Donor layers traced
32 balanced experts per layer; 96 neurons each.
Measured evidence
Model fracture milestone
Real Qwen3 0.6B fracture, exact K32 engineering reconstruction, and a completed protected four-way trace for material-sparsity alignment.
Recorded finding: All 28 donor layers produced balanced expert assignments, K32 matched five frozen engineering cases exactly, and all 112 protected trace units completed across four isolated roles.
On this page
Currency and next evidence gate
Decision signals
Key metrics and controls
Each value is mapped directly from the recorded summary artifact and retains its experimental, historical, or control context.
Exact frozen cases
First logits, first-token top-five order, and complete greedy generations at K32.
Maximum first-logit error
Observed across the five registered engineering cases; broader token-level comparison remains a next gate.
Protected trace units
4 isolated roles across all 28 layers; no failed or interrupted unit was promoted.
Source-linked visual evidence
Readable at a glance. Inspectable in detail.
Charts are generated from normalized source fields. Every visualization includes its exact values in an accessible table.
Status matrix
Exact reconstruction checks
Matched and compared case counts for the three frozen K32 engineering-parity checks.
Finding: All three checks matched in all five cases; this is exact observed reconstruction within the frozen set, not broad quality evidence.
First-token logits
First-token top five
Greedy generations
View source values
| Group | Matched | Compared |
|---|---|---|
| First-token logits | 5 cases | 5 cases |
| First-token top five | 5 cases | 5 cases |
| Greedy generations | 5 cases | 5 cases |
Ordered source series
Assignment structure across 28 donor layers
Minimum, median, and maximum difference between balanced activation-derived assignments and contiguous neuron slices.
Finding: Every layer produced a distinct balanced assignment 96.5–97.6% different from contiguous slicing; this establishes non-arbitrary organisation, not semantic superiority.
Minimum observed
Median observed
Maximum observed
Connecting strokes aid comparison between published ordered points; they do not represent measurements between those points.
View source values
| Point | Assignment difference |
|---|---|
| Minimum observed | 96.5 % different |
| Median observed | 96.9 % different |
| Maximum observed | 97.6 % different |
Progress and control
Native adapter update scope
Trainable and frozen parameter shares in the 671.315M-parameter one-step bootstrap.
Finding: The update covered 577 injected modules while training 1.552% of parameters; it proves local numerical execution, not convergence or generation quality.
Trainable adapter
Frozen parameters
View source values
| Group | Parameter share |
|---|---|
| Trainable adapter | 1.552 % of parameters |
| Frozen parameters | 98.448 % of parameters |
Method
How to interpret this pack
- A frozen real-model activation trace covered all 28 feed-forward layers and assigned every source neuron exactly once into equal-width experts.
- The complete K32 expert set was reconstructed and compared with the donor on five pre-registered engineering cases; a separate native checkpoint then completed one bounded adapter step.
- A separate protected trace completed all 112 layer-role units across train, calibration, validation, and final-holdout roles, with distinct split identities and no fit access to the final holdout.
- The bounded alignment execution path is implemented and its protocol is executable; this pack does not present a complete real aligned checkpoint or terminal alignment receipt.
- The machine-readable source JSON retains all 28 layer rows; the visual chart presents their minimum, median, and maximum for legibility.
Claim boundaries
The bounded alignment path is executable, but no complete aligned checkpoint is promoted; K8/K16/K24 quality, savings, broader generalisation, and production readiness remain unevaluated.
- Exact parity is limited to five frozen engineering cases at complete K32 reconstruction.
- The activation curriculum was a small synthetic engineering set and does not establish semantic or routing superiority.
- The adapter result is one finite update, not convergence, benchmark quality, or sustained training evidence.
- The protected four-way trace is complete and the bounded alignment path is executable, but no complete real 28-layer aligned checkpoint or terminal alignment receipt is promoted here.
- K8, K16, K24, and adaptive-K quality, latency, throughput, memory, and energy remain unevaluated on this converted model.
- No production-readiness, cross-model portability, continual-learning, or capability-marketplace claim is supported by this pack.
Inspection and reuse
Downloads and provenance
Source data
Machine-readable summary
The recorded JSON artifact is staged from the governed public evidence pack during the build. Its currency notice determines whether it may be read as current.
Open source JSONProvenance
Build and evidence receipt
Use the public build receipt to compare the deployed site source, evidence digest, claim posture, and inference release state.
Open build receiptContinue reviewing
Related evidence
measuredreviewed
Refresh required- Observed
- Small controlled probes provide measurement signals suitable for engineering decisions and follow-up experiments.
- Supports
- Step counts, throughput ranges, parameter footprint, and pair-work reduction from bounded probes.
- Does not support
- Probe-scale behavior is not a substitute for matched end-to-end production benchmarks.
- Next decisive evidence
- Regenerate from the repaired exact artifact identity and close or retain the phase-41 contract-discovery failure.
Decision relevance: research · engineering
Review evidencecontrolreviewed
Superseded- Observed
- Critical system surfaces are governed by explicit pass/fail controls rather than narrative assurance alone.
- Supports
- Configuration validation, critical gate coverage, phase-control checks, and API observability posture.
- Does not support
- Control coverage describes engineering discipline; it does not independently prove model capability.
- Next decisive evidence
- Rerun and close—or retain—the three failing raw-discovery validations, then publish an exact-SHA quality-gate pack with executed assembly CI.
Decision relevance: governance · diligence
Review evidence