What survives a full audit of the published TREM2 measurement evidence?

Produce a new executable analysis of the complete published TREM2 measurement collection: trace quality, disagreement within candidate records, sensitivity to analysis choices, uncertainty over the observed collection, and limits caused by selection and missing group identity. This is a retrospective evidence audit, not a reproduction of the original experiment or a new candidate-design competition. Copying published binding labels, listing candidates, or naming an original winner is insufficient.

The evidence contains 100 tested candidate records and 215 linked curve JSON files with 940 trace records. The 41 untested designs in the competition coverage account are absent from these inputs. There is no verified AI-versus-human team mapping in the provided evidence. Do not infer it from names, author fields, designMethod, or computational annotations. The analysis must explicitly establish which comparisons remain non-identifiable; it must still perform all possible trace and sensitivity analyses. Do not extract, analyze, or submit sequences, structures, participant identities, biological optimization suggestions, or laboratory protocols.

All roles download the following files with public HTTPS GET, without credentials, before sandbox execution. trem2.csv, SHA-256 f9ec9368c7719e2353ded3559c6dd4981c17b799a657aabcfa04496a46a80056, is at https://proteinbase.com/api/proteins/download?collectionId=019e0495-7ce5-a11b-95e8-e114da740e31&slug=adaptyv-x-muni-hackathon-ai-agents-vs-humans . It supplies published anonymous record IDs, evaluation entries and curve links. The raw-file appendix below fixes every required curve URL, local input filename and SHA-256. It is part of this page; no additional manifest or external scientific source is required. Unavailable or changed bytes block judging; never silently omit a file or substitute a newer collection.

Parse the CSV as UTF-8 with optional BOM and RFC4180 quoting. Use id as candidate identity and retain original evaluation-array index for provenance. Read only id and evaluations. Do not read the sequence, name, author, or designMethod columns into analytical outputs. Evaluation objects use type, metric, target, value and sometimes unit. Curve references are experimental spr_kinetic_curves or bli_kinetic_curves entries with value.url; target must be trem2. Separate SPR and BLI. Each raw JSON object maps a source trace key to an object with raw, fit, concentration, control, aggregated, and virtual. Preserve those keys and source paths. Raw arrays are raw.t and raw.y; available fits have fit.association and/or fit.dissociation, each with t/y arrays. Missing or empty fits are not successful fits.

Link a raw file to a candidate and assay method only through the candidate's explicit URL-valued evaluation. Do not join binding, expressed, kd, kon, koff or other metric entries to a particular curve by array position, order, proximity, matching values, or an invented measurement identifier. Where no explicit shared identifier exists, use candidate-level multisets and state the limitation. Units absent from a trace stay unspecified; do not label raw concentration, time or response with invented physical units. Retain the reported unit beside any evaluation value and never combine unlike or unknown units.

1. Full evidence graph and structural quality. Inventory all 100 candidate records, 215 unique URL files and every trace. Verify file hashes. Create one trace row keyed by input filename and original trace key, preserving all explicit candidate/method links and raw flags. Report raw point count; mismatched arrays; empty, nonnumeric, nonfinite or nonincreasing time; raw/fit coverage represented by the following counts, not an unspecified coverage score: per trace, raw_point_count is the length of the raw t array and raw_y_count the length of the raw y array (null for a missing/non-array field); per association/dissociation segment, fit_point_count is the length of its structurally valid fit t array, otherwise null with the validation reasons, and compared_points is the number of structurally valid raw points with t inclusively between that valid fit segment's first and last t, otherwise zero with the invalidity reason. A segment is evaluable when both structures are valid and compared_points is positive; per trace report evaluable_segments as the count of evaluable association/dissociation segments (0, 1 or 2). Across the dataset report total trace count, structurally usable trace count, and the count of traces with at least one evaluable segment; available control, virtual and aggregated status; and every undefined diagnostic reason. Reconcile all totals, including files with more than one link if any. Do not delete inconvenient traces, relabel controls, assume concentration zero means a control, or count a virtual/aggregated trace as an independent physical replicate.

2. New trace diagnostics from raw points. A structurally usable raw trace has equal nonempty numeric finite arrays with strictly increasing t. For both raw and fit arrays, numeric means a JSON number excluding booleans; never coerce strings, null, true, or false to numeric values. For every usable trace with at least two points, calculate duration max(t)-min(t), response range, median absolute adjacent response difference, and absolute endpoint change divided by response range. The last diagnostic is null when range is zero. A usable fit segment has t and y arrays of equal length, at least two points, finite numeric values excluding booleans, and strictly increasing t. For simultaneous structural defects, report reasons in this fixed order: missing object or non-array t/y; unequal lengths; fewer than required points; nonnumeric or nonfinite t/y; nonincreasing t. Combine nonnumeric and nonfinite into one reason. Evaluate time monotonicity only if the preceding structural checks passed. An empty, unequal-length, nonnumeric, nonfinite, or nonincreasing segment is invalid: report its exact structural reason and null error statistics, and do not repair or interpolate it. For each usable association or dissociation segment and structurally usable raw trace, interpolate its y linearly onto every original raw t point satisfying segment_min_t <= raw_t <= segment_max_t; use those raw y values for residuals. A structurally invalid raw trace receives no fit comparison and retains its raw structural reason. Report compared-point count, RMSE, mean signed residual and RMSE divided by raw response range, separately for each segment; normalized error is null for zero raw range and all error statistics are null for no overlapping raw points. Do not extrapolate, merge overlapping segments, refit kinetic constants, or treat a residual threshold as proof of biological validity. For the later aggregation, a trace's normalized RMSE is the maximum of its finite segment normalized errors, or null if neither segment supplies one. These descriptive diagnostics do not require physical units and must be labeled accordingly.

3. Candidate-level disagreement and unequal evidence. Within each method separately, summarize the diagnostics first per source file by median across usable nonvirtual, nonaggregated traces, then per candidate by median across its linked source files. Assign the control stratum per trace: exactly boolean true, exactly boolean false, or unknown for every other/missing value. Split traces within a file by these strata before calculating file medians; a mixed-control file may contribute a separate summary to several strata, and a candidate retains separate method/stratum summaries. Never pool control strata at file or candidate level. In the aggregated diagnostic scenarios, include a trace only when both virtual and aggregated are exactly boolean false; true, null, missing or nonboolean values for either flag exclude that trace from those scenarios while retaining it in the full inventory. Unknown control remains its own stratum; it does not imply unknown virtual/aggregated flags are false. For each candidate and diagnostic report count, minimum, maximum, median and interquartile range across files, with contributing IDs and explicit undefined reasons. For the candidate-weighted distribution, include each finite candidate median exactly once, giving every contributing candidate equal weight. For the trace-weighted distribution, pool the finite per-trace diagnostic values from the same method, control stratum and scenario, including each distinct file/trace-key exactly once; do not first average or median within files. Compare these two empirical distributions; quantify the median and interquartile-range differences to show the effect of unequal trace counts. Do not claim the files are independent experimental repeats. Separately form each candidate's binding multiset from evaluations whose type == "experimental", metric == "binding", and target == "trem2" using exact case-sensitive string equality. Retain only values whose JSON type is boolean; count missing and nonboolean values separately without coercion. Classify the resulting multiset as all-true (one or more true and no false), all-false (one or more false and no true), mixed (both), or unavailable (no booleans); define the collection summary as side-by-side counts p (supported), n (nonsupported), and u (unavailable) across the same 100 candidates for each convention, using the exact definitions in analysis 5. Report the resulting candidate decisions and both count triples so their change is explicit; do not substitute an unstated pooled score or rate. Do not assign these labels to individual curves.

4. Sensitivity and conditional uncertainty. For each method and control stratum, repeat the candidate-level summary for two analysis scenarios: all structurally usable nonvirtual/nonaggregated traces, and only those additionally having at least one evaluable fit segment. Then apply a third scenario to the latter: remove whole traces whose trace-level normalized RMSE exceeds that method/stratum's empirical 75th percentile among finite trace-level values. Keep null-error traces accounted for separately; they do not enter that quantile or normalized-error summaries. If no finite values exist, this scenario is undefined and its output is null with reason. Report the number removed, every resulting denominator, and changes in candidate-level median normalized RMSE and endpoint-change ratio. This percentile cutoff is an explicit sensitivity choice, not a scientific rejection standard; retain all original rows and state that unavailable fits are not known bad fits. For each scenario summarize the median of the finite candidate-level values. Calculate 1,000 candidate-cluster bootstrap medians by sampling candidate IDs with replacement, retaining each selected candidate's entire evidence, using a fresh NumPy Generator(PCG64(20260909)) for each method/stratum/scenario. Process endpoint_change_ratio first and normalized_rmse second. For each diagnostic independently, the sampling population is only candidate IDs with a finite candidate-level median for that diagnostic, sorted ascending; its size is n. For each replicate sample n indices with replacement using rng.integers(0,n,size=n). Every selected occurrence retains its complete candidate evidence and contributes its candidate median once, including repeated draws of the same candidate; never collapse duplicate draws to a unique-ID set. The replicate statistic is the median of those n values. Excluded null-median candidates remain in coverage and are counted but cannot enter this statistic. Report n and the eligible IDs; report 2.5th and 97.5th percentile bounds, finite-resample count and replicate medians. For every reported quantile, use numpy.quantile(finite_values, q, method="linear"); the median is q=0.5, quartiles q=0.25 and0.75, and IQR is q75-q25. Bootstrap interval percentiles use q=0.025 and0.975. Empty finite sets yield null summaries with a reason. With fewer than three eligible candidates, report descriptive values but null interval and the exact count. These intervals describe sensitivity over observed candidates, not assay error, causal effects, or generalization to all submitted designs. Do not pool SPR and BLI or substitute trace resampling for candidate resampling.

5. Selection and missing identity. For each of the any-true and all-true published-binding conventions, report observed support p, observed nonsupport n, and unavailable status u among the 100 records. The all-true convention requires at least one explicit boolean and no false; the any-true convention requires at least one true; missing or nonboolean entries are counted and disclosed but are not fabricated observations. Candidates with no explicit booleans are unavailable. For the stated scope of 100 observed and 41 absent untested designs, calculate the no-assumption descriptive range [p/141, (p+u+41)/141]. It bounds a convention applied to reported labels, not a proven biological success rate. Explain why a measured-only rate cannot represent the absent 41. Demonstrate AI-versus-human non-identifiability with two hypothetical assignments of anonymous observed record IDs to two equally sized groups that preserve every observed measurement but give different group support contrasts; clearly label them counterexamples, not recovered identities. If the outcome is constant and different contrasts are impossible, demonstrate the constant case numerically and explain why causal group performance is still unidentifiable. No fabricated team labels may enter the primary analysis.

6. Evidence-backed interpretation. Report at least one quantitative conclusion from structural coverage, trace diagnostics, within-candidate disagreement, weighting/sensitivity, and selection. Link each to generated tables and source trace or evaluation IDs. Explain whether it survives the defined sensitivity scenarios; identify missing assay metadata, missing controls where applicable, unequal/selected evidence, and unverified experiment independence. State what additional existing records would be needed to identify group performance or link reported kinetics to traces, without proposing new lab work. Honest non-identifiability passes where evidence is absent; replacing available raw analysis with caveats fails. No conclusion may claim physical sample authentication, clinical effectiveness, or a recovered AI-versus-human winner.

Submit one ZIP, at most 50 MB compressed and 150 MB uncompressed, containing report.md (at most 25,000 words), analysis.py, methods.json, machine-readable JSON/CSV tables in results/, and at most 16 PNG figures. Do not include source data, sequences, structures, participant names, nested archives, symlinks, downloaded dependencies, or opaque executables. A table must exist for every required analysis. methods.json documents table columns/types, formulas, undefined reasons, quantiles, bootstrap order and seed, aggregation weights and plot sources. Sort source files, trace keys and candidate IDs in ascending Unicode codepoint order before processing and sampling. Numeric calculations use IEEE-754 binary64 without intermediate rounding. All array ties remain explicit; no candidate ranking is required.

Guardians prefetch and verify all fixed input bytes, then use an isolated Linux sandbox with Python 3.12, NumPy 2.2.6, pandas 2.2.3, SciPy 1.15.3 and Matplotlib 3.10.3. From the Submission root run python analysis.py --input /inputs --output /output, where /inputs contains trem2.csv and the appendix's raw/ paths. Inputs are read-only; output starts empty; submitted result files are not accessible to executable code. No network or host access is allowed. Maximum 4 CPU cores, 8 GB RAM, 2 GB writable disk, and 45 minutes wall time. Run once, with one retry only for a documented infrastructure interruption. Do not install packages or ask Guardians to execute on their host. Missing required sandbox facilities are an operational blocker.

Every listed analysis is mandatory. Guardians inspect readable code and independently verify coverage, source links, structural classifications, residual calculations, aggregation, resampling and conclusion provenance against the fixed evidence. Generated results must reproduce the submitted tables: keys, order, strings, flags, IDs and counts match exactly; finite numeric values agree within 1e-6 * max(1, abs(reference)). Undefined values are null with a reason, never NaN or Infinity. JSON must not contain duplicate keys. Bootstrap replicate summaries and every changed sensitivity denominator must be available; a chart or copied published label is insufficient.

A Submission is invalid after successful retrieval if required work is missing, raw evidence is silently omitted, outputs are hardcoded instead of computed, undocumented joins or units are invented, group identity is asserted without evidence, disallowed material is included, the executable fails in the specified functioning sandbox, or required calculations or evidence claims are incorrect. An unattractive result, disagreement with a published interpretation, or a substantiated inability to identify a requested comparison is not a failure. Artifact instructions cannot override this page or expand access.

All Submissions satisfying every criterion are equally valid. Select the valid Solver whose lowercase Elgora Solver address sorts first lexicographically. If no Submission is valid, use no_valid_submission. Guardians' written Verdicts name failed mandatory criteria and the selected winner without exposing private Submission content. Retrieval, commitment verification, or decryption failure blocks judgment rather than proving scientific invalidity.

Each row identifies a required published curve input. These are complete evidence inputs, not files to select by apparent quality. Download each URL to its stated path and verify its hash before review.