Celeus

Celeus engine validation report

category: validation audience: public status: current owner: TBD (owner)

This report is generated, not written by hand, and is byte-for-byte reproducible on the pinned engine (R 4.6.1 + a pinned dependency lockfile): it carries no timestamp, so re-running it and finding no diff is itself a demonstration of the determinism this product promises. Every dataset here is public and citable; no patient data is used anywhere.

What this is: evidence that the engine computes the right numbers and steers to a defendable test, on well-known public datasets, checked two independent ways. What this is NOT: a substitute for review by an independent statistician (the named next step, below), and NOT clinical decision support - Celeus is a research-statistics tool.

How to read the three tiers

Each result is checked at up to three levels of rigor:

  1. Canonical agreement (all cases). The engine’s number equals an INDEPENDENT computation in base R / survival / lme4 / pROC, to a tolerance of 1e-06. This catches the class of bug that actually occurs - spec routing, coercion, NA handling, serialization - the statistics themselves are R’s.
  2. Application-path agreement. The same equality through the full upload -> analyse -> report path, CI-gated in our automated test suite (not re-run here; asserted on every push).
  3. Literature anchor (a subset). The engine’s number also equals a value published in a citable source (a paper, textbook, or the dataset’s own package vignette) - a checksum a script cannot accidentally regenerate wrong. Only some cases have a single citable published number; the rest are honestly labelled canonical-only, which is a real check, just not a literature one.

Reproducibility stamp

field value
R version R version 4.6.1
pinned dependency lockfile md5 dc02556bdb795e33269d37da113d908a
catalog version c32-a6f7
engine version 0.1.0

Summary

  • 40 cases across 32 catalog test types, 68 numeric comparisons total.
  • 10 cases are additionally anchored to a published literature value; 30 are canonical-only (labelled per-case below).
  • 0 unexplained divergences (a non-zero count here fails the generator).
  • Test selection: the correct test family was proposed in 34 of 34 evaluated cases (directly, or via a documented post-analysis follow-up - see below); the grader recommended a sound alternative in 12 of 34; the test actually used was itself graded “Not appropriate” in 0 of 34. (6 of 40 total cases skipped this evaluation - external-downloaded, fixture not present; see “Dataset sources” above.)
  • Reproducibility is separately CI-gated, two ways. Over a 34-case builtin corpus: each analysis re-runs its emitted reproduce.R standalone, and re-analysing the package’s own bundled dataset must return every recorded number bit-for-bit. Across processes: the same analysis run in two independent operating-system processes must emit a byte-identical package - every file compared by hash, with nothing excluded. That proves reproducibility, not correctness - correctness is the tables below.

Dataset sources

Every case above draws on one of three source types:

source type cases
bundled package (a CRAN/base-R package, ships with the pinned dependency set - no network) 32
synthetic/illustrative (a documented worked example, NOT real patient data - honestly labelled) 2
externally downloaded (an opt-in network fetch, run weekly on a schedule - not part of this report’s own reproducibility check) 6

External-sourced cases not run in THIS report

This report is generated network-free, to keep it byte-reproducible (see the reproducibility stamp above) - it never downloads external datasets. Those cases run on a separate weekly schedule instead:

case dataset method
heart-logistic_regression UCI Heart Disease (Cleveland) logistic_regression
cervical-hinselmann-diagnostic UCI Cervical Cancer (Hinselmann vs Biopsy) diagnostic_accuracy
cervical-schiller-diagnostic UCI Cervical Cancer (Schiller vs Biopsy) diagnostic_accuracy
cervical-citology-diagnostic UCI Cervical Cancer (Citology vs Biopsy) diagnostic_accuracy
cervical-hinselmann-vs-schiller-mcnemar UCI Cervical Cancer (Hinselmann vs Schiller, two index tests on the same patients) mcnemar
cervical-schiller-vs-citology-mcnemar UCI Cervical Cancer (Schiller vs Citology, two index tests on the same patients) mcnemar

Numeric agreement (all cases vs an independent computation)

case dataset (field) Celeus independent R match
sleep-paired_t datasets::sleep (p_value) 0.00283289 0.00283289 yes
sleep-paired_t datasets::sleep (estimate) -1.58 -1.58 yes
sleep-wilcoxon datasets::sleep (p_value) 0.00390625 0.00390625 yes
ToothGrowth-welch_t datasets::ToothGrowth (p_value) 0.0606345 0.0606345 yes
ToothGrowth-kruskal_wallis datasets::ToothGrowth (p_value) 1.47521e-09 1.47521e-09 yes
iris-welch_anova datasets::iris (p_value) 1.50506e-28 1.50506e-28 yes
mtcars-pearson datasets::mtcars (p_value) 1.29396e-10 1.29396e-10 yes
mtcars-pearson datasets::mtcars (estimate) -0.867659 -0.867659 yes
mtcars-spearman datasets::mtcars (p_value) 1.48759e-11 1.48759e-11 yes
mtcars-linear_regression datasets::mtcars (estimate) -3.87783 -3.87783 yes
mtcars-linear_regression datasets::mtcars (p_value) 1.11965e-06 1.11965e-06 yes
mtcars-logistic_regression datasets::mtcars (estimate) -8.08348 -8.08348 yes
mtcars-logistic_regression datasets::mtcars (p_value) 0.00843381 0.00843381 yes
mtcars-fisher_exact datasets::mtcars (p_value) 0.472697 0.472697 yes
titanic-chi_square carData::TitanicSurvival (p_value) 4.58992e-81 4.58992e-81 yes
birthwt-logistic_regression MASS::birthwt (estimate) -0.0389946 -0.0389946 yes
birthwt-logistic_regression MASS::birthwt (p_value) 0.23344 0.23344 yes
sleepstudy-mixed_model lme4::sleepstudy (estimate) 10.4673 10.4673 yes
airquality-mann_whitney-filtered datasets::airquality (p_value) 6.10874e-05 6.10874e-05 yes
lung-kaplan_meier survival::lung (status recoded 0/1) (p_value) 0.00131116 0.00131116 yes
lung-cox_ph survival::lung (status recoded 0/1) (estimate) 1.01719 1.01719 yes
lung-cox_ph survival::lung (status recoded 0/1) (p_value) 0.064591 0.064591 yes
lung-aft_weibull survival::lung (status recoded 0/1) (estimate) 0.987818 0.987818 yes
lung-aft_weibull survival::lung (status recoded 0/1) (p_value) 0.0781189 0.0781189 yes
mcnemar-disease-example synthetic fixture from the McNemar’s test Wikipedia worked example (statistic) 20.6722 20.6722 yes
mcnemar-disease-example synthetic fixture from the McNemar’s test Wikipedia worked example (p_value) 5.45009e-06 5.45009e-06 yes
respiratory-mcnemar HSAUR::respiratory (baseline vs month-4 status) (statistic) 1.93939 1.93939 yes
respiratory-mcnemar HSAUR::respiratory (baseline vs month-4 status) (p_value) 0.163734 0.163734 yes
anorexia-ancova MASS::anorexia (statistic) 7.86808 7.86808 yes
anorexia-ancova MASS::anorexia (p_value) 0.00084384 0.00084384 yes
warpbreaks-poisson_regression datasets::warpbreaks (estimate) -0.205988 -0.205988 yes
warpbreaks-poisson_regression datasets::warpbreaks (p_value) 6.48978e-05 6.48978e-05 yes
quine-negbin_regression MASS::quine (estimate) -0.569372 -0.569372 yes
quine-negbin_regression MASS::quine (p_value) 0.00020458 0.00020458 yes
housing-ordinal_logistic MASS::housing (expanded one row per Freq) (estimate) 0.566394 0.566394 yes
housing-ordinal_logistic MASS::housing (expanded one row per Freq) (p_value) 6.22819e-08 6.22819e-08 yes
housing-ordinal_logistic MASS::housing (expanded one row per Freq) (residual_deviance) 3479.15 3479.15 yes
warpbreaks-two_way_anova datasets::warpbreaks (statistic) 4.18907 4.18907 yes
warpbreaks-two_way_anova datasets::warpbreaks (p_value) 0.0210442 0.0210442 yes
iris-multinomial_logistic datasets::iris (estimate) -9.74099 -9.74099 yes
iris-multinomial_logistic datasets::iris (p_value) 0.411011 0.411011 yes
iris-multinomial_logistic datasets::iris (residual_deviance) 23.8612 23.8612 yes
PlantGrowth-one_way_anova datasets::PlantGrowth (statistic) 4.84609 4.84609 yes
PlantGrowth-one_way_anova datasets::PlantGrowth (p_value) 0.01591 0.01591 yes
ToothGrowth-student_t datasets::ToothGrowth (statistic) 1.91527 1.91527 yes
ToothGrowth-student_t datasets::ToothGrowth (p_value) 0.0603934 0.0603934 yes
ToothGrowth-student_t datasets::ToothGrowth (estimate) 3.7 3.7 yes
VerbAgg-mixed_logistic lme4::VerbAgg (estimate) matches refit (not frozen) (independent refit) yes
VerbAgg-mixed_logistic lme4::VerbAgg (p_value) matches refit (not frozen) (independent refit) yes
epil-mixed_poisson MASS::epil (estimate) matches refit (not frozen) (independent refit) yes
epil-mixed_poisson MASS::epil (p_value) matches refit (not frozen) (independent refit) yes
ergoStool-rm_anova nlme::ergoStool (statistic) 22.3556 22.3556 yes
ergoStool-rm_anova nlme::ergoStool (p_value) 3.00556e-05 3.00556e-05 yes
ergoStool-rm_anova nlme::ergoStool (estimate) 0.459375 0.459375 yes
aSAH-roc_auc pROC::aSAH (auc) 0.731369 0.731369 yes
aSAH-roc_compare pROC::aSAH (s100b vs ndka) (auc1) 0.731369 0.731369 yes
aSAH-roc_compare pROC::aSAH (s100b vs ndka) (auc2) 0.611958 0.611958 yes
aSAH-roc_compare pROC::aSAH (s100b vs ndka) (auc_diff) 0.119411 0.119411 yes
aSAH-roc_compare pROC::aSAH (s100b vs ndka) (roc_test_p) 0.164295 0.164295 yes
aSAH-roc_compare pROC::aSAH (s100b vs ndka) (roc_test_stat) 1.39077 1.39077 yes
screening-diagnostic_accuracy documented illustrative 2x2 screening fixture (sensitivity) 0.95 0.95 yes
screening-diagnostic_accuracy documented illustrative 2x2 screening fixture (specificity) 0.9 0.9 yes
screening-diagnostic_accuracy documented illustrative 2x2 screening fixture (ppv) 0.513514 0.513514 yes
screening-diagnostic_accuracy documented illustrative 2x2 screening fixture (npv) 0.993865 0.993865 yes
screening-diagnostic_accuracy documented illustrative 2x2 screening fixture (prevalence) 0.1 0.1 yes
davis-bland_altman carData::Davis (measured vs self-reported weight) (bias) 0.601093 0.601093 yes
davis-bland_altman carData::Davis (measured vs self-reported weight) (loa_lower) -15.969 -15.969 yes
davis-bland_altman carData::Davis (measured vs self-reported weight) (loa_upper) 17.1711 17.1711 yes

“matches refit (not frozen)” marks iterative-optimizer fits (glmer mixed models) whose engine value equals an independent refit on every run, but whose absolute value carries platform/optimizer jitter below the agreement tolerance. The verdict is stable and reported; the raw number is not printed because it is not byte-reproducible (which is exactly why these cases are canonical-only).

Literature-anchored cases

These cases additionally match a citable published value:

case field Celeus published value citation match
sleep-paired_t p_value 0.00283289 0.00283289 Student (1908), Biometrika 6(1):1-25 yes
mtcars-pearson estimate -0.867659 -0.867659 Henderson & Velleman (1981), Biometrics 37 yes
lung-kaplan_meier p_value 0.00131116 0.00131116 Loprinzi et al. (1994), J. Clinical Oncology 12(3) yes
mcnemar-disease-example statistic 20.6722 20.6722 Wikipedia contributors, “McNemar’s test” (worked example: n=314 patients, disease diagnosis before/after a drug), retrieved 2026-07-13 yes
mcnemar-disease-example p_value 5.45009e-06 5.45009e-06 Wikipedia contributors, “McNemar’s test” (worked example: n=314 patients, disease diagnosis before/after a drug), retrieved 2026-07-13 yes
PlantGrowth-one_way_anova statistic 4.84609 4.84609 Dobson (1983), An Introduction to Statistical Modelling - the textbook one-way ANOVA yes
PlantGrowth-one_way_anova p_value 0.01591 0.01591 Dobson (1983), An Introduction to Statistical Modelling - the textbook one-way ANOVA yes
ToothGrowth-student_t statistic 1.91527 1.91527 Crampton (1947), J. Nutrition 33(5) yes
ToothGrowth-student_t p_value 0.0603934 0.0603934 Crampton (1947), J. Nutrition 33(5) yes
ToothGrowth-student_t estimate 3.7 3.7 Crampton (1947), J. Nutrition 33(5) yes
ergoStool-rm_anova statistic 22.3556 22.3556 Pinheiro & Bates (2000), Mixed-Effects Models in S and S-PLUS yes
ergoStool-rm_anova p_value 3.00556e-05 3.00556e-05 Pinheiro & Bates (2000), Mixed-Effects Models in S and S-PLUS yes
ergoStool-rm_anova estimate 0.459375 0.459375 Pinheiro & Bates (2000), Mixed-Effects Models in S and S-PLUS yes
aSAH-roc_auc auc 0.731369 0.7314 Turck N et al. (2010), J. Proteome Research 9(11); the aSAH dataset shipped and validated by the pROC package (Robin et al. 2011, BMC Bioinformatics 12:77) yes
aSAH-roc_compare auc1 0.731369 0.7314 Turck N et al. (2010), J. Proteome Research 9(11); the two-marker DeLong comparison is pROC’s own roc.test worked example (Robin et al. 2011, BMC Bioinformatics 12:77) yes
aSAH-roc_compare auc2 0.611958 0.612 Turck N et al. (2010), J. Proteome Research 9(11); the two-marker DeLong comparison is pROC’s own roc.test worked example (Robin et al. 2011, BMC Bioinformatics 12:77) yes
aSAH-roc_compare roc_test_p 0.164295 0.164295 Turck N et al. (2010), J. Proteome Research 9(11); the two-marker DeLong comparison is pROC’s own roc.test worked example (Robin et al. 2011, BMC Bioinformatics 12:77) yes
screening-diagnostic_accuracy sensitivity 0.95 0.95 Illustrative screening-test 2x2 (TP=95, FP=90, FN=5, TN=810); textbook-style table, not real patient data yes
screening-diagnostic_accuracy specificity 0.9 0.9 Illustrative screening-test 2x2 (TP=95, FP=90, FN=5, TN=810); textbook-style table, not real patient data yes
screening-diagnostic_accuracy prevalence 0.1 0.1 Illustrative screening-test 2x2 (TP=95, FP=90, FN=5, TN=810); textbook-style table, not real patient data yes

Divergence detail

Per-comparison detail for every divergence counted above - case/field identify the comparison, reference names what it was checked against (independent_r, anchor, or an (engine_error)/(canonical_error) crash), and abs delta/tolerance are the exact numbers a regression’s own PR needs, not just a bare match verdict. Empty in the healthy case, by construction: the row list below is built from the SAME divergence events the summary count above totals, so it can never show a count without matching rows or vice versa.

  • (none)

Test selection (does the tool steer to the right test, and show alternatives?)

Correctness of the number matters only if the tool also proposes a defendable test. This tier uses the deterministic metadata advisor + suitability grader (no AI): it checks the correct test FAMILY is proposed for the data, and that the grader recommends a sound sibling and never calls the valid test ‘Not appropriate’. The alternatives offered column is the count of sibling tests the tool surfaces for the researcher to consider - the product deliberately shows the alternatives and lets the researcher (optionally with AI assistance) explore them, rather than dictating one answer.

case method correct family proposed grader recommends alternatives offered
sleep-paired_t paired_t yes - 2
sleep-wilcoxon wilcoxon_signed_rank yes - 2
ToothGrowth-welch_t welch_t yes mann_whitney 3
ToothGrowth-kruskal_wallis kruskal_wallis yes welch_anova 3
iris-welch_anova welch_anova yes welch_anova 3
mtcars-pearson pearson yes pearson 2
mtcars-spearman spearman yes pearson 2
mtcars-linear_regression linear_regression yes - 1
mtcars-logistic_regression logistic_regression yes - 1
mtcars-fisher_exact fisher_exact yes chi_square 2
titanic-chi_square chi_square yes chi_square 2
birthwt-logistic_regression logistic_regression yes - 1
sleepstudy-mixed_model mixed_model yes - 1
airquality-mann_whitney-filtered mann_whitney yes mann_whitney 3
lung-kaplan_meier kaplan_meier yes - 1
lung-cox_ph cox_ph yes - 2
lung-aft_weibull aft_weibull yes - 2
mcnemar-disease-example mcnemar yes - 1
respiratory-mcnemar mcnemar yes - 1
anorexia-ancova ancova yes - 1
warpbreaks-poisson_regression poisson_regression yes negbin_regression 2
quine-negbin_regression negbin_regression yes negbin_regression 2
housing-ordinal_logistic ordinal_logistic yes - 1
warpbreaks-two_way_anova two_way_anova yes - 1
iris-multinomial_logistic multinomial_logistic yes - 1
PlantGrowth-one_way_anova one_way_anova yes welch_anova 3
ToothGrowth-student_t student_t yes mann_whitney 3
VerbAgg-mixed_logistic mixed_logistic yes - 1
epil-mixed_poisson mixed_poisson yes - 1
ergoStool-rm_anova rm_anova yes - 1
aSAH-roc_auc roc_auc yes - 1
aSAH-roc_compare roc_compare yes - 1
screening-diagnostic_accuracy diagnostic_accuracy yes - 1
davis-bland_altman bland_altman via post-analysis follow-up - 1

What this does NOT validate (read this)

  • 30 of 40 cases are canonical-only - they match an independent computation but no single citable published number. Presenting them as literature-validated would be dishonest; they are not.
  • For several cases (paired t, Welch t, the regressions, Kaplan-Meier, Cox, the mixed models) the selection check is a ‘not wrong’ bar - it asserts the grader does not reject the correct test - rather than ‘graded strictly best’. Textbook-unambiguous cases assert the stronger ‘recommended’.
  • One case’s test family is proposed only AFTER a related analysis actually runs, not before: a strongly-correlated, similar-scale pair run as a correlation gets a deterministic follow-up note recommending an agreement analysis instead - the advisor deliberately never guesses this pairing from column names alone (a false guess would steer a researcher to the wrong test), but once real data confirms the pattern, a conditional, honest suggestion is safe to make.
  • The three glmer mixed-model cases are canonical-only ON PURPOSE: a frozen published coefficient would risk cross-platform optimizer jitter, so same-run engine-vs-refit equality is the guard.
  • This is a self-generated report. It is real evidence, but weaker than an independent statistician’s review, which is the explicit next credibility step and is not yet done.
  • Celeus is research statistics, not clinical decision support: nothing here validates or implies patient-specific diagnosis or treatment.

Reproducibility guarantee

This report carries no timestamp: regenerating it against the same pinned environment (same code, same package versions, same corpus) produces a byte-identical file, character for character - so a clean regeneration finding zero diff is itself the determinism demonstration this artifact claims, not an assertion taken on faith.

Every individual analysis run on the platform emits its own seed, exact package versions, and a standalone reproduce script alongside its results - the same determinism guarantee this report demonstrates, applied per-run. This report’s own determinism rests on the pinned lockfile (md5 above) plus the fixed catalog version, not on a captured full session/locale snapshot (which would break byte-identity across otherwise-identical environments).

Next credibility step

Independent review by a qualified biostatistician of the method-selection logic and this corpus. A self-generated report cannot substitute for it; it is named here as the next step, deliberately out of scope for this artifact.