Research status · Selected, claim-specific evidence summaryPublication status · Pending site freeze confirmationArtifact / release status · Public Technical Evidence Register available · 17 selected records
Evidence · Overview
Evidence Overview
What evidence supports the major public claims of the SIF research programme?
This page is the human-readable evidence overview alongside the public Technical Evidence Register. It summarises selected, versioned results with the scope and limits needed to read each claim.
No combined SIF score is reported here. Each result applies only to its named material, configuration and evaluation.
How to read this page
A claim is bounded by its evidence
Each surface states the question, evidence basis, observed result, supported claim, boundary and related route. Metrics stay with their named test; separate tests are not pooled.
Research status, publication status and artifact or release status answer different questions. Publication status remains pending site freeze confirmation. The Technical Evidence Register contains a selected public subset; unpublished internal evidence and source materials remain outside the public site.
Evidence map
Eight public evidence surfaces
These surfaces connect the major research claims to bounded observations and visible limitations.
Bounded foundation and falsification
Foundations and falsification
Question
Could explicit preservation protect technical carriers, and did that by itself establish relational ordering?
Evidence basis
TRIZ-Aware Loss v1.0 / Corpus A; TRIZ-Aware Loss v2.0 / Corpus B; Demo Module 2.0; and the bounded contradiction-magnitude probe.
Observed result
Explicit preservation was tractable. Demo Module 2.0 was a bounded reproducible proof of concept. The magnitude probe was not sign-consistent across seeds.
Supported claim
Carrier-level preservation and semantic reinforcement formed a research foundation; contradiction magnitude was not supported as the dominant tested driver.
Boundary
These experiments did not establish structural ordering, production accuracy or universal behaviour. TRIZ-Aware Loss itself is not evidence that relational ordering was solved.
Did corpus order affect chain-level representation under a frozen protocol, and could distinct operating regimes be selected?
Evidence basis
Protocol C: MarianMT with LoRA, Corpus C (N = 142), with corpus pair order changed while the remaining declared configuration was held fixed; followed by the Module 3.0 regime study.
Observed result
Random reshuffling produced 2.26× boundary amplification: reported weakest-transition degradation divided by mean degradation was 0.0941 / 0.0416 ≈ 2.26 (Module 0 / COIT), not a 2.26× absolute loss of coherence. Training loss did not separate the conditions, and global embedding entropy remained identical to six decimals (10.972432). Controlled parameter changes produced conservative, balanced and stress-tolerant regimes.
Supported claim
Corpus ordering had a reproducible structural effect in the tested setup; distinct operating modes were selectable on the tested scaffold.
Boundary
The result is bounded to Protocol C and its low-data design. It does not establish a universal neural-model claim or show that contradiction structure is irrelevant.
Did observed internal representation structure reliably appear as admissible generated output?
Evidence basis
Corpus D / Module 5.0 encoder-geometry diagnostics, paired with the separate output probes.
Observed result
A two-tier representation hierarchy was reported in the multi-seed pair-by-variant probe: first-order 250/252 units (99.2%) and second-order 124/252 (49.2%). These are within-protocol observations, not independent experiments; exact scoring-unit construction awaits original-protocol confirmation. The stronger regime did not form a stable third tier. Anchor tokens, formula placeholders and contradiction drivers were absent from free generation across the tested regimes.
Supported claim
Representation diagnostics describe internal structure; they do not establish that structure will survive into generated output.
Boundary
The observations are bounded to Corpus D and the tested output probes. They do not establish general output failure rates or whole-document correctness.
Failure taxonomy, bounded repair and authority transition
Question
Which output failures were observed, what could bounded repair address, and who determines canonical identity?
Evidence basis
Corpus D2 failure analysis; Module 4.0 / ARC-03 repair probes; and the separate Bridge v2 Run03 proposal-to-Registry authority comparison.
Observed result
Corpus D2 distinguished complete loss, partial damage and identity-level substitution. Repair behavior was class-dependent: simple loss was more recoverable than partial damage, while identity-level substitution and complex S4 identity remained unresolved. These findings motivated narrower, span-local repair and the separation of proposal from external deterministic authority.
Supported claim
Failure classes can be made explicit and some cases can receive narrow, span-local repair. A neural proposal remains subordinate to external deterministic authority for canonical identity.
Boundary
Class counts and exhaustive coverage are not supplied. These findings do not establish general identity recovery, and repair does not define canonical truth.
What do the selected frozen and component evaluations show about structural inspection, context qualification and protected carriers?
Evidence basis
Each item below has its own configuration and denominator. They are not combined into one SIF 2.0 score.
Observed result
Frozen Unseen Test · Registry v2.0.3 / Positional EII v1.1: 35 rows, 28 non-negative and 7 negative controls; Anchor Identity Precision (AIP) 1.000. The per-instance anchor denominator is not stated / requires source confirmation.
Multi-anchor Bench · Registry v2.0.3 / benchmark v1.2: 30 rows and 104 anchor instances; primary structural metrics 1.000. FC-5 neighbor-corruption was not tested.
Context Gate · Negative Bench v2.4 / Context Gate v1.2: 30 adversarial sentences, comprising 23 MUST_PASS and 7 KNOWN_RISK. False activations changed from 9/23 before the gate to 0/23 after it, and were 0/7 in the other subset.
Formula Stress · Module 2.0 v2.0: 42 rows (36 protected, 6 null) and 73 carriers. Protected results were 36/36 rows and 73/73 carriers; null results were 6/6. The unprotected baseline was 18/36 rows and 48/73 carriers.
These separate tests support bounded claims about frozen structural inspection, context-gated intervention and literal carrier preservation under their named configurations.
Boundary
E05 AIP 1.000 is not universal or system accuracy; its per-instance anchor denominator remains unstated. Do not infer denominator-complete success, whole-document correctness, general semantic correctness or an aggregate SIF score. Formula Stress does not test mathematical equivalence, automatic carrier discovery or recovery of already-lost formulas. Do not retroactively relabel test Registry versions.
Architecture reuse and specification sufficiency are separate questions. The nuclear study met its frozen criteria with a specification extension; the psychiatry study did not meet its frozen overall criteria.
Boundary
This is not evidence of universal portability or clinical validation. The morphology-aware psychiatry result is diagnostic and does not replace the frozen strict criterion.
The engineering qualification reports 279 tests passed, a 12/12 final frozen gate and zero regressions. The original qualification logs were not recovered for independent website-audit verification. The professional pilot remains pending.
Supported claim
Internal engineering qualification reports bounded implementation results for operationalizing the frozen specification; the public eight-case replay is separate recorded demonstration evidence.
Boundary
Professional pilot evidence is pending. The 279-test and 12-gate source logs were not independently recovered in the website audit. These results do not establish professional-use, clinical or independent external validation, and archive closure is not claimed.
Evidence also includes limits and negative results
These findings are part of the evidence map and carry the same weight as supportive results.
Status and next layer
Evidence, publication and release status stay separate
Research / evidence status
Selected findings are summarised with their own tested scope and boundary.
Publication status
Pending site freeze confirmation. No publication state is inferred from the evidence summary.
Artifact / release status
The public Technical Evidence Register contains 17 selected records. Internal evidence and private source materials remain unpublished.
Technical Evidence Register
The public register presents selected records with their available configuration, denominator, version and status information. It does not publish the complete internal evidence base or private source materials.
An observed result supports only the claim, version and evaluated scope stated alongside it; this page is not a universal correctness score, certification or validation claim.