# WO-04 — Source pins and live model test gates

Status: automatic placement authorized; Qwen/Laya/Kev choice baselines,
Moondream's twenty-capture grounding test and one configured Qwen browser task
measured; owned services released.

The user-supplied model catalog is a design input. Pin selected code and weight
revisions from primary metadata and record checkpoint license, tokenizer,
quantization, runtime, corpus revision and actual hardware for every run. Source
metadata verification is separate from package installation, license sign-off,
GPU compatibility and inference verification. Unselected alternatives stay marked.

Read-only `docker gpu discover` and `nvidia-smi` on 2026-10-11 found three A100s
already used by model/voice services, plus a GT 1030 excluded from inference.
An active voice-service lease is not available capacity. No service was stopped,
lease reclaimed or model loaded for this work.

The user answered the placement question: "automatically decide, automatically
place on whatever is open and useable". Select suitable A100 capacity through the
broker, document exact endpoint/tag/UUID, and verify the actual runner before
inference. This authorizes selection; it does not waive the placement gate or
allow disrupting existing leases. No CPU/GT 1030 fallback.

The earlier temporary Laya environment was cleared by the host restart. It is
not counted as an available reproducible runtime. On 2026-10-12, broker discovery
reports admission blocked, a missing selected A100 and NVIDIA recovery action
`Reboot`. A later recorded preflight at 04:16:56 UTC reports an NVML driver/library
version mismatch (library 580.178) and all three selected A100s unverifiable.
Read-only `/api/ps` has no resident Ollama model. No model load or
inference request is permitted while this condition remains. Existing voice
services and GPU leases have been preserved.

Persistent source pins and the read-only preflight/benchmark harness are under
`computer-use/` and `scripts/rapid-computer-use/`. Their fixture tests validate
contracts and rejection behavior only; no local model latency, quality, visual
grounding accuracy or peak VRAM is claimed from those tests.

The persistent harness passes 69 offline tests and freezes ten finite-action
cases plus twenty distinct local browser PNGs. The images use DOM-measured target
bounds and exact full/crop/high-DPR/letterbox transforms. They are generated UI
fixtures, not app-disjoint production traces or Moondream predictions. Each image
and annotation is hash-bound in the source lock. See
[`computer-use/BENCHMARKS.md`](../../../computer-use/BENCHMARKS.md) and the recorded
[blocked preflight](../../../computer-use/evidence/host-blocker-20261012.json).

Before live tests, follow `/usr/local/share/ollama-unify/AGENTS.md` and repository
AGENTS.md: fresh broker discovery; scoped lease for changed CUDA workloads;
checked server GPU visibility; actual runner placement; no runner on an unintended
device; commentary with preflight evidence. Only the specified zero-token load
probe exception can precede actual runner inspection.

After that gate: Laya and Kev smoke tests, 20 local Moondream captures with exact
coordinate transforms, Qwen at selected context length, repeated cold/warm timing
and peak VRAM. Freeze the same legal-action corpus for policy comparisons and
record unsupported shapes. Report observe/rank/ground/plan/queue/execute/verify
p50/p95/p99, not author latency as local performance. Prefer shadow replay before
live side effects. Do not promote the model cascade without measured baseline.

## Hardware recovery and isolated baseline (2026-10-12)

At 04:48:15 UTC, fresh broker discovery reported healthy admission, no quarantine
and two free A100s. The voice-service lease remained on the third A100. A separate
broker-owned user model subsequently occupied GPU `GPU-fc9d264d-8a20-bf8c-eb72-72a6dcbdaf2a`;
this task left it untouched.

The test selected `GPU-3538758b-6ba7-fa1b-a80c-e315a606c3f0` under the scoped lease
`omnius/rapid-computer-use-qwen-smoke` (49,152 MiB reservation, 1,200 seconds).
An empty isolated Ollama server at `http://127.0.0.1:11469` was inspected through
`/proc/389206/environ` before loading; its CUDA visibility contained only that UUID.
The exact installed baseline is `robit/qwen3.8-27b-obliterated-e03:27b`, digest
`d185382bd866ef361e9825f1818451ca2694e25a4f503339cde4a8071f4544a3`, context 8,192.
It is a comparison baseline, not validation of the architecture's Flash-Next target.

The empty-prompt, zero-token load returned `done_reason: load` after 28.94 seconds,
with no generated text. At 04:54:51 UTC, `nvidia-smi` showed runner PID 393666 only
on the selected GPU (17,620 MiB). Process ancestry, environment, boot identity,
checkpoint digest and context were verified, and the lease was active. The result
was stated in commentary before authorizing any generation. Loading alone is not
a model-quality result.

The gated ten-case finite-choice run returned ten valid responses and six exact
label matches. All four differences were observe versus abstain; none selected a
mutating action in those disagreements. HTTP rank latency p50 was 466.17 ms, p95
14,168.33 ms and p99 23,095.73 ms. The first request took 25,327.58 ms; its cause
was not measured. Native probabilities, cold-start, peak VRAM and task-completion
metrics are unmeasured, not inferred. The report and its immutable tested source
lock are under `computer-use/evidence/qwen-27b-*20261012.json`.

At 05:07:20 UTC the owned runner was unloaded and independently absent from
`nvidia-smi`; the owned empty server was stopped and the lease release confirmed.
The historical hardware blockers, successful preflight, model responses and
cleanup are distinct receipts. At that point Laya, Kev and Moondream runtime
tests remained open; their subsequent results follow.

## Laya comparison (2026-10-12)

The staged server (code `c4a805b5160621bbfc53417126990a67d3163bc5`, checkpoint
`7b928d828b7b0e022f929d9bd2e44165aa270148`) ran at loopback port 47363 under owner
`omnius/rapid-computer-use/laya-phase0`. PID 744899's actual environment and
identity were checked before weights-only loading. Its resident allocation was
2,132 MiB on A100 `GPU-3538758b-6ba7-fa1b-a80c-e315a606c3f0` only. No forward or
warm-up preceded the host placement statement. An expired statement was replaced
only after another current preflight; the expired one sent no inference.

The shared Unix-socket lease helper uses the supervisor's token without exposing
it in argv or receipts. Each request requires active preflight/approval, prepare,
forward, CUDA synchronization and ready. Failed intervals terminate for cleanup.
Thirty-five offline service/lease tests pass. The real ten-case run returned ten
valid choices, four matching labels; native Brier sum is 0.99369639 and ECE is
0.54339. At probability 0.9, coverage is 0.4 and selective risk is 0.75. Confident
wrong choices included an unsupported saved-file claim and unauthorized payment.
No UI input occurred. Laya is not promoted to the default policy.

HTTP rank p50/p95/p99 are 758.64/1176.07/1421.94 ms including the service's
preflight and lease coordination. Neither peak VRAM nor cold-start was measured.
The loader's 11+ choice temperature-clamp warning is preserved; at most four
choices were used. Exact raw results, immutable tested lock, preflight and runtime
receipts are `computer-use/evidence/laya-*20261012.json`. The owned server stopped
at 05:29:15 UTC; independent nvidia-smi and broker checks confirmed no owned
allocation or lease remained. The voice service and other models were preserved.

## Kev comparison (2026-10-12)

Pinned code `2ecafc49fe1e8945cd600443e6c04c1cf0b3a78f`, adapter
`6cfce5c2fa4b4bd64026336ab649c5ca78857d52` and Qwen3.5-4B-Base revision
`1001bb4d826a52d1f399e183466143f4da7b741b` passed 33 offline staged-service
checks. The dedicated runtime pins 79 package versions and 33 installed Kev
source files. Its own loopback listener at port 36155 loaded only after actual
PID 990102/start-time/CUDA-environment verification. The weights-only load
performed no forward or startup warmup. Actual nvidia-smi and the active scoped
lease then proved residence only on A100 GPU0 at 16,734 MiB before approval.

Ten fixed requests returned ten valid choices, five label matches. Four errors
were observe/abstain differences and one selected the disabled-control candidate.
Native Brier sum is 0.578373458 and ECE 0.18016; threshold 0.9 accepts no case.
HTTP rank p50/p95/p99 are 908.10/1042.01/1099.28 ms including host/broker checks.
This eager float32, unmerged-LoRA configuration does not establish upstream
optimized serving parity. No UI action was executed and no default was changed.

The separately tested sampler (13 offline checks) began before load and recorded
548 valid process-bound rows at nominal 200 ms intervals. Maximum observed memory
was 16,960 MiB, with possible missed peaks explicitly documented. At 05:45:23 UTC,
independent checks confirmed the owned process, allocation and lease absent.
Exact source, runtime, raw responses, sampled memory and cleanup are recorded in
`computer-use/evidence/kev-*20261012.json`.

## Photon initial startup (2026-10-12)

The hash-checked 10,505,269,576-byte Moondream checkpoint passed raw tensor
residency under a new GPU0 lease. Actual PID 1025483/start ticks 402545, loopback
port 42803, CUDA visibility and nvidia-smi placement were independently checked;
10,522 MiB were resident. No Kestrel engine or forward preceded that check.
Only afterward was the one-token engine startup authorized.

The upstream kernel shim then aborted with missing `cudaLibraryLoadData` from
the system CUDA 12 runtime. The service's unchanged raw receipt does not prove
engine initialization or completed warmup. No point requests were sent, and no
accuracy is reported. The scoped supervisor exited; independent process,
nvidia-smi and broker checks confirmed the owned process/allocation/lease gone.
The timestamped failure, exact initial lock, raw preflight and memory sample
records are under `computer-use/evidence/photon-*20261012.json`. Correcting the
isolated runtime and repeating all stage checks is required before point testing.

## Photon compatible-runtime retry (2026-10-12)

Static inspection found that the pinned kernel loader requests `libcudart.so.12`,
which resolved to system CUDA 12.0.146. The correction pins NVIDIA CUDA 12.9.79
only inside the isolated Photon environment. Fifty Photon offline tests plus
69 shared harness checks pass. The new manifest binds 65 package versions and
622 installed files; original failed-run manifests and receipts remain intact.

Fresh PID 1179914/start ticks 476189, port 59613, exact GPU0 visibility and raw
10,522 MiB residency were checked before engine authorization. The loader
preloaded the exact checked library inode, verified symbols and actual process
mappings, completed exactly one one-token warmup and returned its scope to ready.
A new actual placement check showed the initialized engine at 13,076 MiB, solely
on the leased A100. A new receipt/approval then admitted the twenty point requests.

The run returned 20/20 single-point target hits, zero errors/ambiguous/missing
responses and zero UI inputs. HTTP ground p50/p95/p99 was
889.52/1028.01/1042.22 ms, including per-request host and broker coordination.
The process sampler recorded 1,392 valid rows and maximum observed 28,384 MiB;
200 ms sampling can miss peaks and is not allocator or worst-case peak VRAM.
This synthetic corpus does not establish real-application task completion.

At 06:03:11 UTC, independent checks confirmed the owned process/allocation/lease
gone while preserving the voice and other workloads. Exact pins, both stage
approvals, actual placement, results, original failure and cleanup are bound by
`computer-use/evidence/photon-cuda129-live-validation-20261012.json`.

## First live configured host session (2026-10-12)

After the local command passed 477 checks and the workspace build, a new scoped
Qwen service at port 46587 loaded the same exact installed 27B checkpoint at
context 8192. Empty-server identity/CUDA visibility was checked before an
empty-prompt zero-token load. Actual runner 1392116/start ticks 577765 occupied
only GPU0 (17,636 MiB), and active lease/current preflight preceded generation.

The shipped form was opened in an owned headless CPU browser. One admitted
planner request reached the strict top-level plan validator but failed with
`Invalid bounded UI plan.` The exact bad JSON shape was not retained and cannot
be reconstructed. There were zero accepted plans and zero semantic input actions;
the returned `needs_replan`, open terminal criterion and absence of a completion
receipt agree. Two actual progress explainers were emitted.

The session took 11,310.65 ms including cleanup; model-only time was not measured.
The 178 valid 200 ms process samples observed maximum 17,654 MiB after model load,
with no allocator-peak claim. The owned browser/process group, model processes,
allocation and scoped lease were independently confirmed gone. The failure and
seventeen hash-bound artifacts remain in
`computer-use/evidence/qwen-host-attempt1-manifest-20261012.json`.
This is rejection/cleanup evidence, not a completed live task. Exact response
schema delivery and retained planner diagnostics are the next repair.

## Configured planner retry (2026-10-12)

Commit `0a9ff3b3` sends the exact host action schema through the native Ollama
response-format field and retains opt-in planner diagnostics. The build, 419
affected checks, 18 independent JSON Schema cases and 34 harness checks pass.
The retry used a fresh owned server at port 50973, the same checkpoint/context,
server 1491074 and runner 1495817/start ticks 631281. Actual environment, zero-token
load, exact GPU0 residency and active lease were checked again before a new
preflight statement. No model or device fallback occurred.

One planner request returned the valid 305-byte plan. Both permitted actions
executed and their expected effects were independently observed: Title became
Report and Apply produced Applied Report. The two trace entries are verified;
six progress explanations precede the final result. The returned terminal receipt
matches the committed completion receipt, satisfied objective frontier and
completed lifecycle. Top-level `ledger.status` is a separate critic/audit
projection; `ledger.terminal.status` supplies terminal authority.

Backend-request time was 3,113.31 ms, whole session including browser cleanup
7,044.95 ms, and launcher duration 7,119.22 ms. The server independently logged
one successful `/api/chat` request. These single observations do not establish
latency percentiles or forward-only time. The sampler recorded 2,574 valid rows
and maximum observed 17,654 MiB after model load, with the same 200 ms limitations.
Root cleanup confirmed the browser, owned model processes, allocation and lease
gone while preserving unrelated workloads. Source/input/response/trace/terminal/
cleanup artifacts are bound by
`computer-use/evidence/qwen-host-attempt2-manifest-20261012.json`.

The original terminal report also exposed a display-only projection defect:
its check formatter expected a command receipt and mislabeled the typed host
verification as unverified. The authoritative typed receipt and objective
frontier were correct. The retained live report stays unchanged; the subsequent
formatter repair recognizes the bound host verification without inventing a shell
command or exit code. Its separate
[offline replay](../../../computer-use/evidence/host-terminal-report-replay-20261012.json)
renders the check as passed, with no spurious limits or next steps. The replay
checks original artifact hashes and does not run a model or change terminal
authority. Regression commands and delivery are tracked in WO-05.

## Visible three-model startup, first attempt

Commit `b7c7f88d` adds the owned headed-browser path, actual finite policy
context/receipts, configured point service and bounded post-completion review.
After its offline checks, the first shared GPU0 scope started three empty pinned
servers. Actual process identities, listeners and CUDA visibility were checked.
Qwen's empty-prompt probe generated zero tokens and loaded its checkpoint.

Photon then rejected raw loading because the new supervisor's `-s` and
`PYTHONNOUSERSITE` setting hid `annotated-doc` metadata from its prepared overlay.
No planner, finite-policy or point request was admitted, no warmup occurred and
no browser started. This is an environment-launch failure, not a model-quality
result. The runtime manifest validation remained enforced. The supervisor closed
all three services and the Qwen runner, verified allocation absence and released
its scope; the voice lease remained active.

Eighteen original artifacts and independent Qwen route counts are bound by
`computer-use/evidence/visible-startup-attempt1-manifest-20261012.json`. The next
attempt must use a fresh state directory and correct the prepared environment's
launch semantics without weakening dependency or runtime identity checks.

The launch repair preserves Photon user-site resolution while retaining Kev's
isolated interpreter flags and removing inherited Python path overrides. A
CPU-only check using the exact resulting Photon interpreter/environment matched
all 65 manifest dependency versions, including `annotated-doc 0.0.4`, without
importing Torch, Kestrel or Moondream. Supervisor regressions pass **25/25**,
including actual Python flag semantics. The retry still requires fresh CUDA
discovery, residency and inference approvals.

## Visible three-model retry: completed (2026-10-12)

After the launch repair, fresh raw load, active-lease and actual-placement checks
bound Qwen runner 2433964, Kev 2423227 and Photon 2423228 exclusively to GPU0.
Photon's separately approved one-token warmup completed once; its new receipt,
actual CUDA 12.9 library mapping and placement were rechecked before fresh
approvals. Exact profiles and source/input hashes are retained.

The owned headed Chromium window on X11 `:0` completed Title=Report, Ready=true,
Preview, then pixel-only Apply. One Qwen request returned the valid plan; four
Kev physical HTTP decisions each selected the exact subsequently dispatched
candidate. Moondream's only point was `(0.34375, 0.75)`, mapping through the
retained 1280×720 capture to CSS `(440, 540)`. Its capture hash/transform match the
click receipt. All four effects and final Applied Report observation verified.
No replan or rules fallback occurred. The valid four-event trace, satisfied
frontier and committed terminal receipt `terminal-59e0f2bf02d01bdf6a36` agree.

Execution took **25,051.82 ms**; the completed GUI remained visible for
**90,022.18 ms**. Qwen HTTP time was **4,758.51 ms**; Kev's four HTTP times were
**2,739.03 / 1,611.38 / 1,564.72 / 1,618.94 ms**; point HTTP time was
**1,705.51 ms**. These include client/service coordination and are single-task
observations, not forward-only timing or an application benchmark. Eleven
progress explanations preceded the final result. Qwen's server independently
logged one `/api/chat`; Kev/Photon disable access logging, so their four/one
completed physical client receipts are retained without inventing server counts.

The 200 ms process samplers observed maxima of Qwen **17,654 MiB**, Kev
**17,554 MiB**, and Photon **28,384 MiB**. Qwen sampling started after raw load.
Kev retains one query error among 1,609 samples; it is not interpreted as zero
memory. These are observed per-process maxima, not simultaneous aggregate or
allocator peaks. After GUI review, independent checks confirmed the browser,
three services, model allocations and owned lease absent while preserving the
voice lease and other workloads.

The [manifest](../../../computer-use/evidence/visible-host-attempt2-manifest-20261012.json)
binds 69 retained artifacts, including the actual visible-window screenshot,
model receipts, input hashes, terminal verification and cleanup. The first
startup failure remains immutable. This closes the controlled visible
three-model gate; application-disjoint comparisons, platform parity, broader
performance profiling and active-bot integration remain open.
