Research · Structural ordering
Same data. Same model. Same loss. Different order.
The first major SIF research question asked whether representation behavior was controlled by the annotated magnitude of technical contradictions or by how the same low-data corpus was structurally organised.
Corpus C as a diagnostic scaffold
Corpus C contains 142 English-Ukrainian sentence pairs constructed for controlled representation diagnostics rather than scale-oriented translation training.
Its design includes single-domain and cross-domain pairs, multiaxial contradictions, fixed diagnostic chains, closed-loop rings, invariant references and separate adversarial NEG probes. The contradiction annotations are frozen structural descriptors used to construct and test the scaffold.
The corpus is therefore treated as a precision diagnostic instrument: the aim is to expose ordering sensitivity, stability boundaries and regime persistence under a governed protocol.
Magnitude hypothesis
The protocol first tested whether contradiction magnitude itself produced measurable representation change.
Systematic ablations included label shuffling, constant-magnitude controls and explicit null tests. Within the declared sensitivity bounds, contradiction magnitude was not detected as a driver of cross-lingual alignment, chain-level coherence, entropy diagnostics or translation performance under passive structural indexing conditions.
Finding: the magnitude hypothesis was not supported as the dominant factor under the fixed protocol.
Ordering intervention
The same micro-corpus was then treated as a deterministic structural scaffold. Model architecture, LoRA adapter, regularization, tokenizer, corpus content and seeds were held fixed while corpus sequencing was changed.
Random reshuffling degraded the weakest chain transitions more strongly than average degradation. The reported boundary amplification ratio compares weakest-transition degradation to mean degradation: Δ_min / Δ_mean = 0.0941 / 0.0416 ≈ 2.26 (Module 0 / Article 1 COIT diagnostics). It does not imply absolute coherence became 2.26 times smaller. Global embedding entropy remained identical to six decimal places across the compared conditions.
The observed ordering effect was reproducible under the fixed protocol even though the pointwise objective did not represent corpus sequence as an explicit variable.
Held fixed
Model architecture · LoRA adapter · regularization · tokenizer · corpus content · seeds
Changed
Corpus sequencing
Observed
Relative degradation: Δ_min / Δ_mean = 0.0941 / 0.0416 ≈ 2.26; global embedding entropy identical to six decimals.
Why Module 0 matters
Module 0 explains why the result is not paradoxical. A pointwise objective can remain invariant to permutation of its terms while the optimisation trajectory remains order-sensitive.
The architectural consequence is narrower than a claim about all neural models: the loss can be minimized without explicitly representing, certifying or enforcing corpus-level ordering relations.
Order-dependent structure must be treated as an explicit protocol-level variable above the pointwise loss.
COIT: the single-factor test
The Corpus Ordering Integrity Test isolates ordering as the manipulated variable under a frozen configuration. Its role is not to optimise translation quality, but to test whether structural sequencing changes representation behaviour when the other controlled variables remain fixed.
What this stage changed
The result shifted the research question away from passive contradiction magnitude and toward controlled structural organisation. The next question was whether the same scaffold could support multiple reproducible operating regimes when only regime-level parameters changed.