Why the LHS 1140 b investigation stopped before experiments
Could bright regions on the surface of LHS 1140 explain the wavelength dependence seen during its planet’s transit, without requiring the planet’s apparent radius to vary with wavelength? AutoScience investigated whether a more general stellar-surface model could turn that possibility into a physically constrained test.
The investigation stopped before any scientific experiment ran. Independent review rejected one proposed comparison because it could reward unequal model flexibility and lacked a consequential physical criterion. A subsequent spectroscopy approach reached an independently reviewed hold because the collected observations and modelling information did not support the required inference. The original question remains unanswered.
This report records both the scientific obstacles and the shortcomings of the autonomous workflow. It documents an unsuccessful attempt to reach an executable research question, with useful screening and provenance work along the way.
The question and why it mattered
When a planet passes in front of its star, the measured loss of light depends on the planet and on the stellar surface. Bright regions, called faculae, and cooler spots can change the spectrum of the visible star. If the planet crosses a region whose spectrum differs from the rest of the disk, those differences can appear in the planet’s transmission spectrum.
The investigation focused on the short-wavelength slope in the published JWST/NIRISS spectrum of LHS 1140 b. The original observations include two transits of planet b and a transit of planet c. Stellar contamination is already part of the published interpretation. Cadieux et al.
A later modelling study incorporated surface geometry and limb darkening, the variation in brightness from the centre of the stellar disk to its edge. Its LHS 1140 application could reproduce the short-wavelength slope with stellar contamination alone, but required hot faculae and a large covering fraction: approximately a 600 K temperature contrast and 35% coverage in the reported solution. The authors regarded those requirements as extreme and favoured an atmospheric contribution alongside more modest faculae. Sumida et al.
That motivated the question recorded in the case:
Under source-faithful stellar radiative-transfer and transit models, do the available LHS 1140 b/c spectra and independent stellar constraints require a wavelength-dependent planetary radius, or can a physically admissible distributed facular surface account for the optical slope; which data or physical priors decide that inference?
The proposed contribution was to determine whether the extreme stellar requirements were robust to the permitted surface geometry and independent physical constraints. Finding another stellar-only fit would have added little: such fits already existed. Nor would replacing one equivalent circular region with many patches, by itself, establish a more plausible star. The source’s equivalent circular radius was not a demonstrated requirement that all bright material form one contiguous patch.
What the agents actually did
Sol agents developed proposals, inspected retained code and data, and performed separate review stages. Dedicated Luna readers answered bounded questions about the primary literature. The controller assembled evidence and handled dependencies. Scientific execution required independent checks of significance, novelty and design; those checks never produced an admission for this project.
The work recovered the published spectral table, retained a fixed version of the author’s modelling code, extracted scientific configuration without executing the full driver, and examined the provenance of possible stellar constraints. It also established limitations that a valid experiment would have to address.
| Work item | Disposition |
|---|---|
| Geometry-based minimum-coverage comparison | Rejected before execution on design and significance grounds |
| Spectroscopic constraint on heterogeneous stellar surfaces | Independently reviewed external-input hold |
| Admitted scientific hypotheses | 0 |
| Scientific experiments or confirmation runs | 0 |
| Answer to the original attribution question | Unresolved |
There were eleven recorded native case invocations, covering refinement, review and one failed intake. They were not eleven experiments. The interval from the first recorded case invocation to the final independent hold was just under five hours, including source work, waiting and technical recovery. That elapsed time is a workflow measurement, not a measure of scientific computing throughput.
Why the geometry proposal failed review
The first developed proposal sought a minimum-coverage boundary: how little projected bright area could reproduce the full numerical spectrum under specified model assumptions? It would compare a distributed surface with a restricted circular-region family while keeping the radiative prescription explicit.
Independent review identified an unfair comparison. The distributed family allowed continuous fractional filling, while the circular baseline used nominal-area steps of 0.01. The proposed test could count an area advantage above 0.002 as evidence of a geometry effect. An apparent advantage that small could instead come from the baseline’s coarser area choices or the more flexible filling prescription. No frontier was computed; this was a defect in the proposed experiment.
A matched, continuously variable circular baseline could have repaired that comparison. It would not have repaired the more important problem: the proposal had no independently justified physical boundary that made a numerical improvement consequential. Reducing the required coverage would matter to the parent question only if it changed what the observations or credible stellar constraints allowed.
An earlier draft’s suggested 0.05 coverage reduction likewise lacked a physical justification. The later proposal abandoned that particular threshold, but still did not establish a meaningful admissibility criterion. Review therefore rejected the geometry formulation instead of approving a more elaborate calculation of the same insufficient claim.
The code audit retained a useful diagnostic: in the specified static model, with no active regions crossed by the planet and intensity depending on radial position, rearrangements that preserve the exact radial filling distribution preserve the light curve. Fragmentation alone is therefore not necessarily a discriminating test. This conditional identity was an implementation check, not a new empirical result about LHS 1140.
Why the observational inputs were insufficient
The simulated depth was not established as the measured quantity
The retained renderer produces a maximum normalized loss of light. The published spectrum contains transit-depth estimates from an observational analysis. The investigation did not recover enough of the fitting and averaging procedure to establish that the renderer’s output predicted those tabulated estimates, or to bound the difference.
That distinction matters for a star with an uneven surface. A fitted light-curve parameter and the deepest point in a simulated transit need not respond identically to stellar structure. Matching the published numbers under a stipulated convention could define a conditional model exercise, but the proposed exercise still needed an independent justification for its scientific value.
There was also a visit mismatch: the supplied spectrum of planet b combines observations on 1 and 26 December 2023, while planet c was observed in the second visit. A joint stellar-surface test could not simply treat both published spectra as measurements of one simultaneous surface. The targeted source reviews did not resolve the necessary visit and measurement transfer. These are limitations of the reconstruction achieved in this investigation, not a demonstrated error in the published analysis. Cadieux et al.
Temperature and luminosity did not supply independent surface bounds
The agents next examined whether stellar temperature, luminosity or spectroscopy could constrain the bright-region coverage.
The source audit traced the quoted temperature, 3096 ± 48 K, to a single-component spectral-line fit. It could not be identified directly with the temperature of the quiet component of a mixed surface. The quoted luminosity was derived rather than an additional independent bolometric measurement. Treating both quantities as independent constraints would therefore have overstated the information available. Cadieux et al., stellar characterization
The retained facular model also did not establish the bolometric normalization needed to turn its component labels into a reliable total-luminosity test. A useful thermal or spectroscopic constraint would require a calibrated prediction of what a mixed surface produces in the actual measured observable, with its uncertainty and parameter dependencies included.
The spectroscopy came from a different year
The targeted observation audit dated the 29 NIRPS spectra to 26 November–6 December 2022, approximately a year before the December 2023 transit observations. The inspected source supplied no quantitative stellar-surface stability bound connecting those epochs. Cadieux et al., stellar characterization
Earlier spectroscopy might still constrain component physics shared across epochs. That possibility remained open. However, the collected material did not provide a validated heterogeneous-surface interpretation or demonstrate such discrimination while allowing the surface fractions to change.
The cited temperature-method paper concerned Barnard’s star observed with SPIRou. It supplied relevant methodological context, but did not establish the exact LHS 1140/NIRPS line data, normalization, instrumental response and uncertainty model needed here. Following a citation to another star’s method did not supply the target-specific test. Jahandar et al.
Where the harness fell short
The independent checks prevented an unsupported experiment from being presented as a discovery. They also exposed weaknesses in project selection and orchestration.
The initial feasibility screen was too shallow. Public spectra and available modelling code made the project appear tractable, but the decisive dependencies were the measurement definitions and independent physical constraints. Those should have been investigated earlier, before developing a detailed numerical comparison.
The first proposal confused additional flexibility with scientific discrimination. Review caught the unequal baseline and the absence of a meaningful physical threshold. A stronger proposal generator would have asked, at the outset, what conclusion would change if the more flexible model fitted better.
Execution monitoring was initially inadequate. An early worker returned an invalid output structure. Later, workers launched from inside a sandboxed model controller could not read their evidence because of a runtime socket error. The run remained at a technical standstill until the operator requested an update. The orchestrator had not connected the case to effective persistent oversight, despite an existing monitoring harness. The operator then explicitly required active orchestration.
The repair used a host-level dispatcher and verified actual worker file access. Five-minute oversight was added through the existing watchdog and controller mechanisms. Further faults appeared in scheduled operation: the cron environment could not locate the model executable, normal resumptions consumed the launch allowance, and a legacy retrieval route could start a worker that immediately exited on an external dependency. Those faults were corrected and checked. They were operational repairs; none resolved the scientific gaps.
Paper-reader validation also failed. One methods report exhausted its allowed citation-validation attempts. The controller preserved only explicitly qualified excerpts whose individual references passed the available checks, and independent closure review treated them as partial evidence. Their inclusion did not make the failed report complete or establish the target-specific calibration.
The workflow took too long to turn accumulating prerequisites into a stopping decision. Several rounds clarified provenance without producing a valid discriminating experiment. The controller eventually stopped expanding the calibration-paper chain and requested an independent assessment of the hold. Future selection should test the critical observable and physical constraint earlier, and require each additional source task to explain how its answer could enable or rule out a concrete experiment.
This case therefore supports a limited assessment of the harness: it rejected a flawed formulation and preserved important distinctions between supporting work and scientific evidence, but needed operator intervention and repeated operational repair to reach that point. It did not demonstrate successful autonomous discovery.
What would justify reopening the project
Reopening requires a concrete change to the evidence or the scientific formulation. Useful inputs could include target-specific spectral line or template products with the normalization, instrumental response and uncertainties needed for a heterogeneous forward model; contemporaneous stellar observations or a validated link between epochs; and the actual transit-fitting and visit-combination procedure, or a validated equivalent.
A different route could also be worthwhile if it established an independently justified physical constraint and a consequential test with accessible inputs. It would need fresh significance, novelty and design review. Another unconstrained fit, an altered numerical grid, a generic calibration citation or a new arbitrary threshold would not repair the reasons for the current hold.
The final disposition is a hold on the investigated spectroscopy route, with the earlier geometry proposal rejected. It does not establish that stellar contamination succeeds or fails, whether LHS 1140 b requires a wavelength-dependent radius, or that all possible approaches to the question are exhausted.
Evidence and scope of this report
This account follows the retained AutoScience proposals, execution records, source-reader reports and separate model review invocations from 4 November 2026. “Independent review” here means a separate AI review invocation with access to the recorded claim and evidence; it does not mean external human peer review. Full-paper questions were delegated to targeted readers, and their inspected scope and limitations were retained. The report makes no claim of exhaustive literature or data-archive coverage.
The principal public sources are linked alongside the claims they inform. The modelling-code audit used the author’s ECLIPSE-Xλ repository at a fixed commit. Those sources describe the underlying research; the proposal rejection, execution failures and stopping decision belong to this AutoScience investigation.