← Complete research archive
Language compilerClosed / no-go95 lines

R12 Referential Gather-Delete Permutation Executor Result

Decision: reject v1; do not generate confirmation

R12_REFERENTIAL_GATHER_DELETE_EXECUTOR_RESULT.mdOpen original Markdown ↗

R12 Referential Gather-Delete Permutation Executor Result

Decision: reject v1; do not generate confirmation

Bottom line

The typed permutation state, source-deletion boundary, and recurrent call are mechanically sound, but the v1 packet does not carry stable entity identity. The primary tied treatment reaches only 48.340% answers, 18.701% exact final assignments, and 17.236% both-transition exact. A favorable untied cell is tied at 48.438% answers, and a gold-pointer/kind training arm remains at 49.170%. No Stage-B promotion gate is close.

The failure is localized. Query and amount classification are 99.707% and 99.780%, respectively, while operation-entity matching is only 51.294%. The compiler almost always points inside the correct multi-token span, but v1 softmax gathering collapses each span to one contextual subtoken. Across the 16,384 operation-entity references on the fresh qualification board:

  • both selected positions lie inside their correct spans in 16,381/16,384 = 99.9817%;
  • selected operation/intro token IDs agree in only 9,720/16,384 = 59.3262%;
  • the complete operation/intro token-ID sequences agree in 16,384/16,384 = 100%.

Thus the next bounded repair is not more recurrence or capacity. It is a set-valued, vocabulary-aligned identity packet that preserves the entire selected span instead of sampling one contextual coordinate.

Frozen execution

All four jobs used commit d69250f, raw-300k, the qualified ordinary compiler, factorized train/development, seed 2026071901, one epoch, and zero confirmation access.

JobArmNodeElapsedExit
693111tied predicted packetevc282m17s0:0
693112untied predicted packetevc294m05s0:0
693113tied gold packetevc33 GPU 04m36s0:0
693114source-retained direct controlevc33 GPU 13m50s0:0

The two evc33 jobs used distinct H100 UUIDs. Base and compiler had zero trainable parameters. Treatment training contained 192,000 atomic operation targets and no full two-step state or answer target.

Development scores

Arm/interventionAnswerFinal assignmentBoth transitionsQueryEntity match
tied predicted48.340%18.701%17.236%99.707%51.294%
untied predicted48.438%17.480%16.113%99.707%51.221%
tied, no-fit gold rescore45.508%19.727%15.625%99.707%46.582%
tied trained/evaluated gold49.170%20.166%19.385%99.902%53.760%
tied operation shuffle31.104%11.084%5.078%99.707%36.157%
tied query shuffle43.066%18.701%17.236%67.188%51.294%
source-retained direct37.988%n/an/an/an/a

Tied per-surface answer accuracy is binding twin 56.836%, canonical 46.289%, order twin 41.602%, and paraphrase 48.633%. Only 24/512 quartets have all four answers correct; one quartet has all four final assignments exact and none has all four complete transition traces exact.

Frozen-gate assessment

The source-retained 95% ceiling fails. Every gold/treatment capability floor fails. Every per-surface/all-four floor fails. The operation and query shuffles fall below their 45% ceilings, but treatment margins are only 17.236 and 5.273 percentage points, not the required 40. The 2,047/2,048 intervention coverage passes. Tied remains within two points of untied and uses fewer parameters, but both are weak. Identity/custody gates pass.

The failure is downstream of pointer accuracy because gold packet training and rescoring do not rescue it. It is not caused by lack of a second cell because untied does not improve it. The source-retained decoder also fails its ceiling, so it is not a useful positive model for this representation.

Consequence

V1 is preserved as a conventional negative baseline. Do not tune it, add epochs, increase width, generate confirmation, or claim source-deleted reasoning. The one admissible repair changes only the packet identity channel:

  1. expose frozen vocabulary embedding states alongside contextual compiler memory;
  2. use normalized sigmoid role masks, which preserve all tokens selected by the compiler's existing multi-token role supervision;
  3. encode entity/literal/query identity from the complete vocabulary-aligned span, while retaining contextual operation-kind information separately;
  4. repeat the same atomic-only training, two-step evaluation, favorable arms, and causal shuffles under a new preregistration.

This repair is inspired by the measured failure, not a post-hoc rescue of v1's claim. It must freeze new source and gates before fitting.