Effective research goal: multi-survey embedded-protostar candidates

Effective research goal: multi-survey embedded-protostar candidates

User-directed revision, 6 October 2026. This replaces the broad ordinary-star detector objective for future work. E29–E160 remain an unmerged, reproducible archive on codex/research-snapshots-20261006; they are not evidence that the Hi-GAL project question has been answered.

Objective

User-directed update, 12 October 2026: “Proceed and update goal”, accepting multi-survey, unsupervised candidate ranking after reviewing the knowledge graph.

Over the next seven days, test whether a shared representation of infrared emission, variability and molecular activity ranks candidate embedded protostars in nearby Hi-GAL catalogue-prestellar clumps better than strong physical-feature baselines, at the same review budget and in held-out sky regions. Establish source associations, coverage, uncertainty and validation eligibility before fitting. Learn without class labels; assess with separately held-out evidence. Deliver a source-bound candidate ledger and a verified positive, null or insufficient-evidence conclusion. A successful embedding fit is not success.

The multi-survey program defines the first bounded checkpoint and the gates for a subsequent prospectively registered comparison. Reuse reviewed core inputs first; retain all requested identities, missing views, upper limits and ambiguous counterparts. This new ranking question does not measure the stopped occurrence excess or reopen N20, c6828 or E29–E160. The historical control and covariance audits below remain unchanged evidence.

The long-term detector objective remains an independently validated system for physical prestellar-to-protostellar transitions at survey scale, with precision, recall, observable-domain limits and review cost stated explicitly. The detector requirements, prepared 8 October 2026, reformulate that end state without claiming it achieved or reopening the closed empirical cycle by itself. On 8 October the user accepted the separate bounded c6828 study described below. A candidate-ranking result or a null research-cycle outcome does not complete the full detector objective.

The immediate scientific target is clump-associated embedded protostars, not generic optical stars, Gaia sources or extragalactic point-source classification. Unsupervised learning is an optional method; fitting an embedding is not an outcome in itself. Acquisition, replay checks, byte counts and successful calibration are operational evidence, not discoveries or independent validation.

Required gates

  1. Scope: name the Hi-GAL prestellar population, the physical question and how the result would change the conclusion. Reject work without that link.
  2. Prospective protocol: freeze the hypothesis, target population, control, primary estimator, comparison, uncertainty method, exclusions, pass/fail criterion, stopping condition and compute/storage caps before readout. Preserve the original registration and all amendments. Never mark all hypotheses preregistered by default or choose the primary after results.
  3. Truth and denominator: retain every requested object, unknown indicator, missing observation and association alternative. Separate catalogue labels, proxy support, physical membership, novelty and temporal birth. Scores are not class probabilities without calibration. Dust clouds and stars are not mutually exclusive classes.
  4. Confounding: test column density/extinction, distance, mass, temperature, surface density, selection and neighbouring sources. A 60-arcsecond offset alone does not remove clump-correlated reddening. A latent-class model must test identifiability and conditional-dependence sensitivity; agreement with the old 19% estimate is not its success criterion.
  5. Comparisons: use the same objects, missingness support and review cost. Include simple physical/tabular baselines with mass and distance. Keep development, calibration, consumed references and untouched evaluation separate. A proxy-conditioned metric is not population precision/recall.
  6. Claims: report the approximately 9–25% estimator point-value spread and approximately 45–55% corrected-share scenarios as descriptive sensitivity, not a confidence interval, lower bound or physical lifetime measurement. Do not quote an unqualified contamination/lifetime result before a defensible column-matched control or validated multi-indicator model exists.
  7. Reproducibility: every headline must have checked-in generation code, input and sample hashes, uncertainty assumptions and independently checked output. Cache reuse must verify sample identity and returned target geometry; changed coordinates or membership must fail closed. Preserve old artifacts.
  8. Bounded work: cached, offline analysis first. One registered scientific question per work unit, at most 30 minutes and 64 MiB new outputs initially, with a 5 GiB disk reserve. No open-ended model sweep, repeated failed network route or expansion to a new survey. An exhausted limit produces a documented incomplete result, not an automatic successor experiment.
  9. Checkpoint: assess the scientific result against the frozen question, record any changed conclusion, and commit the coherent verified change. Do not rank narrative catch-all proposals by the number of phases they list. Stop after a failed question unless a distinct, justified on-goal test is registered. Operational verification alone is not scientific progress.
  10. Publication: retain corrected core findings and their caveats together. Restore the designated Atlas to the E28 core state before selectively publishing reviewed core corrections. Keep the tranche archive separate. Preserve identity/sharing; do not publish queued E156–E160 updates. Respect the rejected E155 readback and E121 upload restrictions.

Current stopping outcome and completion

The separately registered N20 pre-fit gate has now concluded: its empirical coverage worker stopped at ambiguous repeated maser names, and an independently checked exact counterexample demonstrated non-identifiability under the tested assumptions. No latent-class fit, complete joint-support table or new physical fraction was produced. See the gate outcome. The 6 October empirical cycle stays closed, and the failed worker must not be restarted. The user’s 8 October acceptance opens only the distinct c6828 unit below; it does not reopen that failed cycle or the E29–E160 archive. No column-density-matched control is claimed completed.

User-directed continuation, 11 October 2026

The user explicitly requested continued iteration on this goal and maintenance of the blog posts with interesting results. Research execution resumes under the exact-range coalescing amendment. The earlier stopped acquisition and insufficient-evidence outcome remain historical attempt records. The same scientific sample, comparison, uncertainty, thresholds and cumulative resource limits remain frozen. No occurrence excess has yet been measured. The resumed attempt completed the 13-point profile and uncertainty gate below. Sensitivity/crowding and balance must not proceed under its failed support. Any subsequent transport or scientific stop still ends that attempt without automatic retries. The resumed native gate is now independently verified: 939/1,000 have profile matches, but 333 meet the conservative bin-uncertainty bound, failing the required

  1. The primary occurrence readout remains stopped. A separately registered post-read covariance diagnosis finds hypothetical independent-bin support of 814; that assumption is not validated and cannot replace the failed primary bound. See the new evidence. Continued iteration now concerns defensible total-column uncertainty/covariance, not a smaller primary cohort or tuned threshold. The scientific goal remains unachieved, while the original frozen measurement unit records insufficient uncertainty support.

The existing blog may receive verified compact study evidence and progress, retaining failed gates and physical-claim limits; original Atlas restrictions remain separate and unchanged.

The human’s further continuation on 11 October authorized a distinct bounded implementation uncertainty audit. Its verified static result identifies internal joint-matrix calculations followed by marginal-error output, but establishes no archived-map release binding or validated total-column error. The primary readout remains stopped. A pinned-source initialization concern is kept separate from any claim about archival map validity. This audit ends at its missing evidence; no third-party execution, reconstruction or automatic next experiment is authorized by its completion.

Historical first attempt: stopped acquisition, unanswered comparison

On 11 October 2026 the active goal was defined as testing whether nearby Hi-GAL catalogue-prestellar clumps undetected at 70 μm have an excess of compact mid-infrared sources over sightlines matched for dust column, sensitivity and crowding. The user additionally requested a public research-program introduction on the blog. See the program definition for its deliverables, stopping rules and limits. This was the initial distinct scientific attempt; the user-directed continuation above records its later repair. Its detailed prospective measurement protocol now freezes the full 19,328-object cached nearby parent, a 1,000-object outcome-blind draw within the 10,136-object buffered MIPSGAL domain, 15-arcsecond apertures, column/sensitivity/crowding controls, uncertainty and terminal criteria. The sample and input hashes are independently reconstructed. The central-column necessary gate now has 941/1,000 supported targets, independently verified from native bytes. No new source rows or occurrence comparison have been read. The terminal study outcome is now independently verified insufficient evidence: the first profile range request returned a whole-file HTTP 200 response, whose body was refused. No failed request was retried. Cached inputs support no complete thirteen-point profile, and uncertainty, sensitivity, crowding and balance remain unestablished. That first attempt concluded under its frozen transport stopping rule. Its subsequent repair required, and received, separate human continuation on 11 October; it is not an automatic retry or a scientific threshold change. Registration and input feasibility are not a physical excess result or long-term detector success. The requested independent excess measurement has not been delivered. The local insufficient-evidence report records the bounded unit’s stop; publishing that report does not supply the missing measurement or answer the comparison.

The pre-read projection amendment corrects a TAN-only assumption after a native l068 header revealed Galactic CAR. The stopped header-only run is preserved; the amendment reuses verified bytes, carries forward the original acquisition caps and changes no statistical rule. It is not evidence for or against a source excess.

The c6828 unit below concluded with insufficient evidence on 8 October; its frozen registration remains bound to the historical goal-document revision in commit 81e82fa4, not this current program amendment. Neither c6828 nor N20 is restarted. The new program preserves the closed archive and does not authorize restricted Atlas/E155/E121 actions. It can conclude positively, with a precise null, or with insufficient evidence; publication alone cannot complete it.

Accepted next unit: c6828 source association

User acceptance, 8 October 2026: “accepted proceed with goal update”, in response to the proposal for one bounded, 30-minute offline study of c6828, testing whether its changing emission can be separated from pre-existing sources using cached images. This acceptance resolves the research-reopening question for that unit. It is not approval to cancel the E155 job, restore/publish the Atlas, retry its 12-file readback or make the E121 coordinate upload.

The target is c6828, HIGALBM273.5460-1.9810, at the archived ICRS position RA 139.618 degrees, Dec −52.258118 degrees. Its brightening, incomplete pre-event coverage and existing pre-event aperture flux are consumed selection evidence; this is source-specific follow-up, not untouched detector validation.

Before new measurement readout, freeze a separate protocol and commit it. Record the exact cached native inputs, pre/post windows, measurement and association hypotheses, local astrometric/PSF controls, uncertainty method, pass/fail criteria and missing-data stopping rules. Inspect native data availability without substituting display JPEG/PNG values for calibrated photometry. Preserve ambiguous associations rather than forcing one emitter. If required native data or controls are absent, report an insufficient-evidence outcome; do not download replacements or tune the criteria after readout.

Use one offline work unit, at most 30 minutes of execution, 64 MiB new outputs, one CPU thread and a 5 GiB free-space reserve. No network observations, model training, survey expansion, N20 restart or automatic successor experiment. Independently check the result, commit a compact source-bound report and stop the unit. A source-association result cannot establish physical clump membership, second-core formation, survey precision/recall or a confirmed birth by itself. Completion of this unit does not complete the full detector objective.

Any distinct future scientific unit needs a separately justified, prospectively frozen question that could change the physical conclusion. Candidate designs include actual column-matched evidence at both positions or defensible identification restrictions for the existing indicators; availability and dependence checks must precede a fit. Missing column evidence or non-identifiability must be reported; do not substitute a generic-star benchmark. For birth detection, first establish which measured observables could distinguish second-core formation from an existing protostar burst over the same observing window. Far-IR brightness or an embedding alone does not supply that discrimination. The requirements document is a design checkpoint, not a new protocol registration, successor phase or authorization for acquisition/model development.

A research cycle completes with an independently checked on-goal answer, which may be that the hidden-protostar fraction or physical transition is not identifiable from these data. This does not certify a robust detector or a stellar birth. Any physical birth claim requires source-specific pre/post evidence, persistent compact emission, source ownership and exclusion of an existing protostar, accretion burst and instrumental explanation.

Public progress and continuation state, 11 October

The new control-support results and separately frozen covariance diagnosis are published and verified. The blog preserves the original formation explorer and failed acquisition, with twelve current exact source files and the original nine historical files. See artifacts/blog_control_update_20261011/verification.json and blog maintenance. Publication is not goal completion.

The native goal API still reports the old blocked status although the human authorized resumption and this turn completed meaningful research. The available tools cannot activate that native flag; the human /goal resume command controls automatic continuation. The local checkpoint keeps those two facts separate. The next distinct unit must freeze a defensible covariance/total-column error audit and its limits before any new readout. Do not infer independence from the 814 scenario, change the primary cohort, or open the MIR comparison early.

Continued covariance audit and resource monitoring, 11 October

The human again requested continued research and active computer/disk monitoring. The separately frozen cached covariance audit and its independently verified outcome found ETA metadata in the cached uncertainty headers, but no bound full covariance or operator/noise normalisation sufficient to validate total-column errors. A scoped exact witness also rejects positivity alone as a general justification for quadrature; it is not an actual PPMAP response or a measured covariance. The primary 333/800 gate stays failed and the occurrence comparison unopened. Continued progress needs validated integrated errors or a source-bound covariance calculation, with new scientific rules registered before readout.

Use the live resource guard for new local workers, retain its compact samples, and stop only the spawned worker on a resource violation. See storage and resource limits. No native evidence was deleted.

The cached covariance audit is published and independently verified, with all 34 live evidence/visual files byte-matched. This communication does not complete the physical research goal or unlock the stopped occurrence readout.

Accepted multi-survey program, 12 October 2026

The immediate seven-day priority is now the program linked above. Earlier text that names covariance repair as the next unit describes the 11 October control study; it does not override this update. Its 333/1,000 support versus the required 800 remains failed, the occurrence comparison remains unopened, and hypothetical 814 support under independence remains unvalidated. Do not reuse a ranking study to lower that gate, assert a hidden-protostar fraction or claim a stellar birth.

The first checkpoint is an offline, source-identity/coverage/validation-role audit of reviewed core inputs, with no fits or new label readout. Later model execution requires an exact frozen protocol and a defensible evaluation cohort. Consumed E12–E28 references may support development or explicitly labelled retrospective checks; spatial repartitioning does not make their previously inspected outcomes fresh. Keep physical associations distinct from projected catalogue matches and clump/core/source scales distinct throughout the data contract.

The available native goal API was inspected on this update: it still contains the older blocked occurrence objective. Its exposed operations create a goal only when none is unfinished, or change terminal/pause status; they cannot replace an unfinished objective or resume it. Do not falsely complete the old unmeasured objective to install this one. This document records the accepted research scope; it does not claim the application’s native goal was replaced.