Research status · Selected claim-specific recordsPublication status · Shown per record where statedArtifact / release status · Shown per record where stated
Evidence · Technical reference
Technical Evidence Register
This register exposes the controlled public evidence subset behind the Evidence Overview.
Each record is claim-specific and versioned. Missing denominators remain NOT STATED. Research/evaluation, publication and artifact/release statuses are separate axes. This is a filtered public register, not the internal working evidence register.
PUBLIC SUMMARY ONLY does not mean that original execution logs are publicly exposed or independently verified. Records whose primary row-level evidence was independently checked say so explicitly; otherwise the public description remains a bounded report or summary claim.
Negative and boundary findings are first-class evidence.
Scan the result and evaluation status, then expand a record for its scope, metric denominator, claim boundary and public sources.
Filters · Showing 17 of 17 evidence records
Showing 17 of 17 evidence records
EV-FND-01Research Arc
Early invariant-preservation foundation
SUPPORTED WITH QUALIFICATIONRESEARCH PROTOTYPE
Key metric / resultExplicit preservation was tractable; no public numeric performance result is proposed.
PUBLIC SUMMARY ONLY
Technical detailsHide technical details
Question tested
What did the early experiments establish about explicit preservation of technical invariants?
Evidence class
FOUNDATIONAL EXPERIMENT
Evaluated material / scope
Corpus A and Corpus B; TRIZ-Aware Loss v1.0 and v2.0; Demo Module 2.0.
Configuration / version
TRIZ-Aware Loss v1.0 · Corpus A; TRIZ-Aware Loss v2.0 · Corpus B; Demo Module 2.0.
Full observed result
Explicit loss terms and protected representations targeted invariant preservation. The v2.0 stage added semantic anchors, bridge pairs and geometric diagnostics. Demo Module 2.0 provided a bounded reproducible proof of concept for that reinforcement stage.
Metric / denominator
An independent held-out denominator is NOT STATED. No public numeric result is proposed.
Supported conclusion
The experiments formed an invariant-preservation foundation and demonstrated bounded semantic reinforcement.
Boundary
They did not establish structural ordering, SIF 2.0 accuracy or production performance. TRIZ-Aware Loss itself did not establish relational ordering.
Publication status
NOT STATED
Artifact / release status
NOT STATED
Public availability
PUBLIC SUMMARY ONLY
Evidence provenance
Controlled source record; public source locator not available.
Key metric / resultBoundary amplification Δ_min / Δ_mean ≈ 2.26; the magnitude probe was not sign-consistent across seeds.
PUBLIC NOW
Technical detailsHide technical details
Question tested
Did corpus order or contradiction magnitude explain the measured structure under the frozen Protocol C setup?
Evidence class
CONTROLLED STRUCTURAL STUDY
Evaluated material / scope
Corpus C, N=142; MarianMT with LoRA; pair order was changed while the declared remaining configuration was held fixed.
Configuration / version
Protocol C; MarianMT + LoRA; Corpus C, N=142.
Full observed result
Random pair-order shuffling produced boundary amplification Δ_min / Δ_mean = 0.0941 / 0.0416 ≈ 2.26 (Module 0 / COIT); this is not an absolute 2.26× coherence decline. Training loss did not separate the conditions; global embedding entropy matched to six decimals (10.972432). The magnitude probe was not sign-consistent across seeds.
Metric / denominator
Corpus N=142; boundary-amplification numerator 0.0941 (weakest-transition degradation), denominator 0.0416 (mean degradation), as reported in Module 0. The magnitude probe reporting basis is NOT STATED; no magnitude number is included.
Supported conclusion
Ordering had a reproducible structural effect in the tested setup. Contradiction magnitude was not supported as the dominant tested driver.
Boundary
This is bounded to Protocol C and its low-data design. It does not establish a universal neural-model result or show that contradiction structure is irrelevant.
Publication status
Module 0 technical note identified; its DOI mapping remains under authoritative bibliographic verification. Ordering Dominance archived manuscript.
Artifact / release status
Frozen study configuration; public DOI records available.
Key metric / resultFirst order 250/252 (99.2%); second order 124/252 (49.2%); no stable third tier.
PUBLIC SUMMARY ONLY
Technical detailsHide technical details
Question tested
Which hierarchy levels were observed inside the tested representation?
Evidence class
REPRESENTATION STUDY
Evaluated material / scope
Corpus D / Module 5.0 multi-seed representation probe; 252 observations.
Configuration / version
Corpus D / Module 5.0; bounded multi-seed probe.
Full observed result
The first-order pattern appeared in 250/252 observations (99.2%); the second-order pattern appeared in 124/252 (49.2%). A stable third tier did not form.
Metric / denominator
First order 250/252 and second order 124/252 within-protocol pair-by-variant observations; exact unit/dependence awaits original scoring-protocol confirmation; not independent replications.
Supported conclusion
The tested diagnostic showed a strong first-order and partial second-order representation hierarchy.
Boundary
Internal representation structure does not establish that the structure will appear in admissible generated output.
Publication status
NOT STATED
Artifact / release status
NOT STATED
Public availability
PUBLIC SUMMARY ONLY
Evidence provenance
Controlled source record; public source locator not available.
Key metric / resultThree failure classes were distinguished; class counts NOT STATED.
PUBLIC SUMMARY ONLY
Technical detailsHide technical details
Question tested
Which types of structural output failure were distinguished in Corpus D2?
Evidence class
DIAGNOSTIC STUDY
Evaluated material / scope
Corpus D2; complete loss, partial damage, identity-level substitution and null controls.
Configuration / version
Corpus D2 failure analysis.
Full observed result
The analysis distinguished complete loss, partial damage and identity-level substitution. The public source summary does not provide case counts by class.
Metric / denominator
Failure-class frequencies and coverage denominator are NOT STATED.
Supported conclusion
The observed structural failures can be distinguished by class.
Boundary
This does not estimate failure frequency or establish that the taxonomy is exhaustive.
Publication status
NOT STATED
Artifact / release status
NOT STATED
Public availability
PUBLIC SUMMARY ONLY
Evidence provenance
Controlled source record; public source locator not available.
Key metric / resultAIP 1.000 on the stated frozen test; per-instance anchor denominator NOT STATED.
PUBLIC SUMMARY ONLY
Technical detailsHide technical details
Question tested
What did the frozen unseen evaluation show on its stated test composition?
Evidence class
STRUCTURAL EVALUATION
Evaluated material / scope
35 rows: 28 non-negative rows and 7 negative controls.
Configuration / version
Registry v2.0.3; Positional EII v1.1.
Full observed result
The frozen evaluation reports AIP 1.000 across a test composition of 35 rows: 28 non-negative and 7 negative controls.
Metric / denominator
35 rows total; 28 non-negative rows; 7 negative controls. Per-instance anchor denominator: NOT STATED. Do not derive a numerator or denominator from AIP or row count.
Supported conclusion
The result supports a claim bounded to this named frozen unseen test.
Boundary
It does not mean 100% accuracy, universal accuracy, denominator-complete success or an aggregate SIF score.
Publication status
NOT STATED
Artifact / release status
NOT STATED
Public availability
PUBLIC SUMMARY ONLY
Evidence provenance
Controlled source record; public source locator not available.
Key metric / resultProtected: 36/36 rows and 73/73 carriers; null: 6/6; baseline: 18/36 rows and 48/73 carriers.
PUBLIC SUMMARY ONLY
Technical detailsHide technical details
Question tested
Were protected technical carriers preserved in the tested component workflow?
Evidence class
COMPONENT BENCHMARK
Evaluated material / scope
42 rows: 36 protected and 6 null; 73 carriers. Registry layer was outside scope.
Configuration / version
Module 2.0 Formula Stress v2.0.
Full observed result
Protected cases passed on 36/36 rows and 73/73 carriers; null controls passed on 6/6. Baseline exact-row performance was 18/36 rows and 48/73 carriers.
The first untouched run was FIRST_UNTOUCHED_RUN_COMPLETE_REVIEW_REQUIRED and RUN_CANDIDATE_NOT_FROZEN. MAIN decisions: 6 VERIFY_REPAIRED, 5 PASS, 3 ESCALATE:REQUIRES_HUMAN_REWRITE, 1 REJECT; row-all-anchors OK 11/15; AIP 0.7895; wrong-exact row rate 0.0667. RESERVE decisions: 2 PASS, 2 ESCALATE:REQUIRES_HUMAN_REWRITE, 1 VERIFY_REPAIRED; row-all 3/5. NULL: 5 PASS and 0/5 false positives. MAIN numbers and symbols were preserved.
Metric / denominator
25 rows = 15 MAIN + 5 RESERVE + 5 NULL. MAIN row-all-anchors OK 11/15; RESERVE row-all 3/5; NULL FP 0/5. MAIN AIP anchor-instance denominator is NOT STATED in the recovered summary.
Supported conclusion
The result shows partial untouched transfer and an important boundary on the tested VIM3 pipeline.
Boundary
The recovered source is a run summary, not the original 25-row scored dataset, so row-level metric verification remains incomplete. This does not establish full VIM coverage, production readiness or general zero-shot transfer.
Publication status
NOT STATED
Artifact / release status
RUN_CANDIDATE_NOT_FROZEN
Public availability
PUBLIC SUMMARY ONLY
Evidence provenance
Recovered source is the v1.0.1 run summary; the original 25-row scored dataset remains unavailable for independent row-level verification.
The DTE-1 transfer reported 35/35 non-negative cases admissible, 0/10 negative activations, 6/6 protected carriers and 0 architecture modifications. Reported study outcome: SUPPORTED WITH SPECIFICATION EXTENSION.
Strict AIP was 40/55 = 0.7273; the diagnostic result was 46/55 = 0.8364. The probe reported 2/10 negative false positives, 14/14 protected carriers, 1/1 semantic non-identity preserved and unsafe release 0. Reported study outcome: NOT SUPPORTED UNDER FROZEN CRITERIA.
The frozen strict criterion was not met; the diagnostic result remains a separate observation.
Boundary
Diagnostic morphology-aware improvement does not replace the frozen strict criterion and is not clinical validation.
Publication status
NOT STATED
Artifact / release status
NOT STATED
Public availability
PUBLIC SUMMARY ONLY
Evidence provenance
Primary DTE-2 final scorer and adjudication workbooks were independently checked during TASK 13; the internal workbooks are not published by this website.
The engineering qualification reports 279 tests passed, a final frozen gate of 12/12 and 0 regressions. TASK 13 recovered a SHA-256 sidecar for the archival source bundle, but not the bundle or original 279-test / 12-gate execution logs. The professional pilot is pending.
Metric / denominator
279 tests; final frozen gate 12/12; regressions 0 in the stated checks.
Supported conclusion
This is an internally reported bounded engineering qualification result for the reference implementation; independent log-level verification remains open.
Boundary
The eight-case public evidence replay is separate static recorded demonstration content and is not the 279-test qualification archive. The recovered sidecar alone does not verify the original execution logs. This does not establish professional-use validation, independent external validation or general correctness. The professional pilot remains pending.
Publication status
Unpublished research evidence; the working skeleton is not a publication.
Artifact / release status
Archival bundle SHA-256 sidecar recovered; source bundle and original qualification logs not recovered in TASK 13.
Public availability
PUBLIC SUMMARY ONLY
Evidence provenance
The public site exposes an eight-case static replay. The 279-test engineering qualification remains supported by internal reporting rather than independently recovered execution logs.
Evaluation outcome describes the evidence result. Research/evaluation status describes the evaluation condition. Publication status and artifact/release status are recorded separately and are not inferred from the result.
Unstated denominators
Where a source does not state a denominator, this register keeps it as NOT STATED. AIP 1.000 is not a universal accuracy claim or an aggregate SIF score.
Controlled outcomes
PASS, VERIFY and ESCALATE are reported result labels only. Negative, partial and boundary outcomes remain visible alongside supported results.
Public identifier gaps
EV-SIF2-06, EV-SIF2-07 and EV-SIF2-08 are not published in this controlled 17-record subset. No public result is asserted for those identifiers, and their absence must not be interpreted as a hidden positive or negative outcome.