DIVERGE-CCR1 Result: Candidate-Relative Encoding Is Not Semantic Grounding
Status: closed negative at the opened development gate. Sealed confirmation was never evaluated.
Frozen execution
- implementation commits:
ec6d2f3, then pre-result audit-alias correctionc2ee33f; - training job:
744334on one Newton H100; - development evaluation:
744340; - read-only attribution:
744352; - training seed/updates:
2026080623, exactly 1,000; - trainable/total parameters:
546,433 / 1,009,618; - charged source bytes:
21,349,174in56.415training seconds; - checkpoint SHA-256:
dd5287181308b3c8aae3a21cba8a12f873d5294f63158313a4bd6eececa3dc1d; - training-report SHA-256:
2695cd21a8af40cc914e721f7e23d08c13cb116a4dcd359ba9b96c1bf8f1997b.
All 50,000 evidence and 50,000 query training assignments fit exactly. WORLD and numeric-EVIDENCE owner hashes remained bit-identical.
The fresh 256-episode confirmation board was generated before training from
commit ec6d2f3. It contains 3,072 evidence items, 768 balanced queries, and
1,048,576 represented worlds. Source, query, identity, and entity overlap
with training and SRP1 are all zero. Board/report SHA-256 values are
29923706...1ebf7 and 77ce453b...042d. Because development failed, this
board remained unopened.
Development result
CCR1 reaches:
| Measure | Result | Required |
|---|---|---|
| Evidence assignment | 2,097 / 3,072 | >= 3,070 |
| Query assignment | 695 / 768 | >= 765 |
| Fully sealed episodes | 0 / 256 | >= 255 |
Query renderer exact counts are 55, 128, 128, 128, 128, 128 of 128.
Development evaluation SHA-256 is
c87a51ed08a103a73a63f6446554d9a6cea57dee3fe2fb975a785ab92cf85d5f.
The evaluation wrapper exited nonzero, so dependency-held confirmation job
744343 became unsatisfiable and was canceled.
Read-only attribution
The one allowed attribution crosses every development item with every held renderer:
- EVIDENCE renderer 0:
159/3,072; renderers 1 and 2:3,072/3,072each; - QUERY renderer 0:
351/768; renderers 1--5:768/768each; - marker swap changes query accuracy only from
695/768to683/768, so the SELF/OTHER marker identity is not the causal decision variable; - canonical entity renaming changes 122 of 3,840 assignments;
- renderer-0 signed margins are negative on average (
-16.581evidence,-0.834query), while every passing renderer has a large positive margin.
Diagnostic SHA-256 is
86eae09bde4123ef0edf4552753d5dc5dc5f97dfa170e61ffbbacf46336b7178.
An independent data audit exposes the identifiability failure. In both 50,000-row source sets, every renderer has exactly one role order: even renderers always label first mention TARGET and odd renderers always label it DISTRACTOR. A model can therefore fit every training label by recognizing renderer/order without learning the semantic relation.
Decision
Close CCR1 without width, marker, duration, seed, renderer, warm-start, loss, or optimizer variants. Independent candidate encoding does not force candidate-relative semantics and retains mention-length leakage. It does not qualify raw-language PL1.
The successor must remove the confound before fitting and change the representation boundary: collapse every full mention to one anonymous, length-free token; encode one sentence once; jointly match two explicit role slots to the two mention contexts; and train on paired role orders within each lexical family. That successor is DIVERGE-RRG1.