Output Structural Verification
Does the produced result satisfy what is currently governed?
SIF 1.0 SIF 2.0 SIF Inspector
Research programme
A research programme built by successive falsification.
SIF did not begin as a finished architecture. It emerged through a sequence of controlled questions about what could and could not preserve critical structure in low-data English-to-Ukrainian technical translation.
Each stage was motivated by a limit exposed by the previous one: preserving tokens did not guarantee stable structural relations; stable representation regimes did not guarantee admissible output; bounded repair did not establish canonical authority; and deterministic output inspection did not guarantee that the governing representation itself remained valid.
Research lineage
The research lineage
- FoundationTRIZ-Aware Loss v1.0 · Corpus AInvariant-preserving foundations
- Semantic reinforcementTRIZ-Aware Loss v2.0 · Corpus BAnchors, bridge pairs and geometric diagnosticsDemo Module 2.0 — reproducible proof of concept
- Observed limit
Preserving individual carriers and reinforcing anchor geometry did not establish the structural ordering of relations around them.
- Layer 1Structural ordering
- Layer 2Behavioral regimes
- Layer 3Representation diagnostics
- BoundaryRepresentation / output gap
- Layer 4Output failure taxonomy
- Layer 5Bounded structural repair
- Architectural transitionExternal decision authority
- SIF 1.0Terminology admissibility
- SIF 2.0Structural admissibility
- TransferDomain transfer / specification sufficiency
- Assurance boundaryGoverning-State Assurance
1Contradiction-structured neural translation experimentsExperiment and observed limitation
2Pointwise objective and corpus-order limitsExperiment and observed limitation
3Representation-to-output transfer limitsExperiment and observed limitation
4Bounded neural repair and identity failuresExperiment and observed limitation
6SIF 1.0: terminology → SIF 2.0: heterogeneous structuresDistinct bounded implementation stages
Foundations before the five-layer diagnostic pipeline
The experimental lineage began with TRIZ-Aware Loss v1.0 on Corpus A, where explicit loss terms and protected representations targeted invariant preservation. TRIZ-Aware Loss v2.0 extended this foundation through Corpus B with semantic anchors, bridge pairs and geometric regularization. Demo Module 2.0 provided a bounded reproducible proof of concept for this reinforcement stage. These experiments established that explicit preservation was tractable, but did not yet establish the ordering and interaction of structural relations.
The early work used curated low-data metrology corpora to protect placeholders, formulas and other critical technical structures. The remaining problem was relational: preserving individual carriers did not by itself control the structural relations around them.
Module 0 then supplied the theoretical premise for the next stage. Under a fully frozen configuration, changing only corpus order altered chain-level representation coherence while the pointwise training objective did not separate the conditions. The conclusion was architectural: order-dependent structure must be governed above a pointwise loss rather than assumed to be represented by it.
The five-layer diagnostic pipeline
| Layer | Public research question | Primary component | What the stage established |
|---|---|---|---|
| 1 · Structural ordering | What controls representation behavior under a frozen low-data protocol? | Corpus C · Ordering Dominance · COIT | Contradiction magnitude was not confirmed as the dominant factor; corpus ordering produced reproducible structural effects. |
| 2 · Behavioral regimes | Can distinct operating modes be selected without changing corpus structure? | Tunable Operating Regimes · Module 3.0 | Controlled parameter changes produced reproducible conservative, balanced and stress-tolerant regimes. |
| 3 · Representation diagnostics | What structure is actually expressed inside the model? | Corpus D · Module 5.0 | A two-tier representation hierarchy was observed; the stronger regime did not form a stable third representation tier. |
| Boundary · Representation → output | Does representation structure reliably survive into generated output? | Cross-layer finding | No. Representation-level structure did not reliably imply output-level structural admissibility. |
| 4 · Output failure taxonomy | What actually fails in generated output? | Corpus D2 | Observable structural invariant failures became the unit of analysis. |
| 5 · Bounded structural repair | Which failures can be repaired without unrestricted rewriting? | Module 4.0 | Class-aware local repair was possible for some failures, but repair could not define canonical authority. |
The architectural transition
The repair experiments exposed a decisive limit: a model can propose a replacement, but it cannot by itself establish which form is normatively admissible.
That moved the research from model-internal behavior to external decision authority. SIF 1.0 introduced Registry-governed terminology admissibility. SIF 2.0 then extended the controlled surface from terminology identity to heterogeneous structural invariants, including protected technical carriers and context-qualified intervention.
Version-label note: “(v1.0)” in the April 15, 2026 archived Architecture and Diagnostic Pipeline title identifies that five-layer research report. The separate June 17, 2026 SIF 1.0: A Registry-Based Framework record describes the later Registry-governed terminology-admissibility study. The matching version notation alone does not make their experiments or claimed scope identical; the original titles and dates remain unchanged.
The current research programme
The current SIF programme now contains two distinct assurance questions.
Governing-State Assurance
Does what currently governs still have valid support?
SIF Governance representation fidelity source-fidelity assurance expert confirmation
The output-verification line exposed a second assurance problem: the governing representation itself can become incomplete, unsupported or out of date. Governing-State Assurance addresses that question.
How TRIZ is used
TRIZ is not treated as an independently validated causal mechanism inside the neural model. In the early research it served as a structured design heuristic for contradiction-oriented corpus construction and diagnostic questions.
Later representation findings were derived empirically from the frozen protocol rather than attributed to TRIZ semantics. At the intervention stage, TRIZ can help propose candidate transformations; SIF qualifies whether a proposed transformation is structurally admissible.
TRIZ proposes. The system qualifies.
Negative results are part of the architecture
- Contradiction magnitude was not confirmed as the dominant driver under the fixed Corpus C protocol.
- Representation-level structure did not reliably transfer to decoder output.
- Neural repair could recover form without recovering canonical identity.
- Psychiatry reused the architecture without modification but did not meet the frozen overall transfer criteria.
Research pages
The deeper pages follow the causal sequence rather than internal artifact numbering.
- TRIZ in the Development of SIFContradictions as experimental probes and architectural constraints.
- Structural OrderingCorpus C, Ordering Dominance, COIT and the pointwise-loss limit.
- Behavioral Regimes & RepresentationControlled operating regimes and Corpus D diagnostics.
- Representation Output GapWhy internal structure did not settle output admissibility.
- Failure Taxonomy & Bounded RepairCorpus D2 and Module 4.0.
- From Repair to External AuthorityThe transition to Registry-governed deterministic control.
- Domain Transfer StudiesNuclear/radiation safety and psychiatry as paired transfer evidence.