← Complete research archive
Evaluation & auditsClosed / no-go76 lines

DIVERGE-SRP1: Semantic Referent Primitive Result

Newton job 744234 completed exactly 1,000 joint REFERENT updates and fit both immutable training sets at 50,000/50,000. Evaluation job 744246 completed normally on evc32 in 00:16:04.

docs/research/DIVERGE_SRP1_SEMANTIC_REFERENT_PRIMITIVE_RESULT.mdOpen original Markdown ↗

DIVERGE-SRP1: Semantic Referent Primitive Result

Status: closed negative after the one frozen training and evaluation pass. No SRP1 width, duration, seed, renderer, optimizer, warm-start, or loss variant is authorized.

Result

Newton job 744234 completed exactly 1,000 joint REFERENT updates and fit both immutable training sets at 50,000/50,000. Evaluation job 744246 completed normally on evc32 in 00:16:04.

The fresh confirmation result is:

  • WORLD programs: 256/256;
  • natural EVIDENCE receipts: 3,067/3,072;
  • fully sealed episodes: 253/256;
  • natural QUERY transactions: 753/768;
  • sensitive answers: 248/256;
  • exact extensional answers: 253/256;
  • no-evidence abstention: 251/256;
  • invariant answers: 248/256;
  • partial-evidence underdetermined abstention: 249/256;
  • protected TOL3: 1,024/1,024;
  • protected NVE1: 256/256; and
  • invalid queries accepted: 0.

Query transactions by renderer are 128/128, 127/128, 128/128, 114/128, 128/128, and 128/128. Renderer 3 therefore misses the frozen 122/128 floor.

Matched owner comparison

Frozen SOT1 is much stronger on this deconfounded board than on its original confirmation board:

OwnerExact query transactions
Frozen SOT1742/768
SRP1753/768
Delta+11/768 (+1.43 points)

SRP1 needed at least +77/768 over SOT1. It misses that causal inclusion gate by 66 transactions. The semantic primitive improves the weak renderer from 102/128 to 114/128, but the gain is too small and does not meet the renderer floor.

Decision

SRP1 fails the conjunctive gate. It also misses full evidence sealing, extensional parity, universal packet/query swap rejection, and post-seal poison invariance because three episodes never seal. The shared REFERENT owner is not a qualified raw-language interface.

Consequences:

  1. close SRP1 without local retries;
  2. do not launch raw-language branch-local plasticity through SRP1;
  3. preserve TOL3 and NVE1 as protected stage-specific owners; and
  4. allow DIVERGE-PL1 only as an explicitly oracle-typed plasticity-mechanics ceiling while referent/source compilation remains the active bottleneck.

Receipts

  • checkpoint SHA-256: 0e4365d96d86ea791b64e8bc60b46e4441acfc3d2bc043bd9339da3e4bb286cd;
  • model-state SHA-256: 4b54f333394c869a4ac7a8286300176167d7d4e1ddae1cb86e1afcf5a3112182;
  • fresh board SHA-256: 8d5bf36e5eab0d51328a1fbfeabad676efcbe697fdbe1dd64cbe592820aa0216;
  • evaluation SHA-256: 8b68d19bbf67007addc6c9e9ec5287414580b4cd9b108198cb199fd171257ebe;
  • immutable runtime archive SHA-256: 502802d0fca0f33ebc005503cfd7867e7cfc2a67a0ddc68a2e1cbe8c559eaa6c.