Results: Qwen3.5 acceptance v1-v7 (v7 native GO)

[atimics]

Stacked experiment record

This PR is stacked on upstream implementation PR
#31. PR AnOversizedMooseWithSocks#31 holds
the proposed 0.2.11 implementation; this PR holds the preregistration,
terminal executions, ilxyr ledgers, permanent publication receipts, and
committee-facing result artifacts.

Terminal result: v7 accepted — GO

V7 is the final native-qualification run, frozen from implementation commit
cb3b1d2ac71c183bf9307ca7145a2a619ff30c30. Ilxyr resolved the one formal
execution as accepted, recorded exactly one EvidenceRecorded event, settled
both forecasts, and independently verified all 16 ledger events and 15 objects.

• One-page PDF result
• Committed v7 project, result, ledger, and receipt
• Permanent Arweave evidence root

Gate or metric	V7 result
Paired evaluation positions	4,096
Original perplexity	7.906224720549084
Installed perplexity	7.906224720549084
Perplexity delta	0%
Paired 95% block interval	[0, 0] nats
Reference-logit relative error	6.787656774900957e-7
Reload / text / vision	pass / pass / pass
Native C GDN	active in both evaluators
Peak RSS	19.05 GiB
Ilxyr resolution	accepted — GO

Spectral filtering remained disabled, the experimental installer was used, and
the emitted checkpoint reloaded through both leCore and official Transformers
paths.

Native C qualification

Both isolated evaluators compiled and used the C recurrence and completed
fresh- and resumed-state parity checks:

Evaluator	Native calls	Native tokens	Fallback / direct NumPy
Original	180	74,124	0 / 0
Installed	200	82,360	0 / 0

Every recorded native float32/float64 parity row passed. The emitted library
SHA-256 is
bafe42afd0ac8aa136b3ef6c68c35dc90fbfdd5c1cbba046100da5e2817bc075.
This closes the v6 qualification: v6 proved correctness under safe fallback;
v7 proves the repaired Linux compiler path and parity-gated native execution.

Performance, cost, and cleanup

The installer took 8,056.60 seconds (about 2h 14m) and remains the dominant
bottleneck. Parallel paired evaluation took 525.35 seconds (about 8m 45s),
using the shared 512-token schedule. Peak RSS was 19,503.84 MiB, supporting a
measured future 32 GiB right-sizing study.

Estimated v7 compute is approximately $1.32. The $6.86 launch-tracker
figure is the v4-v7 phase subtotal, not all-history spend: the published
estimates for v2-v7 total approximately $14.23, plus the unestimated v1
attempt and a few cents of storage. The proposal explicitly says its $10
ceiling buys one formal attempt. Instance i-096225051167ef900 is terminated.
Root volume
vol-087fdcb57d5c79c09 and security group sg-0a11cb364009ae8d7 are deleted.
No retry or additional instance was launched.

Permanent publication checks:

• publication manifest SHA-256: f97e75341f5913dfd9b91663dd8e2db9bd67862b79339a5df12e2a101a88d10d;
• ledger/events SHA-256: b558cf918abc393f50c2c45b16d214d56aa8245b831c33a85b4f0d47fd41e88e;
• reconstructed evidence SHA-256: 2871b7fa96685bd7f5d75f79368808d00649f9cf91950f4655e6099266319f8f;
• 95 Arweave transactions confirmed at 0 winc; and
• public cache and canonical gateway reconstructions verified; model weights excluded.

Execution history

Attempt	Outcome	Scientific result?	Principal finding
v1	execution_failure	No	Virtual-environment interpreter was not preserved
v2	execution_failure	No	Official Transformers rejected emitted checkpoint layout
v3	execution_failure	No	Diagnostics polluted strict JSON ingestion
v4	preflight failures	No	Linux fixture and ilxyr project-schema gaps
v5	execution_failure	No	Final output contained non-schema source.emitted_checkpoint
v6	accepted	GO	All correctness gates passed; native C safely fell back
v7	accepted	GO	Correctness repeated and native C qualified with no fallback

Committee decision requested

Review v7 as the complete acceptance record for the proposed 0.2.11 scope.
Retain the Qwen loader, tokenizer, configuration, text runtime, diagnostics,
and parity-gated optional native recurrence. Keep spectral filtering
research-only. Keep checkpoint-changing layer prepending experimental until the
maintainers and committee decide whether this accepted evidence is sufficient
for a broader release gate.