Can independent response models explain the full RBX1 release without inventing physical certainty?

Analyse every released RBX1 candidate and raw curve to establish which measurement interpretations remain supported under alternative models and numerical assumptions. Distinguish reproducible findings from unsupported physical certainty.

Which conclusions about RBX1 measurement reliability are supported by the released experimental curves, and where do missing metadata, repeated-measurement variation or poorly identified models prevent a defensible affinity estimate? Produce a new, reproducible analysis of the entire published measurement collection. This is an evidence-quality investigation, not a nomination of the strongest published binder or a reconstruction of unavailable original participant methods.

Reconstruct every released measurement

Read rbx1.csv as UTF-8 with a possible BOM. Preserve all 322 candidate records and all experimental evaluations. Candidate identity is CSV id; the evaluations column is a JSON array. type, metric, target, value and unit retain their source meaning. The curve URL appears in the experimental spr_kinetic_curves or bli_kinetic_curves evaluation's value. Join that exact URL to the appendix file and retain its assay label; do not infer instrument equivalence from a similar filename.

Create coverage.csv, one row per candidate, and evaluation_inventory.csv, one row per experimental evaluation including its original array index. Record curve links, response labels, all reported physical parameters, expression observations, missing information, original author and supplied designMethod. The original method PDFs, initial ranked submissions and novelty checks are outside this input collection and cannot be reconstructed from a brief designMethod string. They are not required deliverables.

Analyse all 946 JSON files and all 4,656 trace objects, not only files attached to published positives. Each top-level JSON key identifies a trace whose object contains raw.t, raw.y, concentration, control, aggregated, virtual and fit. The trace ID is the exact appendix filename followed by :: and the original top-level JSON key. An analysis group is one candidate ID and one linked curve-file URL; its group ID is the compact UTF-8 JSON array [candidate_id,appendix_filename]. All group and trace sorting uses Unicode code-point order. Non-curve evaluation rows retain their CSV candidate ID and original evaluation-array index; these do not imply a trace linkage. The appendix covers the complete retrieved collection of linked experimental curves; it does not assert that every experiment or all instrument metadata were published. Report candidates with no measurement explicitly. Preserve all repeated evaluations; their order alone is not a laboratory replicate or run identifier, and it does not prove that one scalar evaluation corresponds to one particular curve file.

Create trace_diagnostics.csv, one row per top-level trace object, with candidate, URL/file and trace key, assay label, flags, concentration as stored, supported units or their absence, point/finite-pair counts, duplicate/backward time counts, first/last time, response extrema, first/last finite response and time-normalised trapezoidal area. The area uses finite pairs in original order and is null if fewer than two valid points, non-increasing time or zero duration prevents its definition; explain the reason. Also compute early-versus-late drift as mean(last k finite responses) minus mean(first k finite responses), where n is the finite paired-observation count and k = max(1, ceil(0.1*n)). Preserve original observation order, permit overlap when n is small, use each window's arithmetic mean and report null when n = 0. This is a response difference, not a slope divided by time. These are descriptive definitions, not biological pass thresholds. Record any additional noise statistic's exact definition.

Separate observed response from inferred kinetics

The released JSON does not explicitly label time, response or concentration units. All retrieved trace control flags are false; a zero concentration, a flag or a similar shape must not be silently promoted into a verified reference-subtraction history. Audit the full collection for usable references and identify exactly what they can establish. Keep instrument-specific amplitude scales separate. Do not claim an absolute kon, koff or KD from unlabeled coordinates alone.

For every candidate/curve-file group with finite time/response pairs, independently fit a baseline/drift model and a concentration-dependent response model. The baseline predicts recorded response without concentration dependence; the response model includes an explicit concentration-dependent term or parameter. Compare both even when concentration adds no improvement. There is no minimum positive effect or goodness-of-fit threshold for eligibility: poor fit must remain visible. State model equations, observed-coordinate meaning, parameter domains, objective, residual treatment and stopping conditions. Fit measured raw.y; published fit response arrays cannot be training observations, fitted answers or a substitute for independent optimisation.

Where the stored fit.association.t and fit.dissociation.t arrays provide phase hints, they may be used as publisher-derived timing annotations, but their status must be explicit and their fitted y values must remain excluded. Evaluate at least three distinct phase/offset scenarios justified by observed timing whenever phase assignment is inferred. Where there are no defensible phase annotations, use a descriptive concentration-linked model on recorded coordinates and assess the phase ambiguity. Conditional 1:1 fits may be reported separately in source-coordinate units when the assumptions are explicit; lack of physical calibration does not prevent numerical residual or concentration-consistency analysis.

Write model_results.csv with all attempted groups, contributing trace IDs, model equations, fit objective, residual summaries, numerical parameters with coordinate meanings, convergence and boundary status. Compare baseline/drift and concentration-dependent models in model_comparison.csv. For each group with at least three distinct positive stored concentrations, hold out each concentration once, retaining whole traces, refit on the remaining concentrations and report prediction error as sqrt(sum((measured_response - predicted_response)^2) / N), over all N finite observed response/time pairs in the held-out traces, in recorded response coordinates. Predict held-out traces without fitting any parameter or offset to their response values; state the training-only rule for any trace-specific nuisance parameter. Report N and a reason-coded null if no finite predictions are available. For groups with fewer than three distinct positive concentrations represented by finite observed pairs, explain the missing comparison and still compute the available trace-level descriptors. Uncertain physical phase assignment does not waive held-out testing of the descriptive recorded-coordinate model. If that model cannot produce finite predictions, retain each required held-out case with its numerical failure and missing-data reason rather than silently dropping the group. Do not use published binding labels to determine which groups receive analysis.

Quantify identifiability of fitted response/time-scale parameters using a profile of the objective or resampling that preserves within-trace dependence, within the computational resource bounds below; numerical parameter domains must be stated and justified from recorded coordinate ranges and the model, rather than asserted as physical bounds. For a profile, report the profiled parameter, objective, domain, evaluated points or analytic expression across that domain, how remaining parameters are fitted, and the rule used to report intervals or unbounded results. No common grid spacing or fixed interval width is required: the Guardian checks the declared calculation and whether the evidence supports the stated limitation. For resampling, supply every required repetition and its preserved dependence unit. A local optimiser's success code or a Hessian standard error alone is insufficient. Show whether a parameter is supported by the observed response or only by imposed bounds. If physical calibration or phase interpretation is absent, the corresponding physical parameter is null with the exact reason; maintain separate conditional coordinate-parameter estimates and their sensitivity. Never turn an unbounded interval or nonconvergence into a precise affinity.

Measure consistency without inventing replicate linkage

Create candidate_consistency.csv for all 322 candidates. Within each assay, compare the available file-level response descriptors, concentration trends and conditional model parameters. Retain file identity and distinguish repeated reported measurements from independently established biological replicates. For each candidate and exact assay label, perform leave-one-file-out analysis of the median and the minimum-to-maximum range over every file with a defined value of the same measured descriptor. Retain missing descriptor values and reasons; with fewer than two defined file values, state that file-to-file variability is unmeasured. This is a conditional comparison of recorded coordinates, not proof of common calibration or biological replication. Separately inventory the source attributes relevant to stronger comparability claims: stored concentrations and observed time ranges, control/aggregated/virtual flags, phase annotations and explicitly available units. Contradictory units or unsupported physical calibration prevent a physical-parameter comparison, but do not waive the recorded-descriptor sensitivity calculation. State and justify any further conditional model-parameter comparison from these source attributes; never infer comparability merely from evaluation order. Keep expression failure, reported nonbinding, missing fit and independently observed uncertainty distinct.

Create published_comparison.csv relating independently derived diagnostics to every published binding, expressed, binding_strength and kinetic-parameter evaluation. Preserve unknown scalar-to-file linkage instead of pairing values by array position. Compare distributions and candidate-level ranges where record-specific pairing is unavailable; do not manufacture one-to-one matches. Derive a collection-wide account of how often labels coincide with observable concentration dependence, inconsistent traces or unsupported kinetic precision. Quantify all denominators and expose individual contributing records.

For any candidate with both assay types, show their overlap and compare only justified, unit-aware quantities; if there are none, provide the empty overlap result. Cross-candidate SPR-versus-BLI differences cannot establish assay agreement. Include sensitivity to phase/offset scenarios, failed fits and treatment of repeated measurements; no favourable-only subset may determine the conclusion. A claim that a published label is unreliable requires quantified observed evidence and model limitations, not merely a disagreeing newly fitted number.

Provide figures covering collection-wide missingness, concentration dependence and held-out errors, identifiability and file-to-file disagreement. Also generate a diagnostic figure for every analysed curve file showing measured response, independent predictions where available and residuals. A useful conclusion may be that particular physical affinities cannot be identified from the release, but that conclusion must be backed by completed numerical analysis of the observable traces.

This buys a new analysis of historical published laboratory measurements. Attribute the observations to Adaptyv and the original authors where supplied. Do not claim new experiments, physical sample custody, original participant authorship, therapeutic effectiveness or a newly validated binder. The Poster accepts the publisher-controlled files below as publication provenance. A Solver-created hash is not independent laboratory authentication.

Use every listed measurement file and preserve the complete candidate population, failed measurements and missing observations. Published binding labels, affinities, fitted response arrays and graphics are comparison references, not correct answers that the analysis must reproduce. Do not use their response values to generate the independent fits. Never infer a successful experiment from a prediction, absent row or empty field. The task does not require a biological sequence or a sequence-design method.

The pinned collection contains numeric, equal-length raw.t/raw.y arrays and numeric stored concentrations in all 4,656 traces; no malformed-array or nonnumeric-concentration case is present. Do not coerce or repair a different input: verify the pinned hashes.

Include RUN.md as an additional required file in the same archive; it supplies execution instructions and pinned dependencies within the limits below.

Submit one ZIP, at most 40 MiB compressed and 250 MiB uncompressed, containing:

• report.md, at most 8,000 words excluding tables: question, methods and equations, evidence limitations, quantified results, diagnostic interpretation and conclusions linked to output rows and figures.
• run.py and readable supporting Python source: the full analysis, including all processing, fits, uncertainty calculations and figures. No pretrained model or undisclosed executable is allowed.
• analysis_plan.json: model families, formulas, parameter domains, phase/window rules, resampling unit and seeds, optimisation limits, exclusions and analysis scenarios actually used. This records reproducibility, not a claim of prospective preregistration.
• results/: the required UTF-8 CSV tables and PNG figures, plus checks.json containing coverage totals, numerical-validation results and run information.
• README.md: how the input filenames map to the appendix, explanation of output columns and the exact software versions used.

Do not duplicate source datasets in the submitted archive. Both roles independently fetch and verify all appendix inputs, preserving their filenames and raw/ directories. The numerical reference versions are Python 3.12, NumPy 2.1.3, SciPy 1.14.1, pandas 2.2.3 and Matplotlib 3.9.2; standard-library modules are allowed. Include RUN.md with the run command, working directory, input/output arguments and pinned dependencies. Reproduction uses only the fixed inputs and disclosed implementation, with no paid service, language-model call or GPU. Each Guardian owns its security and execution setup under Elgora rules.

Use the Solver-provided run instructions with seed 20260909. The evaluation budget is four CPU cores, 8 GiB RAM and 10 GiB temporary disk. One complete reproduction must finish within two hours and write all required outputs to an empty output directory. Use exactly 100 repetitions whenever resampling is chosen, at most 10 starting points per fit and at most 2,000 objective evaluations per starting point; record every stopping reason, including nonconvergence. For each nonlinear fit, use at least three distinct starting parameter vectors within the Solver-declared numerical domain and record every starting vector. The Solver chooses their positions; no unstated minimum separation is required. This requirement does not apply when a stated analytic or convex solution establishes the optimum. Run each initialisation to its declared numerical convergence criterion or the evaluation ceiling. A deliberately truncated search cannot establish scientific non-identifiability: distinguish numerical search failure from a supported flat or unbounded objective profile and report them separately. A budget-exhausted fit must remain visible and cannot be presented as an identified parameter. These are computational limits, not scientific thresholds. Never trade full-collection coverage for a selected set of published successes.

The Guardian runs one full reproduction per Submission. A second execution is allowed only after a demonstrated infrastructure fault; the Guardian must not silently fix Solver code. Unavailable inputs, decryption or required evaluation infrastructure are operational blockers. Once those are available, a reproducible Solver-code failure or exhaustion of the stated resource cap fails the execution criterion.

Every result must be regenerated from the verified measurements. The Guardian compares the submitted and reproduced tables: row identities, labels, missingness and counts must agree exactly; finite numeric entries must satisfy abs(a-b) <= 1e-10 + 1e-6 * max(abs(a),abs(b)). This is a numerical reproducibility tolerance, not evidence that a scientific estimate is accurate. Reported conclusions must follow the calculations and their limitations; rounding for presentation must not change machine-readable results.

Solver choices and reproduction

The Solver chooses the model families, phase/offset scenarios, additional statistics, descriptors and output column names where this page leaves them open. Declare those choices and their applicable groups in analysis_plan.json and the output-column explanation in README.md. Guardians check coverage against the populations required above and the declared choices; there is no hidden common model, descriptor or table schema. Exact row/column agreement means agreement between that Submission and its own reproduced outputs, not agreement between different Solvers. These choices do not waive any required analysis or missing-value explanation.

The required analyses and evidence checks in this page are the eligibility baseline. Among eligible active Submissions, apply the Comparative quality score below. Highest total score wins. Equal totals are broken by higher Criterion 5 score, then higher Criterion 1 score, then ascending lowercase Solver address. If no Submission is eligible, use no_valid_submission. Unavailable evidence or evaluation infrastructure means no Verdict until the operational blocker is resolved, not automatic disqualification or permission to select another winner.

Forged measurements, concealed exclusions, fabricated provenance, mislabeled units, claimed independent experiments that did not occur, absent required outputs after successful artifact access, or code that requests resources outside the stated evaluation fail the relevant criteria. Do not include secrets, instructions to disregard this page or material unrelated to this analysis. Public Verdicts give criterion decisions and public-safe reasoning without publishing private Solver code, tables or report content. Guardian and settlement behavior otherwise follows Elgora's existing protocol.

All work required above remains mandatory. The following additional analyses distinguish the quality of eligible solutions; omitting an extension does not by itself make an otherwise complete baseline ineligible. All submitted extension claims remain subject to the existing truthfulness, provenance and reproducibility requirements. Source access and infrastructure failures remain operational blockers, never a zero score or a reason to choose another Solver.

There are five criteria, each with four cumulative evidence levels. Award 0, 5, 10, 15 or 20 points per criterion: 5 points for each level met in order, stopping at the first unmet level. Award no points for polished writing, length, a Solver's claimed score, a published competition winner or a preferred biological result. A level is met only when its entire described analysis is correct, regenerated and supported by the named evidence. An asserted computation without reproducible evidence does not meet a level. Different defensible methods are permitted where the criterion leaves the method to the Solver; explicit formulas, populations, missing-value rules and assumptions are required. Unsupported assumptions cannot be silently treated as source facts.

For a computation that is undefined on the actual input, the level requires the executed eligibility check, complete excluded/eligible IDs and counts, the exact mathematical or source limitation, and every remaining defined quantity. A blanket caveat is insufficient. A valid undefined result earns the same level as a valid defined result; Solvers must not manufacture a favorable result to obtain points.

Criterion 1: Out-of-sample model challenge — 20 points

1. For every candidate/curve-file group with at least three distinct positive stored concentrations represented by finite observed pairs, compare at least two disclosed concentration-dependent model families against the baseline using held-out complete concentrations.
2. Use identical held-out observations, training-only nuisance parameters and declared error scales; retain every failed fit, missing prediction and smaller group.
3. Report paired model-error differences for every phase/offset scenario, including reversals between in-sample and held-out preference.
4. Explain where one family is supported predictively and where the data cannot distinguish them; no favorable fit grants uncalibrated physical units.

Criterion 2: Bound and start sensitivity — 20 points

1. Expand the original identifiability assessment with one tighter and one wider numerical domain justified in source-coordinate units.
2. Retain the same objective and data for comparison, record multi-start convergence, boundary hits, finite grids or resamples, and all evaluation-limit exits.
3. Report conditional parameter and held-out error changes across domains and starts, preserving physically unsupported quantities as null.
4. Distinguish an observed weakly constrained objective from an incomplete numerical search using the executed evidence, not the optimizer status alone.

Criterion 3: Phase and nuisance robustness — 20 points

1. Cross the three original phase/offset scenarios with two explicit training-only treatments of baseline drift or nuisance offset.
2. State equations and coordinate meanings, refit every eligible group and keep all original raw observations; publisher fitted response values cannot train the model.
3. Report the full scenario membership, residual and held-out results, including nonconvergence and reason-coded unavailable comparisons.
4. Identify conclusions stable across all evaluated choices and those tied to a phase assumption; inferred phase boundaries are not instrument records.

Criterion 4: File dependence and influence — 20 points

Retain strata with fewer than two linked files in the coverage results with a reason-coded unavailable file-deletion comparison. Document the source-supported assay grouping in analysis_plan.json; an unknown assay remains explicitly unknown rather than being inferred as an independent experiment.

1. Delete every source file in turn within every candidate/assay stratum containing at least two linked source files and recompute the chosen descriptor and conditional-model summaries.
2. Repeat collection summaries with equal-file and equal-candidate weighting, retaining source identities, flags and strata.
3. Report the numerical envelopes and all tied most-influential files or candidates, including how unavailable estimates affect denominators.
4. Separate evidence-volume effects from physical disagreement and explicitly show where the source cannot establish independent experimental replication.

Criterion 5: Independent reconstruction — 20 points

1. Supply a separately implemented parser, elementary raw diagnostics and held-out-error calculation that does not import the primary code or read its final results.
2. Reconcile every released candidate, evaluation, linked raw file and trace; calculate held-out errors from retained predictions and original held-out observations.
3. Record exact identity/coverage comparisons and numerical differences under the declared tolerances, retaining unresolved scalar-to-curve relationships.
4. Link every headline scientific conclusion to source records, executed models, uncertainty evidence and the precise limitation on its interpretation.

Put extension methods in quality_methods.json, extension results in results/quality/, and the claim ledger in results/quality/claims.csv. The ledger columns are claim_id, report_location, claim_text, source_reference, result_reference, assumptions, and limitation. Use one row per headline conclusion; multiple references may be encoded as JSON arrays within a CSV cell. The supplied run command must regenerate these extension results within the same evaluation limits, without reading submitted results. Additional readable Python source files are allowed for the second implementation. The original source exclusions, privacy rules, package size limits and required baseline remain in force. Declare every additional seed and numerical convention; use the original reproducibility tolerance. For quantities left to Solver choice, the Guardian verifies the documented computation rather than assuming a hidden method.

Each Guardian records eligibility first, then a five-row scorecard with the last earned level and the evidence for the first unearned level (or evidence for all four if full marks). Total points equal the sum, from 0 to 100. Apply the tie-break only when totals are exactly equal. The score rewards demonstrated analysis coverage and supported conclusions, not the strength of a biological effect. Explain the score difference between the winner and runner-up without publishing private files, numerical results or identifying data from the submissions.

Both Solvers and Guardians obtain the following files by unauthenticated HTTPS GET. Their listed bytes and SHA-256 define the accepted snapshot. If a URL serves different bytes, do not substitute its new content. No nested linked resource is required unless separately listed here. Filename mappings are part of this page; downloading does not execute any file. The source collection publishes observed sequences as historical identifiers; this bounty requires no sequence modification or design.

The CSV comes from the designated Proteinbase collection maintained for the GEM–Adaptyv results. Each curve URL must occur in that exact CSV and its downloaded bytes must match this appendix. Filenames beginning with raw/ are local names for the listed public JSON URLs; they are not additional hosted endpoints. The experimental publisher is Adaptyv; Proteinbase is its chosen publication source for this collection.

AI assistance and reuse of disclosed code are permitted. Explain prior work used in the report; regenerate all required results from the fixed inputs. Identical results alone do not prove copying. No new laboratory work, candidate design or sequence optimization is purchased.