Sanitized package integrity
Exact signed Git object, strict file allowlist, outer hashes, canonical self-manifests, and disclosure scan.
Engineering controls
Terminal evidence handoff
An exact-source, sanitized public bundle joining current phase and claim reconciliation with a 2M synthetic context candidate and two retained sparse-geometry negatives.
Recorded finding: The package records 38/41 raw phase discovery, 44 reconciled claim rows, a guarded 2,048K synthetic CPU candidate, and two component-geometry failures that narrow the next refracture and recovery experiments.
On this page
Currency and next evidence gate
Decision signals
Key metrics and controls
The recorded values, with their controls and scope.
Raw phase discovery
Three failures remain. The separate 41/41 result checks canonical registry and declared-artifact integrity only.
Claims / evidence links
49 unique hash-registered paths with 0 reconciliation issues; this does not prove the claims.
Largest completed context
50 replicates per mode on the Tiny CPU synthetic all-zero K/V candidate; 4,096K was refused before allocation.
Turbo sequence agreement
At 2,048K. The semantic-changing turbo path is retained as a diagnostic and is not promoted.
Fine geometry first pass
Unrecovered donor-derived component geometry; both registered geometries failed at equal active width 1,152.
Source-linked visual evidence
What the evidence shows
Open any chart’s data table to inspect the exact values, or download the CSV.
Status matrix
Registry integrity and raw discovery are separate
The sanitized phase derivative retains both views rather than converting canonical integrity into an all-pass experimental claim.
Finding: Canonical registry integrity is 41/41, while raw discovery remains 38 pass and three fail.
Canonical registry integrity
Raw discovery
View source values
Scroll sideways if any columns are out of view.
| Group | Pass | Fail |
|---|---|---|
| Canonical registry integrity | 41 phases | 0 phases |
| Raw discovery | 38 phases | 3 phases |
Source-linked comparison
2,048K synthetic decode throughput
Mean current-host CPU decode throughput across 50 replicates per mode for one 2.473M-parameter Tiny candidate.
Finding: The semantic-changing turbo mode measured faster, but exact sequence agreement was only 82% and first-step logits were not bitwise equal.
2,048K context
View source values
Scroll sideways if any columns are out of view.
| Group | Full vocabulary | Turbo (not promoted) |
|---|---|---|
| 2,048K context | 357.8 tok/s | 441.9 tok/s |
Source-linked comparison
Unrecovered component approximation error
Calibration mean and p95 per-row relative FFN-output L2-squared error against the frozen 0.02 mean / 0.05 p95 gate.
Finding: Learned and oracle K8 remained above the gate; record-shared oracle K16 improved the result but still failed. Geometry error dominated router error.
Learned K8
Oracle K8
Shared oracle K16
Frozen maximum
View source values
Scroll sideways if any columns are out of view.
| Group | Mean | P95 |
|---|---|---|
| Learned K8 | 0.246 relative L2² | 0.459 relative L2² |
| Oracle K8 | 0.219 relative L2² | 0.393 relative L2² |
| Shared oracle K16 | 0.145 relative L2² | 0.270 relative L2² |
| Frozen maximum | 0.020 relative L2² | 0.050 relative L2² |
Method
How to read this result
- The importer reads only eight allowlisted blobs from the exact signed ERAIS commit and tree; it does not read the ERAIS working tree.
- Outer file hashes, the package hash list, six canonical JSON self-manifests, contracts, source identities, and a disclosure scan are verified before an atomic site-side promotion.
- Successful CI receipts remain attached to the three source evidence families. The final publication commit is explicitly not labelled CI-verified because its jobs did not start under an external billing limit.
- Canonical phase-registry integrity and raw phase discovery remain separate signals. Negative component results are preserved as decision evidence, not hidden or softened.
- Only aggregate public derivatives are staged. Raw rows, prompts, token identifiers, weights, protected identities, and moving experiments remain outside the public site.
What this result establishes
Source-family CI passed, but final assembly CI did not start because of an external billing limit. The package does not promote language quality, sparse serving, end-to-end performance, production readiness, or commercial superiority.
- The final assembly commit is not CI-verified; successful source-family CI does not silently transfer to the package assembly.
- The 2,048K result uses synthetic all-zero K/V on one CPU host and a 2.473M-parameter Tiny candidate. It does not establish language quality, generalisation, production context support, cross-host speed, whole-process memory, energy, training throughput, or commercial superiority.
- Turbo changed semantics: 82% exact sequence agreement and non-bitwise-equal first-step logits prohibit equivalence wording.
- The route-ceiling and oracle-ladder results are unrecovered donor-derived component diagnostics, not language rollouts. They do not establish recovery impossibility, router readiness, serving performance, or production readiness.
- Claim-to-evidence hash reconciliation establishes registry consistency, not claim truth or promotion authority.
- Canonical 41/41 registry integrity does not mean all experiments were rerun and does not erase the three retained raw-discovery failures.
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
measuredbounded
Historical- Observed
- 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.
- Supports
- Real Qwen3 0.6B fracture, exact K32 engineering reconstruction, and a completed protected four-way trace for material-sparsity alignment.
- Does not support
- 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.
- Next decisive evidence
- Use the retained negative geometry to test refracture and recovery, then rerun the material-sparsity ladder under quality-bearing controls.
Decision relevance: research · investment · diligence
Review evidencemeasuredbounded
Historical- Observed
- Tiny and Small completed the declared ladder through 524,288 tokens, while Medium completed every safely admitted point through 131,072; 4,991 metric rows closed with zero observatory errors.
- Supports
- A replicated current-host context ladder across three CPU-only synthetic candidates, with cache-representation counts, semantic diagnostics, and admission boundaries.
- Does not support
- The history is synthetic all-zero K/V. These results do not establish language quality, generalisation, energy, whole-process memory, cross-host performance, end-to-end throughput, or superiority.
- Next decisive evidence
- Clear final assembly CI, then advance from synthetic all-zero K/V to representative histories and quality-bearing workloads.
Decision relevance: research · investment
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