and the completed `/astro/armin/koji/syndiff/dev_runs/neighbour_priors_20261002/` stress tests; do not count them as independent real-data validation.

The current preferred experiment keeps hp_d fixed and explicitly models faint neighbours. For partially retained neighbours, the model must account for the component already subtracted through the template. A full PSF added indiscriminately for every ledger entry can double-count light. Catalogue prior accuracy must be calibrated on real sources; injected-truth prior centres are not a real catalogue calibration.

**Source roles must be explicit:** nuisance sources fit their contaminating light; contributors train the shared ePSF. Adding the former is not automatically approval to promote all faint sources into the latter. Compare any changed contribution rule independently, with leak-free initialization and the same evaluation population.

## What else remains

| Gate | Minimum deliverable | Responsibility | Status |
|---|---|---|---|
| F1 Input handoff | Ledger and scene use the same template generation, WCS, units, masks and source IDs; source completeness/partial survival checked | Two research tasks; coordination accepts the handoff | Active |
| F2 Background decision | With the neighbour-aware scene, establish whether the existing background treatment preserves flux/wings across crowding and position; only run another targeted alternative if a material bias remains | User owns the outstanding preparation decision; coordination assembles evidence | Open |
| F3 Fit reliability | Convergence/stop-reason review, alternate-start agreement and five-fold held-out validation; uncertainty, counts and failed fits reported | Coordination owns the final campaign and acceptance review | Pending pilot |
| F4 Template and kernel checks | Quantify residual cross-projection seam impact; require explicit per-band sum/normalization checks; matched-geometry Hotpants comparison and final-image refit consistency | Coordination/recovery | Partly implemented; final-generation checks pending |
| F5 Faint-star endpoint | Exact target add-back with the actual segmentation, WCS and per-band kernels; flux/uncertainty checks and difficult crowded/red cases | No separate active executor identified; coordination owns the gap | Pending final-image inputs |
| F6 Final products | One accepted recipe across the five selected FFIs, with explicit exceptions and reproducible manifests | Coordination | After F1–F4 decisions |
| F7 Figure and claim package | Every main claim tied to final figures/tables and evaluation domains; inspected measured images beside summary metrics | Existing publication workstream; coordination supplies accepted inputs | Layout can proceed now |
| F8 Manuscript and release | Full draft, limitations, primary-source citation check, coauthor review, reproducibility and submission materials | Coordination and user | Structure/method drafting can proceed; numbers await acceptance |

Targeted faint-star photometry is a distinct promised endpoint. Modelling faint neighbours to improve bright-star ePSFs does not close it. If that endpoint is removed, that must be an explicit scope decision; this plan retains it.

The original unseen-image comparison also needs either evidence or an explicitly narrowed claim: the current dataset config has `unseen_stem=null`. Five independently fitted FFIs establish breadth of the procedure, not transfer of frozen coefficients. C1/C4 need appropriate external flux-comparison coverage if the paper claims that result over all five fields.

## Sequence that avoids unnecessary reruns

1. **Now:** finish the two existing research pilots. Continue C1 recovery and preserve verified v3 products. Prepare the claim/table/figure manifest and manuscript structure while the experiments run. Do not launch a second neighbour experiment.
2. **At their handoff:** determine whether only the catalogue/scene changed or whether actual template pixels changed. If only the scene changed, retain the current verified hp_d/template inputs for the first comparison. If pixels changed, rebuild only affected downstream stages.
3. **One acceptance pilot:** integrate the ledger and scene treatment on F1, with C4 as the already-planned independent neighbour-fit comparison. F2 remains the template seam/glint stress check. Fix evaluation populations and useful improvement/no-regression criteria before inspecting the final held-out scores.
4. **Freeze once:** select the neighbour treatment, source roles and background recipe from those checks. Preserve A3, lp7s and the no-width baseline unless a specific failed check warrants changing them. Code implementation, numerical correctness and scientific acceptance remain separate labels.
5. **Final five-field campaign:** run the accepted all-star products and K=5 validation with the essential matched controls. This is a future stage, not a new launch authorization while the user is settling preparation. Regenerate fitted-WCS products only when geometry changes. Complete F4/F5 endpoint checks.
6. **Write from accepted evidence:** replace provisional plot inputs, complete the manuscript and circulate internally. No Paper 2 light-curve result is required to close Paper 1.

## Claim to evidence map

| Intended claim | Evidence that must support it |
|---|---|
| Reference preparation preserves the intended stellar signal and accounts for neighbouring light | Ledger/paired-source accounting, common-coordinate scene images and a measured bound on unmodelled light |
| Neighbour-aware chromatic ePSF/WCS fitting improves prediction | Same-star/same-pixel held-out comparison against the matched baseline; residual grids, astrometry and flux errors with uncertainties |
| Per-band matching improves the final image | Same-WCS/template-domain Hotpants control, independent faint-source evaluation, residual images and numerical flux checks |
| The procedure works beyond the headline image | Per-FFI table for all five preselected fields, with coverage, failures and limits |
| Targeted faint-star de-crowding yields useful measurements | Validated add-back, external/injection flux checks, uncertainty calibration and difficult-case images |

Existing WCS correction maps show changes to a model, not held-out astrometric accuracy. Refit agreement measures stability, not external truth. Low training loss and a visually smaller ring alone do not establish better photometry.

## Finish boundary

The goal is validated, useful improvement over matched baselines in a stated domain—not zero residuals or a unique physical explanation of every feature. Residual brightness/color/CCD structure may be a stated limitation if bounded and consistent with the actual claims. Further brightness-width terms, a complete detector-physics explanation, all-camera generality and Paper 2 temporal performance stay outside the critical path unless they are necessary to repair a failed Paper 1 claim.

Conclusion: the two active investigations should feed one controlled pilot and one final recipe. The remaining obligations are background acceptance, independent validation, seam/kernel limits, targeted extraction and manuscript production; they do not require a new round of open-ended model exploration.