Results: Qwen3.5 acceptance v1/v2 (execution failures)

[atimics]

Stacked experiment result

This PR is stacked on the implementation branch behind upstream leCore PR #31. PR AnOversizedMooseWithSocks#31 holds the proposed 0.2.11 implementation; this PR holds the preregistration, terminal executions, ilxyr ledgers, evidence exports, permanent publication receipts, and committee discussion.

Terminal status

COMPLETE — v2 ilxyr outcome: execution_failure — no scientific acceptance/rejection result

The authorized v2 run is over. It will not be retried under the same identity, and no v3 run is authorized.

V2 corrected the v1 interpreter-path bug and completed the substantive CPU work:

1. exact source/model/corpus preflight passed;
2. all 12/12 ilxyr admission gates passed;
3. pre-install tokenizer and reference-logit gates passed;
4. the original checkpoint was evaluated across 4,097 input tokens;
5. the experimental installer emitted a 2,104.2 MB checkpoint;
6. leCore reloaded it and evaluated the same 4,097 input tokens; and
7. the required official Transformers checkpoint reload failed.

Transformers reported that the two prepended layers contain packed linear-attention tensors where the official model expects split projections, and that their MLP tensors use width 128 where the Qwen configuration declares 3,584:

linear_attn.in_proj_qkvz.weight | UNEXPECTED
linear_attn.in_proj_ba.weight   | UNEXPECTED
linear_attn.in_proj_a.weight    | MISSING
linear_attn.in_proj_z.weight    | MISSING
linear_attn.in_proj_qkv.weight  | MISSING
linear_attn.in_proj_b.weight    | MISSING
mlp.up_proj.weight              | MISMATCH: [128, 1024] vs [3584, 1024]
mlp.down_proj.weight            | MISMATCH: [1024, 128] vs [1024, 3584]
mlp.gate_proj.weight            | MISMATCH: [128, 1024] vs [3584, 1024]

This is the exact interoperability failure the mandatory reload gate was intended to expose. The leCore runtime accepts its own heterogeneous representation; the official Transformers architecture does not.

Because the exception happened before the runner's atomic metric envelope was emitted, ilxyr correctly recorded execution_failure with an empty metric set. Partial internal observations are not promoted into preregistered metrics. Official text generation and the vision-input smoke did not run.

V2 result and permanent evidence

• Committed result: v2-20c3330d0b3e-execution-failure
• Permanent Arweave root: f35OgXuOoulk9EcndJBNWIgZ5tmaQ6DRzeINP7ARZTU
• Canonical publication manifest SHA-256: 21874555c7f9e73d20b9d11599c79236f8f40267ee975f703ffc053eb6b1f9ab
• Publication receipt: publication-receipt.json
• ilxyr evidence ref: artifact://sha256/15885a3e5fbb01a504473e3076c25c9d5d938c13f96faa340ed9600a245845cc
• ilxyr run ref: artifact://sha256/dee0b19888937eb4b4109ef6558ce3e5c6a530469019b94f4905c01d144a6d8f
• Verified ledger: 15 objects, 16 events, valid: true
• Arweave verification: all 47 transactions confirmed at 0 winc; publication-manifest.json and ledger/events.jsonl were downloaded through the public path manifest and matched their committed SHA-256 values
• Original AWS evidence bundle: s3://zero-training-022118847419/qwen35-acceptance/2f32151-2fc06364-a6c9d113/v2-formal/qwen35-evidence.tgz, SHA-256 df2ba31608f9039e9528a43a93d4766c8e8feb023668a1b9ffc1c85553210c1c

The r7i.4xlarge formal execution ran from 00:26:36.737Z to 04:51:28.293Z (4h 24m 51.556s). Estimated on-demand compute is $4.71, plus a few cents of storage. The instance self-terminated, its encrypted root volume was deleted, and the temporary zero-ingress security group was removed. The total remained well below the $10 ceiling.

Execution history

Attempt	Instance	Outcome	Scientific result?
preflight-001	i-0b49b8662723706f0	Rejected a mistyped nonexistent source SHA	No formal run
v1	i-0c549643b970e896a	Wrong Python interpreter; execution_failure before model evaluation	No
v2	i-059aab8bf1d3d87c9	Official Transformers rejected the emitted layer shapes/names after paired evaluation	No

Frozen v2 identity

• Experiment: lecore.qwen35.install.20c3330d0b3e.d6905f043e78.a6c9d1135a04.2f32151c6d44.v2.acceptance
• leCore source: 2f32151c6d44c80a7e2ff98c0696650cd1d9c6f1
• ilxyr base identity: e92382ff2a5e8714466533a160f6609b4ef9cee8
• Model: Qwen/Qwen3.5-0.8B@2fc06364715b967f1860aea9cf38778875588b17
• Installation corpus SHA-256: d6905f043e7856b93b2dd72dac5fa0dc593898c55d6c54c51f3153f4317d6b7f
• Held-out evaluation SHA-256: a6c9d1135a04d10955fe11d210b7f642e1c2341d4f2c8369b9a832cc97839d94
• Spectral filtering: disabled
• Installer: explicitly experimental

Gate record

Gate	V2 record
Frozen input preflight	Pass
ilxyr admission	Pass (12/12)
Tokenizer/reference-logit pre-install gate	Pass by execution ordering; exact partial values not promoted
Original paired stream	Completed (4,097 input tokens)
Installed paired stream	Completed (4,097 input tokens)
Exact metric envelope / statistical result	Not emitted
leCore checkpoint reload	Pass
Official Transformers checkpoint reload	Fail
Official text generation	Not reached
Official vision-input smoke	Not reached
ilxyr ledger verification	Pass
Permanent publication and public hash check	Pass

Validation

• Focused Qwen and selector suite: 22 passed.
• Every committed JSON document parses.
• The AWS evidence archive matches its signed SHA-256 sidecar.
• The ilxyr ledger independently verifies all 15 objects and 16 events.
• Native, RO-Crate, and in-toto exports were generated from a read-only isolated ilxyr snapshot.
• The canonical publication manifest rebuilds deterministically with SHA-256 21874555…f9ab.

Committee decision requested

Accept this PR as the terminal evidence record. Keep the Qwen loader, tokenizer, configuration parsing, text runtime, and diagnostics in the 0.2.11 review. Keep spectral filtering research-only and keep the layer-prepending installer experimental; the official-reload failure is direct evidence that it is not ready for promotion.

Any future run should require a new implementation that passes an official Transformers reload fixture before seeking explicit authorization for a new experiment identity.

PR AnOversizedMooseWithSocks#30 remains the separate long-term liblecore track.