← Complete research archive
Evaluation & auditsClosed / no-go77 lines

DIVERGE-RRG1 Result: Query Semantics Transfer, Shared Evidence Does Not

- implementation commits: 8246fcb, f23d032; - H100 job: 744379 on evc22; - schedule: exactly 2,000 updates, 128 EVIDENCE plus 128 QUERY rows per update; - trainable/total parameters: 733,249 / 1,196,434; - fit elapsed: 85.681 seconds; - checkpoint SHA-256: 05c7b6c7ee68e14133231d5…

docs/research/DIVERGE_RRG1_RELATIONAL_ROLE_GROUNDER_RESULT.mdOpen original Markdown ↗

DIVERGE-RRG1 Result: Query Semantics Transfer, Shared Evidence Does Not

Status: closed negative after one frozen fit, one opened-board admission, and one read-only attribution. Confirmation was never opened.

Frozen execution

  • implementation commits: 8246fcb, f23d032;
  • H100 job: 744379 on evc22;
  • schedule: exactly 2,000 updates, 128 EVIDENCE plus 128 QUERY rows per update;
  • trainable/total parameters: 733,249 / 1,196,434;
  • fit elapsed: 85.681 seconds;
  • checkpoint SHA-256: 05c7b6c7ee68e14133231d572b8c66aced5e3a72a6ba25827498560f723336ae;
  • training-report SHA-256: 0e0b681f3bd65c876fbdbc094de3385de6369352f934fac75ebf540c2cff57db.

The immutable corpus contains 100,000 EVIDENCE and 100,000 QUERY rows. Every one of 50,000 semantic items per stage has TARGET-first and DISTRACTOR-first realizations. Every family/form/order cell is exact and all overlap receipts are zero. Data/report SHA-256 values are 56f0e6e5...d67fe, 2d325c86...78c1c, and 7439147d...36b3.

Training reaches 100,000/100,000 exact in both stages and 50,000/50,000 exact counterfactual pairs in both stages. WORLD and numeric-EVIDENCE owner hashes remain bit-identical.

Development result

The opened SRP1 board rejects promotion:

GateResultFloor
QUERY768/768765
EVIDENCE2,048/3,0723,070
Fully sealed episodes0/256255
WORLD256/256256

Every QUERY mode is 256/256 and every QUERY renderer is 128/128. Every episode fails sealing because exactly one of the three evidence renderers is inverted. Development evaluation SHA-256 is 6ca42dabb4da119e26d7159c340750a4f01161efe65d634ceba0b2f865fd96db.

Read-only attribution

Diagnostic commit 7038868 changes no weight, threshold, source, split, or prediction. It scores the frozen checkpoint on every established NVE1 training and confirmation surface:

SurfaceExact
confirmation renderer 00/3,072
confirmation renderers 1/23,072/3,072 each
legacy training renderer 00/3,072
legacy training renderers 1--53,072/3,072 each

Renderer 0's mean signed margin is -59.806 on confirmation and -2.774 on the legacy training surface. All passing surfaces have positive margins. Arbitrary entity renaming changes zero assignments and produces bit-identical logits with maximum absolute difference 0.0. Owner hashes remain exact. Diagnostic SHA-256 is 0095b8ca08ed14e2e26534b00eb6da0a4fb1ef10305c1514d458f5a91e865409.

Decision

RRG1 disproves the universal shared REFERENT owner. The representation removes identity and order shortcuts and produces a qualified natural QUERY owner, but EVIDENCE semantics remain stage/syntax dependent. Do not run RRG1 width, duration, seed, grammar, renderer, loss, marker, or optimizer variants.

Preserve the RRG1 QUERY owner as a qualified component. The structurally different successor is STI1: immutable NVE1 owns all EVIDENCE bindings and the immutable RRG1 owner handles QUERY, joined only by typed transactions. STI1 uses no new learning and must independently pass development before opening confirmation.