DIVERGE-HSC1 Support-Rank Diagnostic
Status: frozen read-only diagnostic; no HSC1 training or promotion standing
Decision date: 2026-08-05
Question
HSC1 is closed: its single Viterbi parse reaches 96.094% exact packets on the train distribution but only 8.594% under lexical shift and zero under renderer and composition shifts. This audit does not repair or rescore HSC1. It asks one question needed to choose a different successor architecture:
Does the frozen HSC1 checkpoint still assign enough structured posterior mass to the valid shifted interpretation that a compact packed support lattice could preserve it, or is the valid interpretation absent even from a broad bounded envelope?
Frozen assessor
The assessor loads the exact failed checkpoint with no gradients and regenerates the same four 256-episode cohorts at fresh seeds. Record boundaries remain the frozen learned boundaries. For each gold record it measures:
- exact rank of the gold monotonic phase triple among legal cut triples;
- exact rank of the gold cue-position/kind assignment;
- exact rank of each gold semantic template among all 128 templates, scored by the exact log-partition over every legal token alignment; and
- whether the gold token alignment is Viterbi within its gold template.
The candidate packed envelope is deliberately not a top-K list of complete parses. It retains the complete cut and cue lattices, K semantic templates per option, and every finite-state alignment inside each retained template. The audit reports packed dynamic-program cells separately from the corresponding unmaterialized Cartesian interpretation count.
This use of gold spans and labels is assessor-only. It measures whether the frozen scores contain a recoverable interpretation; it is not an executable compiler and has no capability standing.
Frozen decision
Evaluate K = 1,2,4,8,16,32,64,128.
- If some
K <= 64retains every gold fault-line interpretation in at least 95% of episodes in each lexical, renderer, and composition shift, a new support-lattice compiler is justified. That successor must encode each record once, retain uncertain record membership/cuts/templates as guarded variables, prove exact extensional parity on CPU, and compare packed execution against full-particle and single-Viterbi controls. - If no
K <= 64clears that floor, do not build the lattice successor from HSC1 scores. Replace the language interface/backbone or training task instead. K=128is reported only as an exhaustive fixed-grammar diagnostic. It cannot satisfy the compact-envelope gate.
No seed, width, duration, loss, cue, template, threshold, source-layer, or optimizer change is permitted. A positive rank audit would qualify a distinct architecture experiment, not reopen HSC1 or the failed broad DIVERGE claim.
Result
The frozen 1,024-episode audit passes the decision rule. It covers 6,144 raw records and 12,288 complete options with the exact failed checkpoint and no weight updates.
| Cohort | Cue top-1 | Template top-1 | Gold alignment Viterbi | Episode support K=1 | Episode support K=2 |
|---|---|---|---|---|---|
| Train | 100.000% | 100.000% | 100.000% | 100.000% | 100.000% |
| Lexical shift | 73.959% | 99.316% | 100.000% | 93.750% | 100.000% |
| Renderer shift | 67.128% | 99.802% | 100.000% | 97.656% | 100.000% |
| Composition shift | 13.845% | 100.000% | 100.000% | 100.000% | 100.000% |
Record segmentation is 100% in every cohort. Gold phase rank is at least 99.751% at K=1 and at least 99.934% at K=2. A K=2 packed envelope retains the valid fault-line interpretation in all 768 shifted episodes, while the hard all-component K=1/Viterbi path retains complete shifted episodes in only 10.156% lexical and 0% renderer/composition cases. The failure is therefore principally premature cue commitment, not missing option semantics.
For composition shift, K=2 stores 324,877 dynamic-program cells across 1,524 records. The corresponding unmaterialized Cartesian upper bound is 1,909,871,477,785,735,488 interpretations. This is precisely the regime where factorization is necessary rather than cosmetic.
Report SHA-256 is
d2fea259a9ef68f6bd7414de8364350ff68d028e7ecc7cb27ae769cb00fd437a.
The result authorizes DIVERGE-ULC1 as a distinct uncertainty-lifted runtime
gate. HSC1 remains closed and its exact-packet result is unchanged.