Qwen → HDC.
All 64 blocks.
Both quality gates pass.

The latest verified steps.

This log grows as meaningful results are checked. Preparation milestones do not count as new model-quality measurements.

1. Verified · both held-out gates pass

   8 September 2026 · 01:47 UTC

   The full 64-block chain preserves held-out token choices

   Known-test: 491 / 512 matches (95.8984%), mean KL 0.00594440. Fresh article-disjoint test: 125 / 128 (97.65625%), mean KL 0.00601305. Each independently passes the predeclared ≥95% agreement and ≤0.05 mean-KL thresholds. The frozen model was not fitted or selected using these scores.

   All 64 replacement blocks ran in sequence, without a native transformer tail or teacher-state resets. Embedding, final RMSNorm, and vocabulary head remained native. The run completed at 01:44 UTC; final candidate states were archived before scoring.

   The independent audit checks final state hashes and Q24 decoding, source and frozen-payload bindings, selected original endpoints, hidden-state errors, and score-row aggregates. A separate recovery audit verifies all 2,811 published files. It does not independently recompute logits, replay the full chain, or locally rehash the large base payloads.

   The instance is terminated and both scoped controllers retired. Cumulative compute estimate: $23.02 of the same $50 cap, not an invoice. This is a short-context teacher-fidelity result, not task accuracy, generation quality, compression, speedup, or successful fusion.

2. Verified · execution freeze

   7 September 2026 · 20:45 UTC

   All 64 bases rebuilt; held-out testing begins

   All 64 rebuilt base receipts match the frozen originals. The independent recovery audit also confirms that all 64 correction archives are byte-exact, all 107,186 frozen payload records are unchanged, and all 276 original source files are preserved. The five approved continuation and output-head files are the only source additions.

   The first four blocks reproduce their saved integer fixtures exactly: 163,840 values checked, zero mismatches. The execution freeze binds the selected original teacher endpoints and the explicit reference policy. Known-test workers have now started after that freeze, with no refitting.

   This establishes a verified continuation checkpoint, not all-64 held-out quality. The host compares base receipts, not the large base payload bytes; it does not independently replay the full chain or recompute head logits. Estimated cumulative compute is approximately $16.54 under the same guarded $50 cap, not an invoice.

3. Verified · continuation milestone

   7 September 2026 · 20:20 UTC

   Past the reference stop; original bases are rebuilding

   The A10’s ordinary head reproduces the earlier saved-state report byte for byte: 124 / 128 agreement; mean KL 0.00751416. This repeats the validation diagnostic; it is not a fresh student trajectory.

   A fresh read-only audit rehashes all 352 uploaded inputs and verifies all eight selected native endpoints against the original archives. Fresh initial embeddings remain exact; the differing regenerated validation endpoint is preserved separately. 22 of 64 rebuilt base receipts match the original receipts retained locally. The guest checks base payloads; this host audit compares receipts, not the large payload bytes.

   One guarded A10 is active within the same cumulative $50 cap. Estimated compute through this check is approximately $16.00, not an invoice. No all-64 held-out result yet.

4. Verified locally · policy versioned

   7 September 2026 · 19:49 UTC

   Keep the original teacher targets fixed

   The explicit frozen-native-endpoints-v1 policy selects archived original h0 and h64 for every split. Fresh h0 must match exactly. Fresh h64 and intermediate native states are diagnostics only; all fresh endpoints are retained.

   All 55 focused tests pass, including a replay of the actual previous mismatch and rejection of a corrupt selected target. The historical strict-reproduction test still fails on that fixture: selection is not successful reproduction. The model, corrections, original 276 source files, corpora, no-refit rule, exact base gates, and held-out thresholds stay fixed. A single frozen continuation is launched.

5. Historical · run closed

   7 September 2026 · 16:38 UTC

   Head repair succeeds; native-reference equality fails

   The A10 successfully scores the saved all-64 validation states. The following strict gate stops on the last eight positions of one regenerated validation sequence, before rebuilding bases or executing held-out student inputs. The failed run is recovered and closed; all eight allocations up to this point are absent. Cumulative compute estimate: $15.33. The later entries above describe the authorized continuation.

Full-depth composition holds
on both held-out tests.

We have executed a chain of 64 HDC replacement blocks on both predetermined held-out sets. Each block receives the preceding replacement’s output; there is no native transformer tail and no reset to the teacher’s intermediate state.

The frozen replacement chain passes both original quality gates when scored through Qwen’s ordinary native output stage on an NVIDIA A10. No corrections were refitted after seeing these results.

Agreement with a teacher is not task accuracy, factual correctness, or free-running generation quality. These are 80 short-context sequences, at most 64 tokens each; broader generalization remains to be tested.

Historical comparators: the first-four replacement test matched 493 / 512 choices (96.29%, mean KL 0.00448743), with 60 native blocks remaining. The all-64 saved-state validation diagnostic matched 124 / 128 (96.875%, mean KL 0.00751416), on data used for correction selection. Neither is a substitute for the new held-out results.

All transformer blocks replaced.
Not every model component.

The source is the pinned mlx-community/Qwen3.8-27B-4bit checkpoint: a 5,120-dimensional residual stream, with 48 linear-attention / GatedDeltaNet blocks and 16 full-attention blocks. We preserve the learned information and causal structure while changing the arithmetic used by the replacement operators.

The evaluated composition

Text transformer scope

Tokens → embeddingNative · unchanged

→

HDC blocks 0–63Compiled base + learned corrections

→

Final norm + headNative · unchanged

→

Token distributionCompared with native Qwen

64 / 64 replacement coverage · zero-based block indices

48 linear attention16 full attention

00010203040506070809101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263

The artifact is a reconstruction package, not a ready-to-ship model. All 64 correction archives, source bindings, saved calibration and final held-out states, scores, and frozen payload manifests are preserved. The large compiled base payloads were not retained locally. They were rebuilt on the A10 and passed the guest’s exact payload gate; the host audit confirms matching original receipts. The cloud instance has now been terminated.

Recompile the base.
Learn the residual error.

A compiled block carries most of the original transformation. A smaller, SVD-guided response mesh learns to correct its error. Importantly, we fit against the states produced by the actual preceding student blocks, so calibration sees accumulated drift.

Inside a replacement block

Causal input → corrected output

Current student state x_t + causal historyNo teacher-state reset between blocks

↓

Compiled base blockInherited operators → integer / population arithmetic

↓

Base output b_tSent to both the correction encoder and the final sum

↓

Encode [x_t; b_t]Training-activation SVD coordinates

↓

Linear + ReLU + prefix meanLearned mixer → residual-SVD decoder

↓   Next student state = b_t + learned correctionThis output becomes the next block’s input

1. Integer arithmetic—not float bit strings

A scalar is encoded as a count plus a group scale. The main projection representation uses P = 2,046, 11 packed count bits per coordinate, and groups of 64 coordinates. Residual interfaces use Q24 fixed point. This is multibit numerical coding, not a one-bit sign vector.

For a group G with n coordinates, weight counts a_i and activation counts c_i give an exact integer accumulator:

A_b and C_d are packed bit planes of the counts. The last expression is before specified rounding and requantization. The AND/popcount path and optimized integer-count path express the same algebra. Exactness is for encoded numbers; it does not remove quantization or nonlinear-approximation error relative to Qwen.

2. SVD coordinates plus a causal learned mesh

Let b_t be the base block’s output on the student’s causal input. E extracts standardized training-activation SVD coordinates; μ is their center. U_r contains dominant output-error directions.

With teacher targets Y and base outputs B, let R = Y − B, β be the mean residual, and R_c = R − 1βᵀ. For row-wise feature matrix Φ, the mixer is a ridge-regression fit:

Rank, ridge penalty, and correction strength are selected on validation. The frozen first four blocks are reused; deeper corrections are fitted on composed student trajectories. Training uses floating-point SVD and linear algebra. Evaluation uses the reloaded compiled correction, with integer ReLU, fixed-point rounding, and signed truncation for prefix means.

These SVD directions are data-dependent reconstruction coordinates. They are not established “meaning eigenvectors” of a nonlinear transformer. Native reference targets supervise fitting; the deployed replacement does not read teacher targets.

3. Why cross-layer fusion is worth testing

For two strictly linear operators, SVD exposes an exact intermediate coordinate map:

The map V₂ᵀU₁ may offer a compact connection between subspaces. Real transformer blocks also contain normalization, attention, gates, nonlinearities, recurrent state, and rounding. A learned fused mesh must approximate that intervening computation; the linear identity alone does not justify deleting a block.

From one block to a full chain.

Model-quality evidence and engineering recovery are separate milestones. A repaired runner is valuable, but it is not a new accuracy result.

1. Verified · local replacement

   Compile and learn a block

   Establish an integer/population base and an SVD-guided correction; reload and evaluate the compiled artifact. Exact encoded arithmetic is distinct from approximation to the native block.

2. Verified · held-out four-block chain

   Compose four replacements

   493 / 512 teacher choices match (96.29%); mean KL 0.00448743. The other 60 transformer blocks remain native.

3. Exploratory · matched-quality gate failed

   Remove a block and train fusion

   Initial two-block fusion removes 49.59% of that HDC pair’s payload, but validation agreement falls to 117 / 128 (91.41%). A separate tiny training probe shows that compiled parameters respond to an output-KL signal.

4. Verified · full-depth calibration

   Execute and freeze all 64 replacements

   Training and validation trajectories reach depth 64 without a native tail. Validation hidden-state update-relative L2 is 8.4701% over 512 tokens. All 64 correction archives and the model freeze are recovered and checked.

5. Verified · saved-state output diagnostic

   Repair and run the ordinary native head

   An initial Metal-only helper failed on NVIDIA. The repaired ordinary head scores saved all-64 states: 124 / 128 agreement on both Mac and A10. On the A10, mean KL is 0.00751416. No refitting follows this diagnostic.

6. Historical · failed continuation closed

   A native-reference equality check stops the earlier run

   The final eight positions of one regenerated validation sequence differ from the archived native reference. Initial embeddings and train / both held-out final native endpoints match exactly. That strict gate remains a failed historical result; no compiled-base rebuild or held-out student evaluation follows in that run.

7. Verified · reference policy implemented

   Select the immutable original endpoints

   Version the choice of archived h0/h64 as scoring authority, require fresh h0 equality, and preserve fresh endpoint differences. Local failure replay and 55 regressions pass; the new A10 continuation passes selection and starts exact base rebuilding.

8. Verified · all bases rebuilt and frozen

   Bind the unchanged execution package

   All 64 base receipts and correction archives match the originals. The first-four integer fixture is exact, and the execution freeze binds original payload records, sources, selected endpoints, and reference policy.

9. Verified · both held-out gates pass

   Run and audit all 64 replacements

   Known-test491 / 512 choices match; fresh-test125 / 128. Mean KL is 0.00594440 and 0.00601305 respectively. Final states and score aggregates are independently checked; cloud recovery and shutdown are complete. No test fitting occurred.

What did the fusion experiments actually establish?

Related exploratory probes—not full-depth quality acceptance

Probe	Measurement	Interpretation
Initial pair fusion	49.59295% less HDC pair payload 91.40625% validation agreement Mean KL 0.111719	Real base removal; quality gate failed.
Compiled-parameter training	KL 0.00586420 → 0.00148589 74.66% reduction	One eight-token training prefix, two scored positions, with native blocks downstream. Positive trainability evidence, not validation recovery.
Later basic fusion mesh	15.2923% hidden-state update error 45.6347% pair payload reduction	Failed the 10% state-error gate. No output score measured. 15.29% error does not mean 84.71% accuracy.

The unfused four-block HDC payload was 3,178,181,201 bytes versus 856,404,288 bytes for those native 4-bit blocks—about 3.71× larger. Pair-relative fusion savings do not establish compression versus the original checkpoint.

Why not extrapolate a four-block result to 64?

Each replacement sees the accumulated error of earlier replacements. Even a small local approximation can grow or shrink downstream. Under a local Lipschitz assumption for the native block F_ℓ:

‖e_ℓ+1‖ ≤ L_ℓ‖e_ℓ‖ + ε_ℓε_ℓ = ‖F̂_ℓ(Ĥ_ℓ) − F_ℓ(Ĥ_ℓ)‖

We have not measured certified L_ℓ values. This motivates testing the whole chain; it is not a numerical guarantee.

One frozen model.
Two held-out gates.

The completed quality experiment restored and froze the existing model, without fitting a new one. Both predetermined splits have now completed the 64-block student trajectory and passed their output gates.

1. 01 · Passed

   Restore and verify

   Check source and input hashes; restore corrections and saved states.

2. 02 · Passed

   Run the A10 head

   Score the detached saved-state validation fixture before expensive rebuilding.

3. 03 · Passed new policy

   Select frozen references

   Original h0/h64 are authoritative. Fresh h0 exact; fresh h64 retained as diagnostics.

4. 04 · Passed

   Rebuild and freeze

   64 / 64 base receipts match origin. First-four integer fixtures exact; execution freeze independently verified.

5. 05 · Completed

   Execute held-out inputs

   Both splits ran through all 64 replacements, with no native resets or fitting. Final states archived before scoring.

6. 06 · Both gates passed

   Score, audit and close

   Independent final-state and score audit passes. All 2,811 recovery files verified; instance terminated.

Score the last eight valid positions in each sequence. Each held-out split must independently achieve at least 95% teacher top-1 agreement and mean KL ≤ 0.05. No held-out results may be used for fitting or selecting this frozen candidate.

The earlier blocker, and the implemented reference policy

In the previous run, regenerated native validation h64 differed at sequence 15 (zero-based), positions 24–31: 33,712 of 2,621,440 valid values changed, maximum absolute difference 2.5, relative L2 0.001259757934334714. The difference is numerical, not archive serialization. Its specific cause is not established.

All four initial-embedding files and final native train, known-test, and fresh-test endpoints matched the original exactly. These are native reference checks, not held-out HDC model scores.

Implemented and verified: frozen-native-endpoints-v1 explicitly selects the archived original teacher endpoints. Fresh initial embeddings must still match exactly; fresh final and intermediate native states are diagnostic, not replacement targets. Fresh endpoints are preserved separately. Policy and selection hashes are bound to the execution freeze and final audit. Model payloads and held-out thresholds remain unchanged.

This is a documented change to the reference contract, not evidence that the original byte-equality gate now passes or that the numerical cause has been cured.

What would make this useful?

• Established here: bounded held-out fidelity. Both all-64 gates pass on the frozen short-context protocol. Extend coverage with separately preregistered data and longer contexts.
• Next: actual compaction. Use this unfused chain as the control when training a fused mesh that removes large inherited operators. Reserve new held-out data for any newly selected candidate.
• Then: execution value. Measure storage, latency, throughput, and energy on a declared platform against a comparable native baseline.
• Still unproven: full-model storage savings, speedup, robust generation quality, matched-quality fusion, and complete FPGA inference. No new paid run is launched by this result.

A dated record, not a live monitor.

This page summarizes retained experiment artifacts and independent checks. All nine allocations in this campaign are closed. The last instance was confirmed absent at 01:47 UTC, and its two scoped controllers were retired. Cumulative compute estimate is $23.019946 under the same $50 cap ($34.529919 including the 50% reserve), not an invoice. Shutdown preceded the unchanged compute cutoff of 8 September 08:30:51 UTC and provider deadline of 09:30:51 UTC.

The final audit checks frozen source/payload maps, base receipts, correction bytes, selected endpoints, final-state hashes, exact Q24 decoding, independently recomputed hidden-state errors, and score-row aggregates. It does not independently recompute all head logits, locally rehash large base payload bytes, or replay the full trajectory. Those limits remain part of the result. The now-retired shutdown guard depended on the supervising Mac and network, not a provider-hosted TTL.

Download evidence snapshot ↓

Pinned checkpoint

mlx-community/Qwen3.8-27B-4bit
Revision 3e6447f082e89cc7f0bc6e5441afd38dfce760ff

Original all-64 model freeze · SHA-256

6ab97d38fbf61e0230b59916030ac35784b9fea7210a6185091620183c001641

A10 validation head result · SHA-256

1a6294a6bf124aa926ddbdd20399ecc7b093c098912a532bed69120d2e6fd16d

Independent head verification · SHA-256

ed9f9b5cb309345d7350aac956091dde7d5b96e47abca44df7a4935731a943d4

Native endpoint reproduction diagnostic · SHA-256

9e88889f53c1f0d3a50c8c1ead209a4aedd724da196023ee3fe06d45bfc84c6d

Hashes identify retained internal evidence; they are not a public replication package or a proof by themselves. The downloadable snapshot contains the publication-safe measurements and their scope, not checkpoints, raw corpora, or infrastructure credentials.

Frozen native endpoint policy · SHA-256

698e29f7fc754adf690c5ea6a518e43539f39bb2472236c1a3e52d726841a1fd

Real mismatch replay · SHA-256

9673118878971983b02613a75eeab0833fe275b3010356d31cd389f5c7656021

Selected endpoint receipt · SHA-256

e6986daa145f628b6928827b05f24c7010e82a45dc5f209d119839bc3f29196d

Read-only continuation audit · SHA-256

aeee19c6afb54bb5ecd16fcbe3480cfc527b907fef5cb4c128670151184341d1

Continuation execution freeze · SHA-256

3ea634f6c3f40fb01fdcf1e2e4903042f49563249ec4cd051db898f2dcf5a8e8

Independent execution-freeze audit · SHA-256

c8242023adb762d4bb38217adff3902ffd5f8c87b982608991fbc39b6dd0dc47

Known-test all-64 result · SHA-256

543b0047c90e74aaf1a0e63292af9ae550902a84cc84f5a1d6b142c03eed477d

Fresh-test all-64 result · SHA-256

677b4940e427f64724990e529b2d0e44440897cdd0b60d1f83e6dd3f0a50edf3

Independent final-results audit · SHA-256

768ba96e49dd2503580a0f3a3f11b841e342a948cd94a4432d267f166a55a8d0

Final-state archive · SHA-256

836e39b04ce1164c474aa1684843d1075ea1d73afb99c3a9494ccea0e4ae5d3c

Completion marker · SHA-256

275357cdad53d961aaf989388818bb435256dc0d2d359de1c3537fc20a4ee5ba