Research · Behavioral regimes
One structural scaffold, multiple reproducible operating regimes.
After ordering became the dominant structural condition under the fixed Corpus C protocol, the next question was whether model behaviour would collapse to one stable state or could be selected through controlled regime-level parameters.
What was held fixed
Corpus composition and ordering, contradiction annotations, model architecture, tokenizer, loss formulation and diagnostic suite were held constant. The fixed seeds were 42, 123 and 456.
What was changed
Only regularization strength, sequence-closure penalty and the timing of NEG exposure were varied.
What changed across regimes
All evaluated regimes preserved global embedding invariants, including complete entropy invariance, but differed systematically in sentence-level self-consistency, robustness under NEG exposure and history-dependent persistence during passive zero-regularization evaluation.
Within the fixed low-data scaffold, regime parameters therefore selected reproducible operating behaviour without changing the corpus structure or model architecture.
Finding: regime differentiation was controllable, but global invariants remained stable.
| Regime | Operational character | Public interpretation |
|---|---|---|
| Conservative | Low regularization / no NEG exposure | Maximal invariant preservation. |
| Balanced | Canonical regularization / controlled NEG exposure | Controlled local adaptation. |
| Stress-Tolerant | Higher regularization / controlled NEG exposure | Boundary probing with bounded sensitivity. |
Why representation diagnostics were needed
Controlled regimes established that model behaviour could be selected. They did not yet show what structure the encoder actually represented under contradiction stress.
Corpus D and Module 5.0 therefore moved from regime behaviour to representation structure.
Corpus D: representation-level probing
Corpus D was used as a representation-diagnostic probe rather than an output-failure corpus.
Across the canonical multi-seed probe, the central finding was a two-tier hierarchy: the stressed configurations separated robustly from the Conservative regime, but the stronger stressed configuration did not form a stable third representation tier.
First-order dominance — {ST-CANON, ST-STRONG} over Conservative — was reported in 250/252 non-negative pair-by-variant observation units (99.2%). The second-order ST-STRONG over ST-CANON split was reported in 124/252 of those units (49.2%), indistinguishable from chance under the declared analysis.
These within-protocol observations are not 252 independent experiments. Exact construction of the 252 units remains subject to confirmation from the original scoring protocol.
First-order dominance
{ST-CANON, ST-STRONG} over Conservative
250/252 · 99.2%Second-order split
ST-STRONG over ST-CANON
124/252 · 49.2%Indistinguishable from chance under the declared analysis.What the two-tier result means
The encoder expressed a robust first-order structural signal but did not preserve a stable graded distinction between the two stressed regimes. This is a representation limit within the tested protocol, not evidence that the regimes were behaviourally identical.
What this stage could not establish
A stable internal representation signal is not the same as a structurally admissible generated output.
The next stage therefore tested the transfer boundary between representation and decoder output.