EGFR measurement reliability: independent kinetics and functional-evidence analysis

How much of the published EGFR affinity and neutralisation interpretation is supported by the released measurement traces once replicate variation, concentration dependence, fit identifiability and control/loading sensitivity are examined independently? Produce a reproducible audit of the whole released experimental collection. Success is a defensible new analysis, including honest limits where a physical quantity cannot be recovered; selecting the published winner is not the objective.

Reconstruct the evidence

Create coverage.csv with one row per result-summary name and additional clearly marked rows for otherwise-unmapped sample names. Link summary name, replicate name, sequence-similarity id, and measurement ZIP member names explicitly. After removing the .csv suffix, the penultimate underscore-separated filename field is the replicate identifier and the final field is the stored concentration: <sample>_<replicate>_<concentration>. The sample is the entire preceding prefix, including any earlier underscores. Apply this same field order to kinetic and neutralisation trace filenames; reading_data.csv is metadata, not a trace filename. Preserve the two reference controls as controls. Account for all 402 summary rows, all 953 replicate rows, all 3,815 kinetic CSV members and every neutralisation member. These totals are completeness checks on the fixed source, not biological claims. Show candidates and replicates with no released curve instead of implying they have measurements.

Create trace_diagnostics.csv, one row per time/response CSV member. Record input/member identity, sample, replicate, concentration as printed, supported units or their absence, number of points, finite-pair count, duplicate/backward time counts, first/last time, response minimum and maximum, first/last finite response, and time-normalised trapezoidal area. The area uses finite pairs in their original order and is null if fewer than two valid points, non-increasing time or zero duration prevents that definition; report the reason. Also compute early-versus-late response drift using the means of the first and last 10% of finite observations, rounding the count upward. These window choices are descriptive audit definitions, not laboratory acceptance thresholds. Additional noise and residual diagnostics must disclose their definitions.

Independently fit and challenge kinetic interpretations

A usable curve has at least three finite paired t,y observations, strictly increasing finite time and positive duration. Nonfinite pairs are excluded with their count retained; no further exclusion based on response magnitude, label, fit outcome or score is permitted. A kinetic analysis group contains all usable curves sharing the exact filename-derived sample and replicate. Its group ID is the compact UTF-8 JSON array [sample,replicate-as-a-string], and its trace ID is the exact ZIP member path prefixed with the input ZIP filename and ::. Sorting IDs always means Unicode code-point order. Diagnostic figures are required for every group; groups without any usable curve remain in the coverage ledger.

For every sample/replicate with usable curves, fit concentration-linked response models to measured t,y, preserving source time coordinates, offsets and concentrations. Compare a baseline/drift model with a 1:1 association/dissociation model wherever assay phase and concentration interpretation can be defended from the listed evidence. Report all attempted groups, including failures; no exclusion may depend only on a published binding label or published KD.

The source README describes corrected BLI traces and kinetic versus equilibrium estimation, but it does not establish every per-file phase boundary or unit. Record what is explicit, what is inferred from trace structure and what remains unknown. If a physical kinetic fit depends on an inferred boundary or offset, label its parameters conditional and evaluate three fixed analytical timing scenarios. Let T be the shortest usable trace duration in the group and each trace use elapsed time from its first finite timestamp. The phase boundaries are T/4, T/2 and 3T/4, named early, middle and late; each model estimates a separate additive offset for every trace. For each trace define elapsed time u = source t minus its first finite source t. Association uses 0 <= u <= the chosen boundary, and dissociation uses u > that boundary through that trace's last finite elapsed time. An observation exactly on the boundary belongs only to association; the model prediction must be continuous there, with the final association prediction supplying the initial dissociation response. Do not crop longer traces at T: T sets the three group boundaries only. Record source timestamps and elapsed timestamps separately, with no silent time rescaling or moving the observed phase data to fit a preferred outcome. These deliberately broad recorded-time scenarios test sensitivity and are not claimed instrument phase annotations. Any additional source-supported phase estimate is reported separately and cannot replace these three scenarios. When phase assignment cannot be defended, fit descriptive response models in the recorded coordinates and quantify the phase ambiguity; do not report physical kon, koff or KD. A missing phase annotation does not excuse ignoring the observed concentration response, repeatability or residual structure.

Write model_results.csv with the model equation reference, sample/replicate, contributing trace IDs, fitting domain, point count, fitted parameters and units, objective value, residual summaries, convergence/boundary diagnostics, and status. Write model_comparison.csv with baseline-versus-response comparison and a held-out-concentration evaluation whenever at least three distinct nonzero concentrations are available. Hold out each concentration once, retain whole traces, refit on the others and report prediction error separately from in-sample error. With fewer concentrations, record why that comparison cannot be made.

Quantify parameter identifiability using a profile of the fitting objective or a parameter-resampling analysis whose treatment of time dependence is justified. The Solver chooses and justifies the model family and identifiability method; there is no hidden required equation or target parameter value. Before execution, analysis_plan.json must specify each equation and parameter meaning, numerical domain or transform, objective and weighting, initialisation/convergence rule, and the finite exploration schedule actually executed. For profiling, list the finite parameter grid and the nuisance-parameter re-optimisation rule; for resampling, specify the dependence-preserving resampling unit and algorithm and use exactly 100 resamples. Include the observed coordinate scales used to justify numerical ranges and a recorded sensitivity check to imposed range limits. All fits remain subject to the existing start/evaluation/resource ceilings. A finite explored range does not establish an unbounded physical interval, and exhausting it cannot establish scientific non-identifiability. The Guardian checks that equations are defined on the claimed coordinates, source assumptions and units are explicit, the declared fits and exploration reproduce, and conclusion strength follows the reported evidence; the Guardian does not require a particular scientifically defensible model family. A Hessian-derived standard error alone is insufficient. Show whether an estimate is constrained by measured information or only by the numerical parameter bounds. Profile ranges and finite resampling counts must be recorded. Biological uncertainty cannot be inferred from repeated time samples. Physical parameters unsupported by units, phases or identifiability are null with a reason; conditional model estimates belong in separately labeled columns and must retain their conditions.

Write replicate_comparison.csv comparing independently derived response descriptors and supported/conditional parameters across every available replicate of each sample. Assess leave-one-replicate-out sensitivity when two or more replicates exist and preserve the complete range rather than selecting the most favourable repeat. Compare your results with the published summary and replicate values only after the independent computations. Explain agreements and discrepancies without treating publisher values as ground truth or treating your model as the original instrument analysis.

Analyse neutralisation and its relation to binding

Read all 57 neutralisation traces and the complete reading_data.csv metadata, including 53 candidate records and four control records. Preserve its sample, replicate, concentration, run_id and loading values. Analyse the observed response separately from the kinetic experiment. Calculate candidate-versus-control contrasts using both an endpoint-window response and the time-normalised area on a common observed time interval; define the interval, baseline treatment and interpolation explicitly. For each candidate, report contrasts using each of the four controls separately and using their arithmetic mean as the pooled centre. All four controls must yield finite values on the same interval for the pooled contrast; otherwise that pooled value is null with each missing/invalid control identified, while individually computable contrasts remain. Do not silently pool fewer controls. A ratio is null if its denominator is zero or the required baseline interpretation is unsupported; retain the underlying numerical responses and differences.

Show sensitivity to observed loading by reporting the contrasts before any loading correction and under a clearly labelled loading-normalised analysis, stating that normalisation is an analytical assumption rather than a proven correction. Do not clip inconvenient negative contrasts or values beyond an expected coefficient range. Distinguish an exploratory response contrast from the publisher's neutralisation coefficient unless the exact source definition is recoverable. There is one recorded candidate replicate and one recorded concentration in this release: do not claim replicate confidence intervals, dose-response potency, clinical activity or general biological neutralisation from it.

Write neutralisation_analysis.csv and binding_function_comparison.csv. Join the parsed neutralisation filename sample to reading_data.csv sample and replicate by exact string equality (normalise a numeric replicate only to its integer string when the source value is integral). Join that sample to kinetic sample, summary name, replicate-summary name and similarity id by exact sample-name equality. Preserve all four metadata control rows as controls, never candidate pairs. List missing, duplicate and otherwise ambiguous joins with their source identities; do not fuzzy-match or resolve collisions by row position. Each correlation input contains at most one binding value and one neutralisation value per candidate per declared scenario. State the deterministic across-replicate aggregation and finite replicate-sensitivity scenarios in analysis_plan.json, retaining the underlying replicate values and leave-one-out results. Match candidates to the independent binding results, report unmatched and unsupported cases, compare their association using Spearman rank correlation (Pearson correlation of average ranks) and exactly 100 candidate-level bootstrap resamples with replacement, each drawing N complete paired candidate records from the N usable pairs, with NumPy default_rng(20260909). Preserve all conditions for a drawn candidate. Report the original coefficient, all resampled coefficients, the count of finite coefficients, and a two-sided descriptive 95% percentile interval using NumPy quantile at 0.025 and 0.975 with method="linear" over finite resampled coefficients. If fewer than two finite resampled coefficients exist, the interval is null with the count/reason. A coefficient is null for a constant ranked variable, and repeat across the control/loading and kinetic-identifiability scenarios above. Tied values receive average ranks. For fewer than three paired finite observations, correlation is null with a denominator explanation. These intervals describe sensitivity within this selected historical set, not a random population sample. No ranking is required; any optional ordering must allow scientifically unresolved ties and use sample name only as a display tie-break.

Write scenario_inventory.csv enumerating the full Cartesian product of the three timing scenarios (early/middle/late), two neutralisation measures (endpoint/area), five controls (each original control plus pooled arithmetic mean), two loading treatments (uncorrected/normalised), and every declared binding descriptor or supported/conditional parameter and replicate-aggregation/sensitivity scenario. Include at least one independently measured binding-response descriptor across the complete available candidate set; unsupported physical parameters remain separate reason-coded null rows and cannot replace that descriptor analysis. Each inventory row links to its candidate pairs, finite pair count, result rows and null reason where applicable. Every combination must be computed or explicitly reason-coded; equivalent results caused by algebraic invariance are retained and explained. This matrix fixes analysis coverage, not a required scientific outcome. Provide figures covering collection-wide missingness, concentration/model residuals, replicate disagreement and neutralisation/control sensitivity. Supply a per-group diagnostic figure for each fitted kinetic group, so individual failures can be inspected without selecting only attractive curves.

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.

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 the source datasets in the submitted archive. Both roles independently fetch the appendix inputs before execution, verify SHA-256, and mount them read-only under /inputs with the appendix filenames, preserving raw/ directories. Source ZIP files remain ZIP files; the analysis must read them without extracting outside its scratch directory. Python files run only in a fresh isolated Linux sandbox, without network, credentials, host directories or a container-control socket. Required runtime: 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. No further package installation, paid service, language-model call or GPU is required or allowed during reproduction.

From the extracted submission root run python run.py --inputs /inputs --output /output --seed 20260909. The sandbox has four CPU cores, 8 GiB RAM and 10 GiB writable scratch space. One complete reproduction must finish within two hours and write all required outputs to a previously empty /output. 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. Use at least three dispersed initialisations for each nonlinear fit, unless an 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 sandbox 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.

A Submission passes only if all of the following hold:

1. Input coverage and provenance: the full appendix is verified, every original candidate has a coverage row, every released raw trace has a trace row; non-curve experimental evaluations have inventory rows and do not require fabricated trace mappings, duplicate/unmapped records remain visible, and candidate linkage can be followed to the original source. A claim that a field is absent must be checked against the complete relevant source, not a selected example.
2. New computation: the complete numerical analyses required below are executed from the measurement arrays. Copied publisher tables, precomputed fitted arrays, a label filter, a narrative alone or a ranking of published KD values cannot satisfy the task.
3. Statistical validity within the stated claim: model formulas, fitting objectives, constraints, uncertainty methods, data reuse and exclusion reasons are explicit. Correlated time points are not counted as independent experimental replicates. Distinct assays and conditions are not pooled without an explicit justified comparison. Missingness, nonconvergence, boundary estimates and non-identifiability propagate to downstream conclusions.
4. Evidence-limited interpretation: units and phases are supported by the listed source or explicitly conditional. An unsupported physical quantity is null with a specific reason and supporting diagnostic, while computable descriptive statistics remain numerical. A blanket declaration that all data are unusable does not replace the required observable curve analysis. No universal confidence, binding or quality threshold may be invented and described as a laboratory rule.
5. Full reproduction: the prescribed command finishes within bounds and regenerates the required outputs within the numerical tolerance. The Guardian independently implements the elementary trace diagnostics for the first 12 trace IDs in ascending Unicode code-point order, or all traces if fewer than 12 exist, and checks them against the submission. The Guardian also traces source-derived analysis groups, ordered by the group IDs defined below: select the first, lower-middle and last group, with zero-based indices 0, floor((N-1)/2) and N-1; remove duplicate indices if fewer than three groups exist. Inspect every required scenario for each selected group from raw response through independent fitted prediction and residual calculation; when it has no physical estimate, inspect its conditional/descriptive model and the evidence for that limitation. With no usable group, verify the complete source-derived unusability ledger rather than inventing a model result. This audit supplements the full reproduction.
6. Complete scientific conclusion: the report answers the stated question using collection-wide results and uncertainty, includes the required sensitivity analyses, distinguishes disagreements from missing measurements, and explains what the available measurements cannot establish. The Guardian checks each claim against its cited generated table or figure; no external search or unpublished judgment rule is required.

All criteria are pass/fail. This bounty does not rank scientific truth by an invented numerical merit score. Among Submissions that pass every criterion, choose the numerically smallest on-chain Solver address, interpreted as an unsigned 160-bit hexadecimal integer. Only active Submissions count; Elgora permits one active Submission per Solver address and a replacement supersedes the earlier one. This identifies one winning active Submission. When every required artifact and input is available and no Submission passes, use no_valid_submission; zero active Submissions also uses that outcome. 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 access beyond the sandbox 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.

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.

For the GitHub contents-API location of egfr_competition_2_replicate_summary.csv, retrieve the JSON response, base64-decode its content field, and verify SHA-256 of the decoded CSV bytes. The API response envelope itself is not the input. Other listed locations return the file bytes directly. egfr_rules.md is the fixed README snapshot, including its experimental-analysis description; links within it are contextual references and do not add unlisted required inputs.