# WORKSTATION-003 findings: reviewed transformations and measured dispatch

The local Transform Lab is implemented and verified within the registered scope.
On the 48 held-out synthetic streams, actual selected-policy dispatch took
**8.153366 seconds**, compared with **91.578442 seconds**
for the one-thread CPU baseline: **11.2320x** for the measured work units.
All 48 dispatches were faster in this execution. This ratio includes setup,
requests, updates, rebuilds, capture/recapture and teardown; it excludes process
startup, calibration, input/reference generation, evidence writing and verification.
It is not an application-startup, total-study or universal hardware speedup.

The [frozen protocol](TRANSFORM_WORKBENCH_SPEC.md),
[complete 648-row ledger](TRANSFORM_WORKBENCH_RESULTS.csv), and
[curated receipt](evidence/TRANSFORM_WORKBENCH_RECEIPT.json) separate implementation,
exact qualification, numerical parity and performance. No production engine was
changed. Use [the local guide](TRANSFORM_WORKBENCH.md) to launch and inspect it.

## Delivered behavior and exact qualification

The versioned data-only plan declares which observations must survive a rewrite.
The default memory example writes green, overwrites it with blue, and returns
both ordered blue answers. Sharing a pure read preserves its observation slots.
Removing the first write is rejected when ordered write/reset provenance matters;
it becomes eligible only under the explicitly narrower answers/final-bindings
purpose. Original events remain inspectable in both cases.

Preview, explicit review, Start, Stop, Resume, export, import and reset use the
same CLI/service contracts. Edits invalidate review. Imports validate separately
before adoption and cannot select code, paths or services. Stop targets owned
processes; Resume starts empty process/device state and preserves earlier attempts.
Memory uses the original CPU environment and requires no CUDA.

The independent exact interpreter passed **7,200 miniature cases** and **16 longer
development chains**. All five negative controls passed: provenance erasure,
stale reads across reset, observable operation order, missing-key context and
collapsed-state substitution. These are bounded operational controls, not general
language understanding or a proof over every possible input.

## Frozen execution and performance

Calibration used seed 202610061, 24 cases, three repetitions and all five arms.
Minimum median complete cost, the fixed 5% guard and fixed tie order sealed the
policy before seeds 202610062 and 202610063 were evaluated. Every held-out case
ran every candidate plus a fresh actual policy dispatch. The study kept all
candidates and used float32 with TF32 disabled against float64 references.

| Candidate | Complete held-out candidate seconds | Selected held-out cases |
|---|---:|---:|
| `cpu-group-1` | 91.578442 | 0 |
| `cuda-direct` | 18.720768 | 8 |
| `cpu-prefix-incremental` | 13.806079 | 18 |
| `cuda-prefix-incremental` | 17.474996 | 4 |
| `cuda-graph-direct` | 12.728092 | 18 |

The selected dispatch used CPU prefix reuse for 18 cases and CUDA for 30.
Reconstructing selected costs from the candidate runs gives
8.540135 seconds (10.7233x),
which is a separate quantity from the measured 8.153366-second dispatch.
Dispatch reused 20,354 intermediate levels and recomputed
47,326; final answers were never cached from references.
Core changes invalidate downstream levels; index, shape and context changes
rebuild workspaces. CUDA Graph execution is a separate direct-contraction arm.

![Every held-out actual dispatch, with its selected arm and CPU cost ratio](evidence/transform-dispatch.png)

This scientific figure uses every held-out actual-dispatch row in the linked CSV;
it is generated with Matplotlib, not conceptual artwork. Colors encode log2 of
the cost ratio; each cell reports the numerical ratio and selected arm. Artifact
and renderer hashes are recorded in the curated receipt.

## Numerical evidence and retained costs

Trusted verification checked **497,829,888 output scalars** across
600 candidate output files and 48 actual-dispatch files, against 72 reference
files containing 55,314,432 float64 scalars. It checked full
ordered arrays, duplicates, update schedules and nonzero reference controls.
All outputs met `1e-5 + 1e-4 * abs(reference)`; the maximum absolute dispatch
error was 1.95487727e-06. Saved-array payload totals
2,433,835,008 bytes. There are 58,968 completed
request records, including calibration, candidate and dispatch executions.

The three startup probes use distinct fresh processes. Driver and OS caches
were not reset, so these are process-cold observations, not machine-cold ones.

| Probe | Worker elapsed seconds | CUDA initialization seconds | First contraction including transfer/setup seconds |
|---|---:|---:|---:|
| cold-1 | 18.172 | 2.797 | 15.126170 |
| cold-2 | 18.391 | 2.984 | 15.144530 |
| cold-3 | 18.453 | 2.734 | 15.451335 |

The resumed calibration/evaluation worker took **941.25 seconds** overall,
including its 3.343-second CUDA initialization and
937.610-second study portion. This includes all candidates, calibration,
references and evidence work; the 8.15-second dispatch sum is not this duration.
The supervisor recorded 1316.266 cumulative seconds across the retained
run, including cancellation and built-in verification, and 3,415,073,605
output bytes. The later standalone verification is outside that run timing.

Execution used one NVIDIA GeForce RTX 5070 CUDA lane, one numerical BLAS thread
within the existing four-thread cap, Python 3.12.10 and PyTorch 2.14.1+cu130 on
Windows. The resumed session peaked at 1,274,408,960 bytes of monitored
process-tree RAM, 379,593,216 allocated CUDA bytes and
473,956,352 reserved CUDA bytes. Minimum observed device-free
memory was 11,043,602,432 bytes. These fit the frozen 12 GiB RAM,
8 GiB allocator, 1 GiB free-device reserve, 1 GiB explicit workspace and 256 MiB
intermediate-cache limits. Allocator limits and monitoring are not OS isolation.

## Lifecycle, software checks and evidence identity

The official run was deliberately canceled after a completed CUDA calibration
arm. Its incomplete session attempt remains sealed. Explicit resume used a new
worker with no restored device cache, completed every stage and passed built-in
and separate standalone trusted verification. All owned workers had exited.
Verification recomputed references and checked saved outputs through matching
current source; it did not rerun CUDA or independently remeasure performance.

Before official execution, a real browser test exposed a Resume polling race:
the old incomplete receipt briefly hid the newly active controller. The fix
tracks the owned controller while it checks retained evidence. The earlier
partial UI result and completed backend run remain recorded; the final source
was frozen again before any official calibration or held-out execution.

Final validation at the frozen implementation:

- CPU suite: **2,063 passed, 56 skipped, 27 subtests passed**; 235.90 seconds.
- Actual configured CUDA suite: **162 passed**; 21.41 seconds, covering the new
  numerical modules and unchanged tensor dependencies.
- `renigma math-check` passed. Optional dependency/device skips remain skips.
- Browser acceptance: reviewed memory Start/Stop/Resume, retained attempt,
  both blue observations, manual export, reload/catalog/import and re-review.
- Actual guided seed-0 CUDA execution completed and its report was exported.
- Canceled/malformed imports, stale review, keyboard focus, 320px touch layout,
  reduced motion, no-JavaScript information and no external requests were checked.
- Launcher/resource rejection, ownership, bounded records and invalidation
  checks passed. A Windows symlink-creation test was privilege-skipped; the
  separate junction rejection test passed.

Execution occurred on 2026-10-06. Full private machine receipts and synthetic
arrays stay under `G:/RenigmaLLM_Artifacts/transform-runs/transform-767eab459ea2`.
Validation logs stay under `G:/RenigmaLLM_Artifacts/transform-validation`.
The actual browser run IDs are `transform-2c289e25248a` (memory) and
`transform-e3d69d16838a` (development tensor); neither is the official study.
An additional agent reviewed source and recorded ledger arithmetic. That is
software/evidence review, not independent scientific replication or CodeRabbit
approval. All 17 ledger checks and 10 final acceptance checks passed; the audit
receipt hash is retained with the curated receipt. This slice is committed
locally; it was not deployed or published.

| Identity | SHA-256 or commit |
|---|---|
| Frozen implementation commit | `ae41ca768c247ed7b8174faa16f12d7a4956f929` |
| Evaluated source | `976d8fd221605ce7c6d41734064fc00d489eeb4d015062a9a88f6de080b770fa` |
| Frozen protocol document | `5c36cd27b391efcc83b426485bb4f6925400cc25fb58b9915a05c5cf15cbc5df` |
| Executable study protocol | `cd972ed491dc362b4b3a3f21c493a717785f6ef244db8f65416c7ab7fcbb8542` |
| Reviewed run configuration | `e3f94cf7c81461cb65e9c50336d20e447bd6a41a0acc371626990f87c71eaf3b` |
| Completed run receipt | `a966170e03d831c4bdbf67d0668a5cc9a60d185e82727f7a4105ef00ef4dbe53` |
| Complete CSV ledger | `9f83d0a611fb033ddf6ad13e00e17d8dcb27231a4569ae65f94841451bbe548b` |
| Standalone verification output | `0782b6653ce88c6d77d3a4e7e5efb4c0988b050a637c6845294a7fadac95f24c` |

## Disposition and next research gate

TW-001 through TW-006 are verified within this local registered scope. The result
supports testing purpose-aware reuse and calibrated resource selection further.
It does not establish a general semantic optimizer, compression, learned context,
or arbitrary-workload speedup. Two fresh seeds vary supplied data within known
profile/length/regime classes; the router knows those classes and sequence lengths.
Each held-out dispatch is measured once. Timing noise and execution order remain
limitations, despite three calibration repetitions and rotating candidate order.

The next gate should freeze fresh mixtures and unknown sequence horizons, with
startup and calibration amortization, switching cost, tail latency and regression
criteria defined before execution. Compare a fixed CPU policy, fixed CUDA policy
and the calibrated policy on the same required observations. Expose unknown
workloads explicitly and preserve every nonselected arm and failed case. Do not
select or tune methods on this completed held-out set.

Earlier evidence remains valid within its own scope: WORKSTATION-002 had two
routing regressions and reported a reconstructed selected cost, while TENSOR-001
found no memory-storage merit. This study changes neither result. Product
admission, custom fusion, lower precision, additional GPUs/NPU/WebGPU, learned
language, topology/PDE/zeta extensions, cloud provisioning and independent
replication retain separate gates.