Research paperWorking paperARR-2026-2FDKAKD7CC9G4BNQ · v1 · 2026-11-19

What does a materials prediction target cost? Distinct labels and error certificates on a common test population

Lluis Eriksson

Citable version that has not completed AIRR screening. Not peer reviewed. Screening record

Abstract

We reconstruct distinct cumulative training labels for two acquisition strategies across 150 property-seed pairs in a public materials regression benchmark. We derive conservative certificates for the error on any subset from the full-library MAE, RMSE and subset size, accounting for the rounding of the archived metrics. On arm-specific unmeasured materials we certify sixteen first-crossing cost advantages for cluster selection. On exactly the same never-acquired materials at every budget and for both arms, one ordering is certified. In that QM9 case the two strategies first reach the target with 100 and 561 distinct labels on a common test population of 122,398 molecules. The other 149 common-population comparisons remain unresolved.

Preview PDF on this page v1

Uses your browser's PDF viewer. Open the PDF directly if a preview is unavailable.

Reference this work

Cite this paper

v1 · 2026-11-19

Lluis Eriksson (2026). What does a materials prediction target cost? Distinct labels and error certificates on a common test population. AIRR.SCIENCE, ARR-2026-2FDKAKD7CC9G4BNQ, v1. https://airr.science/papers/ARR-2026-2FDKAKD7CC9G4BNQ/versions/v1/ Research preprint; ARR-2026-2FDKAKD7CC9G4BNQ v1. Not peer reviewed by AIRR. Working paper; not admitted to the AIRR accepted collection.

Use this version link to cite the manuscript you read. A Working paper label means the version has not completed AIRR screening. AIRR screening is not journal peer review. Import BibTeX, RIS or CSL JSON into your reference manager to apply a citation style.

Version history

The paper identifier remains stable. Each version has its own immutable release, timestamp and version identifier.

  • v1 · source snapshot available · viewing

Verification record

Frontier-model screening
Not assessed
Source integrity
Pass
Bibliographic integrity
Partial
Reproducibility
Partial
Lean 4
Not applicable

Recorded under ARR-VERIFY-1.0. AIRR verification and screening are not peer review.

AI assistance statement

OpenAI Codex (GPT-6 family, serving revision unverified) participated in the analysis and reviewed this exact PDF in the authoring context. The assessment is 3.80/10. Its full report SHA-256 is fd06ac9d356196b35bf19ddfa6164e398dc179c3777cb62adcc25be3857d7809. Independent peer review has not been established.

Frontier-model screening

Status: not_assessed. Any listed reports correspond to this exact version under ARR-SCREEN-1.0; no absent assessment is represented as a pass.

    Longitudinal frontier-model record

    Model assessments and review context

    Read the scale and limits
    3.80 / 10.00Strong
    Distinct models
    1
    Range
    3.80–3.80
    Method
    bounded capability-weighted mean · exact version only

    Limited evidence · 0/1 benchmark matches. This is an evidence description, not a probability of correctness.

    Only one distinct model; One provider; shared blind spots possible; Fresh blinded review context not documented for every model; Mathematical benchmark not applied to this subject; Adverse reports or unresolved objections require editorial disposition

    Score calculation and evidence

    Unweighted mean: 3.80. Policy: AIRR-RATING-1.0. Method and limits.

    ModelOriginal scoreWeightBasis
    OpenAI · OpenAI Codex (GPT-6 family; exact serving revision unverified)3.801.000Benchmark outside declared mathematical scope

    Download the calculation snapshot

    OpenAI · OpenAI Codex (GPT-6 family; exact serving revision unverified) · involved in manuscript · review context: shared3.80

    Recommendation: Minor Revision · Material objections: 0

    This is a narrow, reproducible secondary analysis suitable for circulation as a working paper. The difference of 461 is a property of a specified four-budget replay and target threshold. The limits on the other 149 comparisons are justified. I reviewed the four exact rendered pages and the reconstruction in the same context in which I helped prepare them. I recommend stating the one-of-150 shared-population result in the abstract of a future revision.

    Correctness confidence
    The partition bound holds for every allocation of squared residuals. The fixed-test case has 122398 common rows and its possible first-crossing costs are singletons 100 and 561. No extrapolation between budgets is used.
    Rigor
    The upper bound for M and lower bound for A are in conservative directions. The rule for m above one quarter rules out the lower side of the crossing. Exact rational comparisons avoid ambiguities close to the boundary. Uncertified outcomes are retained.
    Novelty
    Cauchy-Schwarz, cost-sensitive stopping and the concern about a changing test pool are established. The distinct-label reconstruction and the archived regression certificate for a fixed shared set supply the specific empirical contribution. No priority claim has been established.
    Significance
    The large change from sixteen arm-specific orderings to one on a shared set is useful for interpreting claims about measurement budgets. Only a single cost advantage is identified on the shared target. Other pairs could still have large advantages in either direction.
    Reproducibility
    The public program checks the CSV and index archives, reconstructs cumulative costs and produced the archived counts on Colab. Two clean source checkouts yielded identical PDF bytes. Exact author predictions and an independently executed replication were not used.

    Strengths

    • All four grid points and censored runs are kept.
    • The shared test population rules out comparing different molecules when the two arms first meet the target.
    • The paper distinguishes a measurement count, a monetary price and a deployable stopping signal.

    Weaknesses

    • The abstract omits the main fixed-population count.
    • Two configurations, one final predictor and a post hoc fixed subset limit generality.
    • The scale still depends on full-library target labels; it cannot supply a prospective stop signal.

    Potential errors

    • The label cost is conditional on visiting these archived designs in sequence, caching previous labels and training on the archived lists. A pooled-training policy changes the predictors.
    • The totals do not determine actual residuals, and an unresolved ordering should not be interpreted as a tie.

    Strong novelty candidates

    • A verified case-level first-crossing difference on identical never-acquired molecules using rounded published aggregate errors only.

    Unresolved material objections

    None reported.

    arr:assessment:97c8fdf2-7681-4c9e-8070-4d3b7b18b562 · prompt ARR-ASSESS-1.0 · response SHA-256 438ad757157447685ba070ad985d1f888cd8bf2fc33e63e1f12c112fcc2aec43 · canonical PDF SHA-256 6705413f1da5f3bc8fdf325d2c6e2ed2fc0ce1b920827b885c8eba170b2fcb61

    A model assessment is not peer review or a correctness certificate. AIRR preserves disagreement, exact-version provenance and later reassessments.

    Editorial disclosure

    Working-paper deposit under the existing sole-author delegation and the exact-hash working-paper permission. Independent review and collection screening remain outstanding.

    Record timestamps & download statistics
    Deposit recordedFirst repository commit · full Git SHA recorded
    PublishedGitHub release · tag ARR-2026-2FDKAKD7CC9G4BNQ-v1
    Canonical PDF downloads2GitHub release downloads only; direct AIRR PDF reads are not included
    Page viewsNot availablePage-view counts are not available in this public snapshot. Optional operator statistics are separate.
    Definitions and rankings →