Strata standalone qualification#

BLOCKED — independent operation is an acceptance requirement, not an achieved result. The installed Strata Control opens independently, but its live Start action is not implemented/qualified. Its synthetic process checks do not load a model. The 2026-10-17 inspection found no listener on 127.0.0.1:8080.

This guide covers Strata without an OmniLAB job, OpenHands or a Codex session. The separate OpenHands integration guide becomes applicable only after this standalone acceptance passes. Publication of this guide does not enable either path.

Installed identity and target#

The inspected installation is Strata 0.1.43 with the preserved LAB serving patch, based on upstream revision 7695859aa13afb248bdb4d36f9b2e95707db079d. The native executable SHA-256 is 5f359e2d08e73b11fd7b0edc72652125c12f836394dbf1732cec0d461324cd6b; the patched serve/server.py SHA-256 is 670a36ec733fadfd0880035180e1345483ea21576a90b3501b807d68209d39f0. The patch source and its scope are in strata_serving_patch.py and its tests. Do not silently substitute another release or discard this patch while reusing its old qualification.

The target is Qwen3.8-Flash-Next GSQ-RCO Coder IQ1_M, selected as qwen3.8-flash-next-coder-iq1_m. Strata's official model guide recommends Coder for 32 GB systems. The artifact is published by ISTA-DASLab, not by the Qwen organization: preserve that distinction when using the word "official." The two existing files correspond to the publisher's pinned IQ1_M directory.

The download occupies 58,408,584,928 bytes before the tokenizer, packed data, draft model, runtime and evidence. The publisher identifies the license as Apache-2.0. Both shards already exist; no duplicate download is needed. VERIFIED: fresh full-file SHA-256 reads on 2026-10-17 matched both publisher hashes and historical evidence. The executable, patched server, tokenizer, direct-GGUF pack, expert profile, draft-model pack and frozen configuration also matched their saved identities. Recheck all selected artifacts before loading. A matching filename or GGUF header is insufficient.

This is a pruned Coder variant. Strata's guide describes expert mappings plus an SSD-resident second shard; disk size is not resident RAM size. Its 32 GB guidance does not prove fit on the inspected Windows host with a 12 GB RTX 3060 eGPU. Current free RAM, GPU staging, context, caches, Windows and other applications still determine admission. Performance published for other computers is not a measurement of this installation.

Existing interfaces and boundaries#

The installed source includes the upstream browser UI and API. The relevant upstream baseline is serve/server.py and serve/web/index.html. Local inspection confirmed root-page serving, /health, /v1/models, /v1/chat/completions, load/unload controls and client-disconnect cancellation code. Source presence is VERIFIED; current browser/API behavior is NOT TESTED. Reuse these interfaces; a second browser frontend is unnecessary.

flowchart LR
    U["User: Strata Control"] --> P["Read-only preflight and fresh authorization"]
    P -.->|"BLOCKED: live supervisor not qualified"| S["Independent owned Strata server"]
    B["Existing upstream browser UI"] -.->|"localhost:8080, during approved session"| S
    A["Bounded API test client"] -.-> S
    S -.-> E["Strata engine and verified Coder artifacts"]
    S -.-> V["Private telemetry, cancellation and shutdown evidence"]

flowchart LR
    U["User: Strata Control"] --> P["Read-only preflight and fresh authorization"]
    P -.->|"BLOCKED: live supervisor not qualified"| S["Independent owned Strata server"]
    B["Existing upstream browser UI"] -.->|"localhost:8080, during approved session"| S
    A["Bounded API test client"] -.-> S
    S -.-> E["Strata engine and verified Coder artifacts"]
    S -.-> V["Private telemetry, cancellation and shutdown evidence"]

Dashed arrows describe the standalone test still required. Runtime interface sources are linked above; this diagram does not assert that the private control already implements the live launcher. The existing public session gate and resource guard explain the retained integration safeguards; importing or running OmniLAB must not become a standalone runtime prerequisite.

The frozen serving configuration binds only loopback, allows one request at a time, caps context at 16,384 and output at 1,536, reserves 2,048 MiB VRAM and disables server-side MCP tools. It has lazy_load=false: starting this server loads weights. Opening an empty browser port is not a safe substitute for startup approval, and opening a running UI must not silently trigger an unapproved reload.

Current blockers and required remediation#

Read-only preflight on 2026-10-17 sampled approximately 19.2 GiB available RAM, below the unchanged 22.5 GiB startup screen. Docker/VM processes were present; both inspected LAB stores had no pending jobs, reservations or experimental owner. The port was unoccupied. These are dated observations, not future admission.

The previous startup approval is consumed. Native/mapped memory growth and live cancellation remain unresolved. Strata Control currently dispatches readiness for Strata Start, not a live server. Its shared synthetic supervisor qualification does not establish Strata process-tree containment or memory fit. No new live startup or large download was performed for this inspection.

Before requesting a live run, finish and independently test a Strata-specific owned supervisor: bounded process tree, host watchdog, owner-death cleanup, stop acknowledgement, finite deadlines and evidence limits. Keep the existing admission/ownership checks; do not borrow Janas's 14 GiB process cap or relabel its tests as Strata qualification. Any shared-host admission coordination must not require an active OmniLAB worker or alter its inference configuration.

Then obtain a new, explicitly scoped authorization for an exclusive test window. Closing Docker/Open WebUI affects a separately running application and requires that operational choice; do not terminate unrelated workloads automatically. Repeat RAM/commit/pagefile, GPU, disk, backup, port and ownership checks immediately before launch. Failed gates stop the attempt without lowering thresholds.

Standalone acceptance sequence#

Use fixed small requests and retained expected answers. For the tool-call probe, offer a harmless schema such as lookup_fixture(symbol) and validate its JSON; do not execute returned tool calls on the host. Record raw bounded responses and actual token counters. Report generation-only tokens/second separately from whole-request throughput and time to first token. Stop on driver, paging, ownership, resource or cancellation uncertainty. Do not mask a failed test with an automatic retry or provider fallback.

Each run must end in VERIFIED, FAILED, BLOCKED or NOT TESTED with an exact scope, timestamp and next action. Screenshots, process IDs, machine paths, full configs and unsanitized logs stay in private evidence. Publish only the reviewed summary.