Russell Parrish · compiled in the August 2026 audit with Claude and Codex (OpenAI) · ledger state 2 September 2026 · page published November 2026
What this is. Every claim the pathology work has made or tested, as audited in August 2026: 123 claims, each with the proposition as originally stated, its current status, the evidence behind it, the boundary of what it shows, and the next test that would decide it. Retracted and unsupported claims are kept, not deleted: 9 are retracted and 51 are not supported. The full table can be downloaded as CSV. Source file paths are held back; the underlying data, code and results are available on request.
Since this ledger (November 2026)
External benchmark. The CRC rows describe within-dataset cross-validation (91.83%). On the benchmark's designed external split the same measurements score 83.7%, and a logged development record reaches 94.3% on held-back tiles with a fitted readout (technical note; lab books for round 1, 2, 3, 4 and 5).
Stain normaliser (INS-005, SITE-002). The mechanism of the degenerate historical normaliser, and the site difference it created, are written up (technical note).
TCGA slide scale. Twenty-five TCGA colorectal slides carry a misstated pixel spacing in a public index (technical note).
Transport (CAM, CPTAC). The Camelyon17 and CPTAC rows are summarised in one place (technical note).
Structural persistence (CPTAC-009 to 012, ID-014). The same-patient specimen result and the brain-tissue test are now on the persistence page (addendum).
Survival (SURV). The retracted survival figures (SURV-001, -016, -017) have been removed from the site. No new survival claim is made here.
Status, as used here
supported
17
Established within its stated boundary.
reproduced
9
Recomputed from preserved data or original images and matched.
internally supported
5
Holds within the dataset where it was found; not yet tested externally.
qualified
24
Holds only with the stated boundary or as a post-result sensitivity.
exploratory
2
A direction worth testing, not a finding.
open
6
Not yet decidable with current data or power.
not supported
51
Tested and not established (refuted, confounded or underpowered; the reason is recorded).
retracted
9
A previously stated claim withdrawn: not reproducible or shown to be wrong.
Domain
supported
reproduced
internally supported
qualified
exploratory
open
not supported
retracted
All
Instrument
4
2
9
14
2
31
CRC-VAL
2
1
5
8
EBHI-SEG
1
1
3
2
3
1
11
STHELAR
1
2
3
6
Structural address positioning
1
1
1
3
GTEx/PAXgene
1
3
1
1
6
TCGA survival
1
1
2
1
2
8
5
20
TCGA identity
4
1
1
1
3
10
Structural persistence
3
5
1
9
18
Site effects
1
1
1
1
4
TCGA slide relationships
1
1
1
3
6
The claims
Each line shows the claim as it now stands. Open a line for the original proposition, the evidence, the boundary and the next decisive test.
INS-001reproducedA named and hashed implementation deterministically maps a fixed image and configuration to explicit measurements.
Original proposition
The feature extractor is deterministic and reproducible under a fixed implementation.
Evidence tier
Technical
Key evidence
Same-image reruns and the local verification harness reproduce recorded outputs.
Boundary
Reproducible computation is not biological invariance or clinical validity.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
COMPUTATION OR RESULT REPRODUCED
INS-002not supportedThe measurement extractor is label-free; class centers, scaling, LDA, Ridge, Cox models, thresholds, and calibration are fitted or reference-dependent.
Original proposition
No trained parameters are used at any stage of the complete system.
Evidence tier
Construct
Key evidence
CRC and EBHI reconstructions identify supervised downstream readouts.
Boundary
Use 'deterministic extractor' rather than 'training-free classifier.'
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
Notes
Original global zero-training language.
INS-003qualifiedParallax is a versioned family of explicit structural measurements; every result must name the kernel, configuration, patch scale, and retained axes.
Original proposition
The 58-axis frame is one fixed universal object.
Evidence tier
Construct
Key evidence
Historical runs used different axis subsets and v3/v4.1 kernels; CRC recovered 81 usable fields despite a Full_58 label.
Boundary
Do not imply that every axis was tested in every corpus.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
INS-004not supportedAt least two different quantities shared the name Omega: mask-coherence disagreement and reference-relative deviation distance.
Original proposition
Omega is one universal structural-deviation scalar.
Evidence tier
Construct
Key evidence
Code and column audit separates the two constructions.
Boundary
Namespace the quantity and declare its reference, aggregation level, and implementation.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
Notes
Original single-Omega wording.
INS-005retractedThe historical implementation is mathematically degenerate; corrected normalization materially changes several axes.
Original proposition
The historical Macenko implementation is a valid stain normalizer.
Evidence tier
Technical
Key evidence
Recovered stain vectors were nearly collinear (about 0.17 degrees; condition number about 748).
Boundary
This does not invalidate CRC-VAL, whose distributed images were already normalized and whose notebook did not invoke this normalizer.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
RETRACTED OR INVALIDATED BY RECONSTRUCTION
INS-006reproducedThe audited local and published v4.1 functions match at the function level.
Original proposition
The local v4.1 kernel matches the published kernel.
Evidence tier
Technical
Key evidence
39/39 functions identical in the kernel audit.
Boundary
Documentation discrepancies remain for eligible masks, scaling, and run configuration.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
COMPUTATION OR RESULT REPRODUCED
INS-007not supportedMost mask centroids were never persisted, so some retrospective pairwise constructions cannot be recovered from the archives.
Original proposition
The archived tables retain the complete pairwise centroid geometry.
Evidence tier
Provenance
Key evidence
Only k_dx/dy and b_dx/dy reached disk; 14 of 16 centroids were absent.
Boundary
Use newly generated complete tables for any future centroid-profile analysis.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
INS-008retractedThe survival analysis ran at 512 pixels; geometry depends on patch scale.
Original proposition
The documented 224-pixel configuration describes the TCGA survival run.
Evidence tier
Provenance
Key evidence
Recovered TCGA batch configuration specifies 512.
Boundary
Patch scale must be treated as part of the instrument definition.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
RETRACTED OR INVALIDATED BY RECONSTRUCTION
INS-009qualifiedOutputs are traceable to named image measurements, but biological or diagnostic meaning requires validation.
Original proposition
Axis-level output is an explanation equivalent to a pathologist's reasoning.
Evidence tier
Interpretation
Key evidence
The computations are explicit; several proxies failed or changed meaning under controlled tests.
Boundary
Use 'measurement-level rationale' rather than equivalence to expert interpretation.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
INS-010qualifiedThe implementation excludes combined and edge masks, leaving 11 eligible masks rather than the 13 implied by the appendix.
Original proposition
Published Omega eligibility includes every non-combined mask.
Evidence tier
Specification
Key evidence
Code audit finds an average 9.72 valid masks among 11 eligible on TCGA.
Boundary
This is a documentation correction; it does not by itself invalidate computed archive values.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
INS-011qualifiedd_s uses an Euclidean distance transform with an undocumented 2/min(width,height) scaling.
Original proposition
The published d_s definition fully specifies the computation.
Evidence tier
Specification
Key evidence
Function-level code review against Appendix A.1.
Boundary
The quantity remains traceable, but scale and resize configuration must accompany it.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
INS-012qualifiedThe H-channel mask is degenerate on 64.1% of audited TCGA tiles, more often than L+M-2S at 38.8%.
Original proposition
H-channel segmentation is consistently informative in TCGA H&E.
Evidence tier
Measurement behavior
Key evidence
Degenerate-mask tally across 9,010 TCGA tiles.
Boundary
This is corpus- and configuration-specific; it warns against treating channel names as reliable biological compartments.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
INS-013not supportedThe surviving patient tables describe raw-like aggregation and do not reproduce the published normalized survival construction.
Original proposition
Archived patient-feature tables reproduce the survival analysis spine.
Evidence tier
Provenance
Key evidence
Archive reruns give approximately HR 1.33, a non-monotone tertile gradient, and KM p=0.16.
Boundary
The missing 182x120 table remains unrecovered; use tile-level rebuilds and named specifications.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
INS-014not supportedTile identity and row alignment between the historical raw and normalized archives are unrecoverable.
Original proposition
Historical raw and normalized tile archives can be paired row by row.
Evidence tier
Provenance
Key evidence
Own-row correlations are approximately shuffled correlations; the earlier size-based alignment check was vacuous because every tile was 512x512.
Boundary
Do not perform paired tile-level A/B analysis from archive row order.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
INS-015not supportedPerona-Malik hurts clean CRC tiles and helps only after added noise, so it is a conditional noise treatment rather than an instrument default.
Original proposition
Perona-Malik denoising should be a default preprocessing step.
Evidence tier
Controlled negative result
Key evidence
On clean tiles LDA falls 83.8% to 82.5%; at sigma=0.05 it rises 83.1% to 84.4%.
Boundary
A denoiser changes the chromatic-boundary quantities being measured; trigger only from a validated noise estimate.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
INS-016not supportedThe predeclared profile-shape survival test is null.
Original proposition
Pairwise profile shape adds survival information discarded by aggregate Omega.
Evidence tier
Prespecified negative result
Key evidence
Seven of eight features clear the orthogonality gate; best stage-adjusted p=0.4292 versus Bonferroni 0.00625.
Boundary
This strengthens the null for the tested construction, not for every possible topology representation.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
CRC-001reproducedThe historical benchmark reconstructs at 91.83% accuracy.
The original capped subset was filesystem-order dependent and exact row identity is unrecoverable.
Boundary
Cite as approximate historical reconstruction.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
EBHI-003retractedThe extractor is deterministic; LDA, Ridge, and the automatic segmenter learn from EBHI labels or masks.
Original proposition
The complete EBHI system used no training.
Evidence tier
Construct
Key evidence
Recovered notebook and V4/V5 cross-fitting protocols.
Boundary
Technical portability of the extractor is distinct from supervised grading.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
RETRACTED OR INVALIDATED BY RECONSTRUCTION
EBHI-004internally supportedGrouped V1 gives Ridge r=0.752 (0.707-0.786), LDA 70.6% exact, and 85.4% adjacent after excluding Serrated from the ladder.
Original proposition
The unchanged extractor contains progression-related signal on the full conventional pathway.
Groups are case-like filename proxies, not verified patient identifiers.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED INTERNALLY; EXTERNAL SCOPE OPEN
EBHI-005internally supportedOfficial-mask geometry and ROI summaries provide a substantial information reserve over the image-only frame.
Original proposition
Official masks reveal additional recoverable information.
Evidence tier
Annotation-assisted oracle
Key evidence
V3 best combined arm 85.2% exact and r=0.901 on 2,168 annotated patches.
Boundary
Oracle performance is annotation-assisted and may include annotation convention.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED INTERNALLY; EXTERNAL SCOPE OPEN
EBHI-006internally supportedCross-fitted automatic ROI raises exact accuracy from 69.7% to 74.1% and recovers about one third of the oracle gap.
Original proposition
Automatic localization improves a locked image-only readout.
Evidence tier
Automatic internal
Key evidence
Paired exact gain CI 2.12-6.49 points; held-out Dice 0.756.
Boundary
Segmenter is supervised by EBHI masks; no external site tested.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED INTERNALLY; EXTERNAL SCOPE OPEN
EBHI-007not supportedDice improved from 0.756 to 0.829 while every V5-versus-V4 downstream interval crossed zero.
Original proposition
Higher segmentation Dice improves the downstream endpoint.
Evidence tier
Negative result
Key evidence
Training-mask-only topology calibration and locked downstream readout.
Boundary
Overlap is not a sufficient surrogate for preserved diagnostic structure.
Observed direction
Dice increased from 0.756 to 0.829; no locked downstream V5-versus-V4 interval excluded zero.
Why this status
REFUTED BY ADVERSARIAL CONTROL
EBHI-008not supportedDeep interior and immediate outer context outperform the thin inner boundary alone.
Original proposition
Thin boundaries contain most of the useful information.
Evidence tier
Mechanism
Key evidence
V6 single-compartment comparisons.
Boundary
Bands are annotation-relative coordinates, not validated epithelium, lumen, or stroma.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
EBHI-009supportedIdentical H&E summaries conditioned on ordered distance compartments outperform monolithic pooling and equal-dimensional count-matched random partitions; the ordered-versus-random contrast survives a fixed covariance-shrinkage grid.
Original proposition
Structural address adds information beyond pooled morphology.
Evidence tier
Controlled mechanism with post-result regularization sensitivity
Key evidence
Stable 5/15 representation 85.3% versus 81.0% monolithic oracle; +3.32 points versus the strongest archived random control (95% CI 1.65–5.13). Across auto and nine fixed shrinkage values from 0 to 1, ordered exceeded the strongest of twenty random arms by 2.49–8.03 points; every group-bootstrap interval remained above zero.
Boundary
Strong internal annotation-assisted mechanism. V6.2 is a post-result sensitivity analysis, not external reproduction or optimization of a replacement shrinkage value; absolute accuracy degrades under heavy shrinkage.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED WITH STATED BOUNDARY
Notes
V6.2 extends, rather than replaces, the stable V6.1 result.
EBHI-010qualifiedLegacy V6 is path-reproducible but ill-conditioned; V6.1 shrinkage results supersede it, and V6.2 shows that the ordered-versus-random contrast is not confined to the automatic-shrinkage choice.
Original proposition
Legacy V6 86.6% is a stable final point estimate.
Evidence tier
Numerical audit plus post-result covariance sensitivity
Key evidence
Machine-scale feature perturbations changed 136 predictions under legacy LDA. Automatic-shrinkage V6.1/V6.2 predictions match exactly; across ten shrinkage settings, the conservative ordered-minus-random contrast remains positive with all group-bootstrap intervals above zero.
Boundary
Carry forward stable V6.1 values, not 86.6% as a production number. The comparative representation result is robust; absolute classifier accuracy is not invariant to regularization.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
Notes
V6.2 closes the covariance-parameter sensitivity question without reopening model tuning.
EBHI-011not supportedEBHI is an internal benchmark and mechanism sequence, not external clinical validation.
Original proposition
EBHI establishes clinical or independent-patient grading performance.
Evidence tier
External validity
Key evidence
No verified patient split, external site, or prospective outcome.
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
STH-001reproducedA bounded 120 MB pilot from frozen Hugging Face revision a01e22c…e548 preserves 1,408 patches, metadata, hashes, protocols, and analysis outputs.
Original proposition
The STHELAR structural-address experiment is traceable to a frozen public-data revision.
Evidence tier
Acquisition and provenance
Key evidence
Four spatial blocks per slide were acquired from 11 eligible breast, pancreatic, and skin slides; the local validator passes source and analysis manifests.
Boundary
The pilot is a bounded sample of the public release, not the complete 17.6 GB corpus.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
COMPUTATION OR RESULT REPRODUCED
STH-002not supportedH&E summaries are explicit and deterministic, but the multinomial logistic readout is supervised and the broad labels are RNA-derived and curated rather than pathology gold standard.
Original proposition
The STHELAR result is training-free clinical cell classification.
Evidence tier
Construct
Key evidence
Five broad classes were evaluated with held-out-slide predictions; labels derive from Tangram guidance, slide-specific clustering, marker evidence, and curation.
Boundary
Use 'supervised readout over deterministic summaries'; do not claim clinical cell typing or pathology ground truth.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
STH-003supportedAcross 17,492 high-confidence cells from 11 held-out slides, ordered radial summaries outperform pooled and count-matched random representations.
Original proposition
Nucleus-relative structural address adds information beyond pooled or count-matched random morphology.
Evidence tier
External-data controlled mechanism
Key evidence
Sealed primary ordered-minus-random mean slide macro-F1 delta +0.0537; 11/11 positive; exact two-sided sign-flip p=0.0009766; slide bootstrap interval +0.0336 to +0.0735.
Boundary
External public dataset and different reference geometry; not independent-team validation or unseen-tissue generalization.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED WITH STATED BOUNDARY
STH-004not supportedMeaningful nucleus-relative order exceeds a contiguous shifted anchor, and addressed context remains informative after removing target-nucleus pixels.
Original proposition
The gain is explained merely by any contiguous local partition or by target-nucleus isolation.
Evidence tier
Post-result adversarial controls
Key evidence
Deterministic amendment: ordered-minus-shifted +0.0532 on 11/11 slides; ordered-context-minus-pooled-context +0.0360 on 11/11; full order adds +0.0155 over target plus pooled context on 9/11.
Boundary
Controls were declared after the primary result and are mechanism amendments, not a new prespecified primary endpoint.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
STH-005not supportedThe frozen ordered-minus-random slide delta is positive within breast, pancreatic, and skin material.
Original proposition
The address advantage is driven by only one tissue family.
Evidence tier
Post-freeze descriptive closeout
Key evidence
Breast 4/4, pancreatic 3/3, and skin 4/4 slides are positive.
Boundary
Descriptive support across represented tissues; no claim about unseen tissues, institutions, scanners, or label ontologies.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
STH-006reproducedTwo independent single-thread evaluations match exactly for predictions, contrasts, and complete results.
Original proposition
The final STHELAR analysis is exactly reproducible under a declared numerical environment.
Evidence tier
Numerical determinism
Key evidence
OMP, OpenBLAS, MKL, Accelerate, and NumExpr threads were fixed to one; all equality gates are true.
Boundary
Numerical repeatability does not establish external biological validity.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
COMPUTATION OR RESULT REPRODUCED
ADDR-001not supportedContext-aware cell classification, spatial slide models, tumor interfaces, and spatial null models are established prior art.
Original proposition
Parallax introduced spatial context, tissue compartments, or spatial randomization to computational pathology.
Evidence tier
Prior art
Key evidence
Targeted adversarial audit records 11 primary-source comparators from 2015-2026.
Boundary
Do not claim invention of compartmental pathology, contextual cell typing, or permutation tests.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
ADDR-002supportedParallax operationalizes structural address as a matched ablation: preserve the explicit H&E measurement family and source support while pooling or disrupting positional identity.
Original proposition
The EBHI-STHELAR programme contains a distinctive controlled methods contribution.
Evidence tier
Cross-dataset methods synthesis
Key evidence
EBHI supplies pooled and count-matched random tissue-boundary controls; STHELAR adds nucleus-relative order, count-matched random, shifted-contiguous, target-only, and context-only controls.
Boundary
Distinctive combination, not proof of absolute priority; publication claim should center on the assay rather than 'location matters.'
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED WITH STATED BOUNDARY
ADDR-003qualifiedEBHI and STHELAR provide cross-dataset mechanistic corroboration that explicit summaries gain task-relevant information from meaningful structural address under two reference geometries.
Original proposition
The two datasets establish a universal cross-scale law of histologic information.
Evidence tier
Programmatic synthesis
Key evidence
Concordant ordered-address advantages occur at annotation-boundary and nucleus-relative scales.
Boundary
Call this an organizing principle or cross-scale synthesis; universal wording requires further external tasks and institutions.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
GTEX-001reproducedThe recovered v3 and frozen v4.1 pipelines regenerate all 1,500 rows from source-verified patches.
Original proposition
The historical GTEx computation is authentic.
Evidence tier
Computational provenance
Key evidence
19/25 artifacts byte-identical; 25/25 equivalent within 1e-12; every PNG pixel-exact to source SVS.
Boundary
Authenticity does not establish biological generalization.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
COMPUTATION OR RESULT REPRODUCED
GTEX-002qualifiedPooled sample centers are computed without patch-class labels, but a finite sample mean is not latent class discovery.
Original proposition
Structural centers emerge without labels.
Evidence tier
Descriptive
Key evidence
Three slides each yield a reproducible pooled center; clustering aligns only moderately with donors.
Boundary
Slide identity and row-major geography are supplied.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
GTEX-003qualifiedRandom halves are stable, while geographic halves and quadrants drift substantially.
Original proposition
Each donor has one homogeneous stable center.
Evidence tier
Spatial
Key evidence
All nine v4.1 geographic separations exceed coordinate-permutation nulls.
Boundary
The pooled center depends on sampled geography.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
GTEX-004not supportedWithin-slide standardization induces much of the shared radial scale; donor-specific departures remain.
Original proposition
Omega has a universal donor-independent scale.
Evidence tier
Construct
Key evidence
Fitted correlated-Gaussian null recreates the center-and-tail form.
Boundary
Distribution shape and geography are more informative than a shared mean near 2.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
GTEX-005openPreparation sensitivity is plausible, but fixation is confounded with donor, tissue compartment, stain, scanner, and geography.
Original proposition
Delta-r inversion is caused by fixation chemistry.
Evidence tier
Mechanism
Key evidence
Only three unpaired PAXgene donors.
Boundary
Paired PAXgene/FFPE sections from matched donor, region, batch, and scanner.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
OPEN OR UNRESOLVED
GTEX-006qualifiedThe kernels execute reproducibly on PAXgene material; invariant biological calibration is untested.
Original proposition
The frame generalizes across preparation methods.
Evidence tier
Technical portability
Key evidence
1,500 v3 and 1,500 v4.1 rows regenerated without failures.
Boundary
Use 'technical portability,' not cross-preparation invariance.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
SURV-001retractedNo preserved construction reproduces the exact published hazard ratios.
Original proposition
Published composite OS HR=1.687 and DSS HR=1.872.
Evidence tier
Historical reconstruction
Key evidence
Legacy reconstruction gives OS 1.530; corrected first wave 1.425; all-pool corrected 1.352.
Boundary
Do not use the published point estimates.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
RETRACTED OR INVALIDATED BY RECONSTRUCTION
SURV-002qualifiedHR 1.425 is a reproducible first-wave stage-adjusted estimate, not a stable biological constant.
Original proposition
Corrected-normalizer OS HR near 1.425 is real.
Evidence tier
Specification-bound
Key evidence
HR 1.425 [1.093-1.858], p=0.0089; exact-site and eligibility controls attenuate it.
Boundary
Historical exploratory estimate only.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
External grouped study; prespecify spatial density and neighborhood.
Observed direction
The nominal stage-adjusted adverse association attenuated to null under exact-site control.
Why this status
NOT IDENTIFIED UNDER SITE CONFOUNDING
SURV-007openDirectional tendencies exist, but no frozen distribution predictor survives the full site, redraw, permutation, multiplicity, and eligibility audit.
Original proposition
Distribution broadly beats mean for survival.
Evidence tier
Hypothesis
Key evidence
Smallest exact-site BH q is non-significant; tile draws move conclusions.
Boundary
Independent cohort with frozen summaries.
Observed direction
Several distribution summaries were directional, but no frozen predictor survived site, redraw, permutation, multiplicity, and eligibility gates.
Blinded pathology review of entropy extremes and state composition.
Observed direction
The entropy direction was internally consistent, but the measurement may reflect dominance by an adverse state rather than generic diversity.
Why this status
OPEN — BIOLOGICAL INTERPRETATION UNRESOLVED
SURV-012not supportedOutcome labels did not enter image vocabulary learning, but the K6 state dictionary used cohort-wide images outside fold-local training.
Original proposition
The section-survival analysis is fully inductive.
Evidence tier
Evaluation design
Key evidence
Clinical scaling and references fold-local; image-only vocabulary cohort-wide.
Boundary
Refit states inside training folds or fix them from external material.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
SURV-013not supportedA protective direction recurs, but the published p=0.037 and robust standalone predictor claim do not reproduce.
Original proposition
std_omega robustly predicts overall survival.
Evidence tier
Historical axis
Key evidence
Nested 50/100/200-tile estimates are HR 0.733 p=0.071, 0.761 p=0.104, and 0.756 p=0.105; redraws are unstable.
Boundary
Treat the direction as a hypothesis; do not use std_omega as a validated predictor.
Observed direction
The protective sign recurred across nested tile draws, while magnitude and significance were unstable.
Why this status
UNRESOLVED AT CURRENT POWER AND REDRAW STABILITY
SURV-014retractedEleven matches no recovered criterion.
Original proposition
Eleven axes are directionally consistent and significant for both OS and DSS.
Evidence tier
Historical screen
Key evidence
The audit finds six under the strict reading, 22 under a looser reading, and 90 direction-only axes.
Boundary
The screen was exploratory across many axes and endpoints.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
RETRACTED OR INVALIDATED BY RECONSTRUCTION
SURV-015qualifiedThe historical composite is stronger in the Stage III/IV subgroup, but this is a post-hoc, site-unresolved internal observation tied to a specification-bound composite.
Original proposition
Survival signal is established as concentrated in Stage III/IV disease.
Use the persistence ladder rather than invariant-fingerprint language.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
ID-008qualifiedArtifact-controlled HTAN supplies a six-curve 40–300-micrometre depth-order result at both tested scales, but the complete specimen and block hierarchy remains open.
Original proposition
Structural similarity decays with physical and biological distance.
Evidence tier
Cross-study model
Key evidence
At 0.50 micrometres per pixel: rho=0.453, 5/6 positive, p=0.00758; at native scale: rho=0.391, 5/6 positive, p=0.0157.
Boundary
One calibrated depth rung at two tested scales does not establish a universal decay law or distinct-portion and distinct-block rungs.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
Next decisive test
Balanced repeated depths, exact block provenance, replicate sections, then distinct-portion and distinct-block rungs.
Notes
Supersedes the V1-only scale boundary; preserves V1 as historical provenance.
ID-009supportedSame-portion candidate slides preserve a discriminative section composition and 42-axis structural repertoire beyond within-site patient-disjoint expectation.
Original proposition
Deterministic structural repertoire persists between TCGA slides sharing patient, vial, and tissue portion.
Same-portion candidate relationship is barcode-supported; immediate serial adjacency, cross-block identity, and tumor-wide fingerprinting are not.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED WITH STATED BOUNDARY
ID-010not supportedTissue-mask registration did not reveal reliable additional coordinate-level similarity beyond within-pair shuffled correspondence.
Original proposition
Same-portion persistence is pointwise invariant at registered 512-pixel locations.
Evidence tier
Frozen registered-local endpoint with pair-bootstrap interval
Key evidence
16/18 pairs qualified, 320 paired locations; one-sided Wilcoxon p=0.188; pair-bootstrap mean-advantage 95% CI −0.037 to 0.226.
Boundary
The interval includes both small disadvantage and moderate advantage. This blocks pointwise-invariance language but is not evidence of equivalence.
Observed direction
Registered fields showed a small median advantage, but p=0.188 and the pair-bootstrap mean-advantage 95% CI spans −0.037 to 0.226.
Why this status
NOT DETECTED; NOT EQUIVALENT
ID-011supportedThe frozen TCGA-derived 42-axis readout distinguishes sections belonging to the same tumor in public HER2-positive breast-tumor material.
Original proposition
Adjacent-section persistence is detectable in another public tissue dataset.
Evidence tier
External dataset/tissue corroboration with exact partition inference
Key evidence
Twelve sections from four tumors; median within-tumor distance 6.415 versus 16.506 between tumors; exact partition p=1/15,400=0.0000649; MRR 0.778, exact p=25/15,400=0.001623.
Boundary
The E-H endpoint recognizes tumor-related sections; it does not isolate immediate adjacency from tumor identity. Four tumors are not broad population validation.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED WITH STATED BOUNDARY
ID-012openThe 16-to-32-micrometre contrast is directional but unresolved.
Original proposition
Short-range structural difference increases from 16 to 32 micrometres.
Evidence tier
Prespecified physical contrast
Key evidence
Mean full-repertoire distance rises from 8.501 to 11.386; 3/4 tumors are positive; exact order-permutation p=0.0988.
Boundary
Two separations and four tumors do not establish a decay law.
Observed direction
Distance increased from 16 to 32 micrometres in three of four tumors; exact p=0.0988.
Why this status
UNRESOLVED AT CURRENT POWER
ID-013not supportedA simple coordinate-free edge repertoire outperforms the 42-axis representation in the HER2 adjacent-section cohort.
Original proposition
Parallax uniquely creates or best captures adjacent-section identity.
Evidence tier
Adversarial control
Key evidence
Binary-edge MRR 1.000 and top-1 100%; full structural repertoire MRR 0.778; structural median MRR 0.903.
Boundary
Four-tumor result; the edge control uses local gradient orientation and blockwise density, not tile coordinates or slide silhouettes, and does not establish a universal ranking.
Observed direction
A simple coordinate-free edge repertoire exceeded the 42-axis representation in the four-tumor HER2 cohort.
Why this status
REFUTED BY ADVERSARIAL CONTROL
ID-014qualifiedPositive TCGA, HER2-ST, and artifact-controlled HTAN cohorts support related-section architectural persistence, while DLPFC and simple-image controls prevent universal or Parallax-exclusive wording.
Original proposition
Coordinate-free architectural persistence generalizes across tissues and preparations.
Evidence tier
Cross-dataset synthesis
Key evidence
TCGA same-portion and HER2 tumor-related section retrieval are positive; HTAN reported-depth ordering and 14-piece retrieval survive artifact control at standardized and native scales; DLPFC adjacency is null.
Boundary
Cohorts, preparations, candidate scopes, and physical relationships differ. Existing HTAN pixel controls are pre-Macenko whereas structural axes are normalized, so their magnitudes are not a symmetric representation ranking.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
ID-015qualifiedStructural difference increases across author-reported section separations of 32-160 micrometres.
Original proposition
Frozen structural distance increases across the six-section A-D HER2-ST series on a physical scale.
Evidence tier
Secondary provenance amendment
Key evidence
Four tumors, 24 sections, and 60 dependent within-tumor pairs; Spearman rho=0.322, within-tumor order-permutation p=0.0102; median distance 11.59 at 32 micrometres and 23.17 at 160.
Boundary
The spacing source was recovered after the protocol froze; patient trends are heterogeneous and the result is not a universal decay constant.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
ID-016supportedAcross six independent four-section HTAN tumor curves, artifact-controlled corrected-Macenko structural distance is associated with reported section separation at both tested scales.
Original proposition
Frozen structural distance increases with documented depth across independent human tumor series.
Evidence tier
Prespecified assay with informed artifact sensitivity
Key evidence
Twenty-four sections; three 40-micrometre and three 100-micrometre series. Standardized rho=0.453, p=0.00758; native rho=0.391, p=0.0157; both 5/6 positive.
Boundary
The exact test supports a 40–300-micrometre reported-depth association under the observed acquisition pipeline; it does not isolate biological depth change from depth-associated processing.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED WITH STATED BOUNDARY
Notes
V1 primary was prespecified; V2 artifact gate was fixed after aggregate V1 knowledge but without per-section outcome access.
ID-017qualifiedThe 42-axis depth endpoint converges directionally and inferentially at 0.50 micrometres per pixel and calibrated native scale after clean-field support harmonization.
Original proposition
The exact HTAN full-structural depth result is invariant to analysis resolution.
The structural path is fixed-target corrected-Macenko; pixel controls are pre-normalization. They establish broad source-image persistence but do not provide a symmetric magnitude ranking against normalized structure.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED WITH STATED BOUNDARY
Notes
Artifact-controlled replacement for the V1 expanded-cohort values.
ID-019not supportedThe pooled artifact-controlled HTAN interval slope is positive, but piece-level uncertainty includes zero and interval is confounded with tissue piece and protocol.
Original proposition
All known HTAN section intervals define one common linear persistence rate.
Between-minus-within distance is +4.75 for HT243B1 and −6.37 for HT480B1.
Boundary
Two contradictory cases cannot identify a portion/block hierarchy; exact block provenance and a larger balanced design remain necessary.
Observed direction
The two same-case/different-piece contrasts had opposite signs.
Why this status
UNRESOLVED — CONTRADICTORY CASES
Notes
Artifact-controlled replacement for the V1 contrasts.
SITE-001supportedSite signal is broad and includes Omega-family measures.
Original proposition
Tissue-source site affects structural axes.
Evidence tier
Confounding audit
Key evidence
Patient-level mean_omega p=1.6e-05 and std_omega p=5.6e-06 across four principal sites; many other axes stronger.
Boundary
Site signal may combine technical and biological differences.
Observed direction
Broad, reproducible site association is present; its technical and biological components are not yet separated.
Why this status
SUPPORTED — CANDIDATE QC CHANNEL
SITE-002retractedNo site-clean family was identified in the TCGA measurements.
Original proposition
Omega is the site-clean family.
Evidence tier
Confounding audit
Key evidence
Tile and patient-level site tests reject the original Omega-clean claim.
Boundary
All TCGA outcome associations require site-aware inference.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
RETRACTED OR INVALIDATED BY RECONSTRUCTION
SITE-003not supportedCorrected normalization does not eliminate site-associated structure.
Original proposition
Color normalization removes collection-site signatures.
Evidence tier
Generalization
Key evidence
Corrected composite and Moran vary strongly across site; exact-site control changes inference.
Boundary
Balanced multi-site external study.
Observed direction
Corrected normalization did not eliminate site-associated structure.
Why this status
REFUTED BY SITE-AWARE CONTROL
SITE-004openThe audit establishes site association, not whether it arises from technical preparation, biological population structure, clinical case mix, or a mixture.
Original proposition
The observed site signal is proven to be technical artifact.
Evidence tier
Causal interpretation
Key evidence
Site-stratified attenuation and broad feature differences identify confounding risk but not its cause.
Boundary
Balanced acquisition or paired preparation/scanner study with prespecified site decomposition.
Observed direction
Site association is established, but technical preparation, scanner, population, and case-mix contributions remain unresolved.
Why this status
OPEN — CAUSAL SOURCE UNIDENTIFIED
BLOCK-001not supportedThe 30 V3 companions share patient and tumor sample type with DX1 but use different vial/portion codes; they are patient-level companion preparations, not barcode-supported same-block pairs.
Original proposition
The downloaded V3 TS/BS slides are same-block matches to the DX1 gallery.
Evidence tier
On-disk barcode audit
Key evidence
V3 DX-versus-tissue-slide barcode relationships differ from the eighteen later-acquired V4 same-portion TS/BS/MS pairs.
Boundary
Do not interpret the V3 failure as a serial-section failure.
Observed direction
The V3 companions differ in vial or portion from the DX1 gallery and are not same-block pairs.
Why this status
INVALIDATED BY BARCODE PROVENANCE
BLOCK-002not supportedTCGA-A6-A5ZU DX1/DX2 is a diagnostic-companion pair, but the 01Z-00 barcode does not establish a shared tissue portion or block.
Original proposition
An on-disk DX1/DX2 pair proves same-block persistence.
Evidence tier
On-disk barcode audit
Key evidence
Full mark ranks the correct-patient DX2 28/174; normalized grayscale rank 35/174.
Boundary
Single diagnostic-companion case; its physical specimen relationship is unresolved.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
NOT ESTABLISHED; SEE EVIDENCE TIER AND BOUNDARY
BLOCK-003supportedAll eighteen prespecified GDC companion slides are now local, completing eighteen candidate pairs sharing patient, vial, and tissue portion.
Original proposition
The local inventory contains a tractable same-portion persistence cohort.
Evidence tier
Metadata and acquisition audit
Key evidence
Post-acquisition audit finds 18 complete strong same-portion candidate relationships and zero focused companions still absent.
Boundary
The barcode supports a same-portion candidate relationship; it does not prove immediate serial adjacency or perfect tissue correspondence.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED WITH STATED BOUNDARY
BLOCK-004reproducedEvery metadata-matched local WSI, including the eighteen new companions, passes expected GDC byte-size and MD5 checks.
Original proposition
The locally downloaded whole-slide companions are byte-authentic to the preserved GDC inventory.
Evidence tier
File integrity
Key evidence
Post-acquisition audit verifies 49/49 metadata-matched WSIs; the isolated acquisition reconciliation independently verifies 18/18 new files.
Boundary
File authenticity does not establish biological pairing or adequate tissue content.
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
COMPUTATION OR RESULT REPRODUCED
BLOCK-005not supportedMatching patient, vial, and portion establishes a strong same-portion candidate, but TCGA barcodes do not encode cut interval or exact adjacency.
Original proposition
Matching TS/BS/MS barcodes prove immediately adjacent serial sections.
Evidence tier
Biospecimen semantics
Key evidence
Official GDC hierarchy and local metadata audit.
Boundary
Call the cohort same-portion candidates and use registration/QC to characterize actual tissue correspondence.
Observed direction
Matching TS/BS/MS patient-vial-portion codes do not encode immediate adjacency or cut interval.
Why this status
INVALIDATED BY BARCODE SEMANTICS
BLOCK-006qualifiedNo target was found on currently mounted, accessible local storage, but absolute absence cannot be proven for offline, unmounted, or operating-system-inaccessible storage.
Original proposition
The 18 GDC-listed companion slides are proven absent from every storage location on the computer.
Evidence tier
Filesystem search
Key evidence
A read-only search scanned 641,847 files on the local disk and mounted volumes; filename, barcode, GDC-ID, exact-size, Spotlight, and MD5 searches found zero candidates.
Boundary
The operational conclusion is 'not currently local in searched accessible storage,' not 'deleted' or 'lost.'
Observed direction
See current_claim and key_evidence; no additional directional interpretation assigned.
Why this status
SUPPORTED ONLY WITH QUALIFICATION
ID-021not supportedThe primary depth and expanded related-section findings survive conservative exclusion of the visible perimeter and dark obstructions.
Original proposition
Visible fiducial and dark source-image artifacts create the HTAN depth and retrieval findings.
Evidence tier
Informed artifact-controlled sensitivity
Key evidence
Primary standardized p=0.00758; native p=0.0157; expanded structural MRR p=0.000050 at both scales; 579/579 primary and 340/340 expanded checks pass.
Boundary
The informed sensitivity analysis cannot prove every residual artifact absent; strict-interior sampling also changes tissue support. It shows the visible marks are not necessary for the findings.
Observed direction
Depth order and retrieval survived conservative exclusion of visible perimeter and dark obstructions.
Why this status
REFUTED BY ARTIFACT SENSITIVITY
Next decisive test
Independent serial sections with clean source rasters and balanced known depths.
Notes
New artifact-specific claim created after visual audit of the public TIFFs.
ID-022not supportedPositive reported-depth ordering occurs in both geometry and optical/kernel families at both tested scales.
Original proposition
The HTAN reported-depth association is confined to optical or stain-sensitive structural axes.
Reviewer-motivated post hoc localization using correlated, shared-preprocessing families; it does not isolate a biological mechanism.
Observed direction
Geometry and optical/kernel families were positive at both tested scales; relational results were weaker and scale-dependent.
Why this status
REFUTED BY POST HOC DECOMPOSITION
Next decisive test
Predeclared family-specific replication with balanced intervals and processing controls.
Notes
Added during 2026-09-01 reviewer reconciliation.
ID-023not supportedOriginal raw pixel controls are not directly rankable against normalized structure; the completed symmetric control separately establishes strong normalized chromatic/compositional depth order.
Original proposition
Raw-color HTAN control magnitudes are directly comparable with the corrected-Macenko 42-axis representation.
Evidence tier
Preprocessing audit plus repaired symmetric control
Key evidence
The fair control applies one fixed corrected-Macenko target to 17,400 TCGA reference and 4,773 HTAN fields; standardized color rho=0.653, p=0.000139; native rho=0.730, p=0.0000174.
Boundary
The old raw-versus-structure magnitude comparison remains invalid. The fair control is non-specific and does not identify a chromatic mechanism.
Observed direction
The raw pixel controls and corrected-Macenko structural features were not passed through the same preprocessing path.
Why this status
INVALIDATED BY PREPROCESSING ASYMMETRY
Next decisive test
Balanced acquisition controls or restain/rescan perturbations that separate composition from chemistry and imaging.
Notes
Updates the initially pending symmetric-control disposition.
ID-024not supportedColor summaries remain strongly associated with reported separation when the identical corrected-Macenko target is applied to reference and query fields.
Original proposition
Fixed-target normalization removes the HTAN color depth-order signal.
Normalized color combines tissue composition, residual stain, illumination, and other chromatic content; it does not identify a causal chromatic mechanism.
Observed direction
Normalized color remained strongly depth-ordered after identical fixed-target preprocessing of reference and query fields.
Why this status
REFUTED BY SYMMETRIC CONTROL
Next decisive test
Balanced restain/rescan and composition-matched serial-section controls.
Notes
Added after mounting the preserved external HTAN field store.
ID-025not supportedSeveral pre-normalization stain-vector and concentration parameter distances increase with reported separation.
Original proposition
Estimated source-stain state remains stable across reported serial depth.
Evidence tier
Post hoc technical-channel audit with multiplicity control
Key evidence
Ten of sixteen arm-by-parameter exact tests retain BH q≤0.05, including multiple hematoxylin- and eosin-vector components.
Boundary
The six-curve post hoc result preserves a technical or chemical alternative; it does not prove that source-stain divergence causes the structural response.
Observed direction
Ten of sixteen source-stain parameter-distance tests retained BH q≤0.05.
Why this status
REFUTED BY SOURCE-STAIN ANALYSIS
Next decisive test
Balanced staining-batch, section-order, thickness, focus, and scanner metadata or controlled perturbations.
The frozen CAMELYON17-WILDS partition gate technically qualified for scientific execution.
Evidence tier
Frozen external-corpus technical gate
Key evidence
24 slides, five centers, 12,000 selected patches; 11,996 complete cases (99.967%); no cross-slide exact duplicates; scale-family median resampling concordance 0.977–1.000; independent reconstruction passed.
Boundary
A technically qualified 96×96 patch assay is not native-WSI, patient-held, clinical, or foundation-model validation.
Observed direction
All technical and scale gates passed; four complete-case failures were excluded identically from all arms.
Why this status
SUPPORTED WITH STATED BOUNDARY
Next decisive test
Technical execution completed in CPTAC-CCRCC V1; scientific transport is adjudicated separately in CPTAC-002 and CPTAC-003.
Notes
Four pre-result amendments are part of the frozen record.
CAM-002supportedAdding the complete 42-axis structural block to raw-color summaries improves slide-macro AUROC under buffered-geographic, slide-held, and center-held partitions.
Original proposition
The complete named structural block adds task-relevant discrimination beyond fixed raw-color summaries under separated CAMELYON evaluation.
AUROC increments: P1 +0.0440, 99% CI 0.0043–0.0919; P2 +0.0451, 0.0145–0.0767; P3 +0.0464, 0.0029–0.0939. Balanced-accuracy increments were +0.0800, +0.0590, and +0.0724.
Boundary
This establishes complementarity to one fixed raw-color summary under one supervised readout. It does not isolate biological architecture from preparation or establish universal transport. The later frozen CPTAC patient-held native-WSI pilot did not reproduce the broad increment.
Observed direction
The nested increment is positive in P0–P3 and passes the frozen 99% decision rule in every partition.
Why this status
SUPPORTED WITH STATED BOUNDARY
Next decisive test
Completed by CPTAC-CCRCC V1; the broad patient-held transport claim was not supported.
Notes
PASS-2 was the only frozen passing route. CPTAC-002 limits transport beyond this CAMELYON assay.
CAM-003not supportedThe four partitions do not form a monotone performance ladder; slide-held performance often exceeds buffered-geographic performance.
Original proposition
Random, buffered-geographic, slide-held, and center-held CAMELYON partitions form a monotone leakage or generalization ladder.
Evidence tier
Frozen four-partition comparison
Key evidence
PASS-1 failed under the prespecified all-representation monotonicity rule; full-structural AUROC was 0.8270 at P1 and 0.8485 at P2.
Boundary
Partition difficulty depends on composition and split geometry. The experiment does not refute leakage generally; it refutes this proposed ordinal ladder in this assay.
Observed direction
Performance generally declines from random to center held, but the intermediate rungs are nonmonotone.
Why this status
REFUTED BY FROZEN PARTITION CONTROL
Next decisive test
None within CAMELYON; require a separately justified design rather than relabeling these partitions as ordinal.
Notes
The generic leakage-paper route is closed by this gate.
CAM-004not supportedThe combined-minus-raw increment does not materially increase from random to center-held evaluation.
Original proposition
Structure and raw color show a stable representation-by-partition interaction from random to center-held evaluation.
P3-minus-P0 interaction +0.00386; 99% CI −0.03055 to 0.04370; only one of five centers had the prespecified same-sign condition.
Boundary
A positive nested increment persists across partitions; its relative advantage does not demonstrably change with partition severity.
Observed direction
The nested structural increment is similar, not increasingly dominant, across P0–P3.
Why this status
NOT IDENTIFIED UNDER FROZEN INTERACTION TEST
Next decisive test
No further CAMELYON tuning; the completed CPTAC patient-held pilot did not support broad transport.
Notes
PASS-3 failed.
CAM-005qualifiedTwenty within-center randomized 42-column controls do not reproduce the observed structural increment.
Original proposition
The CAMELYON nested gain is explained by adding 42 arbitrary dimensions.
Evidence tier
Post-result dimension-matched sensitivity
Key evidence
Observed increments exceed the maximum random-control increment in P0–P3; random controls reduced AUROC on average. Empirical comparison floor 1/21=0.0476.
Boundary
Post hoc sensitivity supports content rather than dimension count but does not upgrade or independently reproduce the frozen PASS-2 result.
Observed direction
Randomized structural columns harmed rather than improved performance in all four partitions.
Why this status
SUPPORTED ONLY AS POST-RESULT SENSITIVITY
Next decisive test
The prespecified CPTAC patient-held assay was completed; its analogous broad transport rule failed.
Notes
Added after the frozen gate disposition.
CAM-006not supportedNo individual geometry, relational, or optical/kernel addition reproduces the complete confirmed increment with a 99% interval above zero; normalized grayscale texture is the strongest standalone family under P1–P3.
Original proposition
One structural subfamily uniquely carries the CAMELYON nested gain, or the combined representation is uniformly best.
Evidence tier
Post-result family localization plus frozen standalone comparison
Key evidence
Individual nested-family increments range from +0.0050 to +0.0330 and all 99% intervals cross zero. Normalized-gray AUROC is 0.8962, 0.9066, and 0.8995 at P1–P3, exceeding raw plus structure there.
Boundary
The supported increment belongs to the complete 42-axis combination. This assay does not identify one mechanism or a universally superior representation.
Observed direction
All three nested subfamilies are directionally positive, but none is individually resolved at the frozen confidence level.
Why this status
REFUTED BY FROZEN AND POST HOC FAMILY COMPARISONS
Next decisive test
Do not carry a CAMELYON-selected subfamily forward; CPTAC tested independently defined broad and mechanistic contrasts.
Notes
PASS-4 failed; post hoc localization is descriptive.
CPTAC-001supportedThe prespecified CPTAC-CCRCC patient-held native-WSI pilot completed with qualified image sampling, complete frozen measurements, and independently reconstructed inference.
Original proposition
A native-WSI, patient-held CPTAC-CCRCC assay can be executed under the frozen bounded-pilot design.
Evidence tier
Frozen external-corpus technical qualification with independent reconstruction
Key evidence
80 patients and 160 slides selected; 79 patients, 158 slides, and 10,112 fields qualified after one technical exclusion; all fields measured; zero cross-slide exact duplicate groups; independent validator passed 37 checks.
Boundary
Technical qualification and patient-held execution do not establish predictive utility, biological specificity, or clinical performance.
Observed direction
The assay completed as frozen on 79 technically viable patients.
Why this status
SUPPORTED WITH STATED BOUNDARY
Next decisive test
None for execution validity; interpret the two frozen scientific contrasts separately.
Notes
One patient was excluded without replacement because one of two required slides yielded 43 rather than at least 48 viable fields; the original fold map was retained.
CPTAC-002not supportedThe CAMELYON complete-block increment did not transport as a broad patient-held native-WSI effect in this CPTAC-CCRCC pilot.
Original proposition
The complete structural block adds broad patient-held native-WSI discrimination beyond fixed raw-color summaries, transporting the CAMELYON nested gain.
Patient AUROC raw plus full structure 0.6282 versus raw color 0.6513; contrast −0.0231, 95% CI −0.0846 to 0.0314, 99% CI −0.1045 to 0.0462; one of five folds positive.
Boundary
This refutes transport of this frozen representation/readout/endpoint in this 79-patient pilot. It does not show that structural measurements never add value in other tasks or assays.
Observed direction
The point estimate favors raw color rather than raw plus full structure.
Why this status
NOT SUPPORTED BY FROZEN PATIENT-HELD TRANSPORT TEST
Next decisive test
No full CPTAC expansion under this broad transport claim; any future test requires a new rationale and preregistration.
Notes
Narrows CAM-002, which remains supported within its own 96×96 CAMELYON patch assay.
CPTAC-003not supportedThe non-optical increment was positive but unresolved and did not meet the prespecified materiality gate.
Original proposition
Non-optical geometry and relational measurements add a material patient-held increment beyond a strong optical arm.
Evidence tier
Frozen patient-held mechanistic contrast
Key evidence
Patient AUROC strong optical plus non-optical 0.6891 versus strong optical 0.6750; contrast +0.0141, 95% CI −0.0160 to 0.0474, 99% CI −0.0256 to 0.0583; three of five folds positive; frozen gate required at least +0.020 with 95% lower bound above zero and transport above −0.010.
Boundary
A positive point estimate is not confirmation. The interval includes zero, the magnitude is below the frozen threshold, and the accompanying broad transport condition failed.
Observed direction
Small positive non-optical increment, inferentially unresolved.
Why this status
FAILED FROZEN MAGNITUDE, PRECISION, AND TRANSPORT GATE
Next decisive test
Preserve only as a preregisterable lead for a new independent cohort; do not tune or expand this cohort to rescue it.
Notes
The pilot status is SCIENTIFIC_CLOSE despite the positive point estimate.
CPTAC-004qualifiedTwenty train-only randomized non-optical controls did not reach the observed +0.0141 increment.
Original proposition
The small observed non-optical increment is reproduced by arbitrary added dimensions.
Random-control mean −0.00064, range −0.01090 to +0.01346; zero of 20 controls reached +0.01410; finite-control empirical one-sided comparison 1/21=0.0476; independent validator passed 45 checks.
Boundary
This post-primary sensitivity argues against generic dimension count as a sufficient explanation. It cannot repair the crossed primary interval, missed materiality threshold, or failed transport condition.
Observed direction
Observed increment exceeded all 20 randomized-dimension increments by a small margin.
Why this status
SUGGESTIVE POST-PRIMARY SENSITIVITY ONLY
Next decisive test
If pursued, preregister the exact non-optical contrast and dimension control in a new independent cohort.
Notes
The control was frozen before its outcomes were computed, but it followed the primary result and remains secondary.
CPTAC-005supportedThe frozen result is SCIENTIFIC_CLOSE and closes the planned full CPTAC expansion under this hypothesis.
Original proposition
The bounded CPTAC gate supplies a prespecified decision on full-cohort expansion under this hypothesis.
Evidence tier
Prespecified decision-gate disposition
Key evidence
Neither TRANSPORT_ONLY_PASS nor MECHANISTIC_PASS criteria were satisfied; the disposition was independently reconstructed from the frozen thresholds.
Boundary
Closing this branch is a study decision, not evidence that every alternative structural representation or pathology endpoint must fail.
Observed direction
Both passing routes failed under their frozen criteria.
Why this status
PRESPECIFIED SCIENTIFIC CLOSE
Next decisive test
Package the negative transport result and prioritize already supported structural-address work or a genuinely new externally justified assay.
Notes
Do not analyze the reserved patients, tune on this cohort, or add a foundation model as a rescue operation.
CPTAC-006qualifiedMultiple named structural axes track percent necrosis, total cellularity, and tumor-nuclei percentage within this cohort.
Original proposition
The 42-axis structural panel is sensitive to pathology-annotated tissue composition in the CPTAC-CCRCC pilot.
Evidence tier
Post-primary exploratory metadata localization
Key evidence
At BH q≤0.05 within each 42-axis metadata family, 16 axes associate with necrosis, 16 with cellularity, and 11 with tumor-nuclei percentage; aggregate structural-score rho values are +0.393, −0.406, and −0.325 respectively.
Boundary
Post-discovery associations with collection-provided annotations are an interpretable sensitivity result, not external validation of a tissue-quality assay or a causal biological mechanism.
Observed direction
Structural score rises with necrosis and falls with cellularity and tumor-nuclei percentage.
Why this status
SUPPORTED ONLY AS POST-PRIMARY INTERNAL SENSITIVITY
Next decisive test
Freeze a small composition/QC assay in an independent renal or multi-tissue WSI cohort with independently scored necrosis and cellularity.
Notes
Does not alter CPTAC-002, CPTAC-003, or the SCIENTIFIC_CLOSE grade disposition.
CPTAC-007not supportedNo within-slide standard-deviation feature in the structural, raw-color, grayscale, or edge families survives its stated BH family.
Original proposition
Within-slide feature dispersion supplies a hidden grade signal that the frozen means missed.
Evidence tier
Post-primary all-feature univariate screen
Key evidence
The screen covered 42 structural, 27 raw-color, 47 grayscale-texture, and 14 edge features; zero mean-within-slide-SD rows had q≤0.05.
Boundary
This closes the simple univariate dispersion rescue in this acquired pilot; it does not test every spatial distribution or multivariable heterogeneity representation.
Observed direction
No multiplicity-controlled dispersion feature was detected.
Why this status
NOT IDENTIFIED IN CORRECTED ALL-FEATURE SCREEN
Next decisive test
None in this cohort under the frozen close decision.
Notes
Explicitly prevents reinterpretation of the negative grade result as an overlooked standard-deviation effect.
CPTAC-008qualifiedFive absolute orientation bins pass the edge-family screen, and the complete eight-bin profile remains nonrandom after label permutation within identifier prefix.
Original proposition
Binary CCRCC grade is associated with the absolute edge-orientation profile under the sampled-field assay.
Evidence tier
Post-primary exploratory feature screen plus post-discovery omnibus control
Key evidence
Five of 14 edge summaries have q≤0.05; pooled profile permutation p=0.00464; prefix-stratified p=0.0156; axial-cosine q=0.0163; neither prefix alone resolves the effect.
Boundary
Bins use absolute scanner coordinates. Tissue architecture, section placement, cutting orientation, and acquisition are inseparable; this is not a renal-cancer orientation biomarker.
Observed direction
Low- and high-grade groups differ in absolute edge-direction balance.
Why this status
SUPPORTED ONLY WITH STRONG COORDINATE AND PROVENANCE QUALIFICATION
Next decisive test
Test a rotation-invariant and scanner-randomized version under a frozen independent-cohort protocol.
Notes
Post-discovery lead only; no readout was rebuilt around it.
CPTAC-009qualifiedWithin exact grade and C3L/C3N-prefix candidate sets, the companion portal-distinct specimen slide is strongly retrievable by the 42-axis structural repertoire; substantial relatedness remains after combined composition and gross-extent controls.
Original proposition
Portal-distinct FFPE specimen slides from the same CPTAC-CCRCC patient retain a discriminative coordinate-free structural repertoire within matched candidate sets.
Evidence tier
Post-discovery matched retrieval with exact permutation and independent stress reconstruction
Key evidence
71 FFPE patients, 142 slides; full-structural MRR 0.720 versus 0.282 conditioned chance; top-1 60.6% versus 11.3%; within-group pair-label permutation p=1/50,001; 47-check independent reconstruction. Combined feature- and distance-level composition/extent control leaves MRR 0.634, top-1 47.9%, p=1/50,001; 20-check independent reconstruction.
Boundary
Distinct portal specimen, slide, and container identifiers do not establish independent blocks, preparation sessions, or biological causation. The result is within one collection and is not a universal patient fingerprint.
Observed direction
Patient-paired structural distances are much smaller than matched impostor distances.
Why this status
ROBUST INTERNAL POST-DISCOVERY RESULT WITH UNRESOLVED PROVENANCE
Next decisive test
Repeat the frozen matched retrieval in a cohort with explicit block, preparation, and site provenance plus a site-held external arm.
Notes
A preliminary row-order diagnostic was invalid and discarded. The formal hash-defined-side top-1 is 60.6%, not the preliminary higher estimate.
CPTAC-010not supportedMeasured tissue composition and gross WSI extent carry substantial linkability but are not sufficient explanations for the structural retrieval result.
Original proposition
The distinct-specimen linkability result is fully explained by grade, stage, prefix, measured tissue composition, and gross WSI extent.
Evidence tier
Post-discovery explanatory controls
Key evidence
Composition metadata alone gives MRR 0.375 versus 0.282 chance; extent alone gives MRR 0.627. Feature-level removal of stage, composition, and extent leaves structural MRR 0.709. The combined feature- and pair-distance extent control leaves MRR 0.634 versus 0.282 chance, top-1 47.9% versus 11.3%, p=1/50,001.
Boundary
Residual retrieval still may reflect unmeasured patient-linked tumor biology, specimen handling, block, preparation, scanner, acquisition, or more detailed composition. Nuisance removal is a sensitivity, not causal isolation.
Observed direction
Extent is a major carrier and composition a smaller carrier; a substantial structural residual persists under their combined control.
Why this status
REFUTED BY MATCHED AND CROSS-FITTED COMPOSITION CONTROLS
Next decisive test
Controlled multi-specimen, multi-block, multi-session study with explicit provenance and independent replication.
Notes
Refines rather than replaces CPTAC-009.
CPTAC-011not supportedColor-containing scores vary with C3L/C3N prefix, and the normalized-grayscale score varies with gross WSI size and row extent.
Original proposition
The frozen CPTAC readout scores are independent of coarse provenance and gross WSI extent.
Evidence tier
Post-primary broad manifest audit
Key evidence
Across the all-score BH family, prefix associations retain q=0.0097–0.0108 for four color-containing arms; grayscale score correlates with series size at rho=0.381, q=0.0097, and pixel-matrix rows at rho=0.336, q=0.0298. No manifest field survives the separate grade-label BH family.
Boundary
C3L/C3N is an identifier-derived provenance proxy, not a demonstrated laboratory site. WSI size may reflect tissue extent, scan extent, or file construction; the audit does not identify causality.
Observed direction
Coarse provenance and image extent measurably shift frozen scores.
Why this status
REFUTED BY MULTIPLICITY-CONTROLLED METADATA SCREEN
Next decisive test
Treat provenance and WSI extent as mandatory audit covariates in future native-WSI assays.
Notes
Explains why pooled color-arm grade performance is not safely interpreted as biology.
CPTAC-012qualifiedA four-variable gross-extent representation retrieves paired specimen slides strongly, and direct extent adjustment materially attenuates structural retrieval.
Original proposition
Gross WSI extent is a substantial carrier of portal-distinct same-patient linkability in CPTAC-CCRCC.
Evidence tier
Advisor-motivated frozen post-discovery extent control with independent reconstruction
Key evidence
Extent-only MRR 0.627 versus 0.282 conditioned chance, top-1 44.4% versus 11.3%, p=1/50,001. Pair-distance extent adjustment reduces structural MRR from 0.720 to 0.576, retaining 67.1% of the original excess over chance. All four extent-control outcomes independently reconstruct in 20 checks.
Boundary
Extent combines tissue amount, specimen geometry, scan extent, file construction, and acquisition. It is a technical or gross-sampling carrier, not patient biology or a privacy capability by itself.
Observed direction
Paired specimens share gross extent, which explains part—but not all—of structural linkability.
Why this status
SUPPORTED ONLY AS POST-DISCOVERY INTERNAL MECHANISM
Next decisive test
Acquire explicit block and acquisition provenance and repeat with tissue-area segmentation, scan-extent normalization, and site-held validation.
Notes
Added after advisor review; refines CPTAC-009 and CPTAC-010 without changing the frozen grade disposition.