Stronger-model OpenHands qualification — 2026-10-09#

Workflow documentation publication — 2026-10-10#

Publication scope is the existing development branch and a draft PR, not a main merge, tag or release. The configured former repository URL redirects to nishdel/OmniLAB-Local-Agent-Bridge-; ownership and the default base main were independently verified. The PR/remote branch records establish publication state; the older sections below retain their historical no-push/no-activation statements at their original cutoffs.

The complete workflow and README overview now explain entry points, policy/context, real stored states, protected tool execution, data locations, independent review, the dock and separate accounting. Five canonical Mermaid views were added; the two legacy architecture diagrams remain. No executable source, active configuration, routing scope or runtime installation changed. Strata stays disabled and OpenHands remains SDK-only.

The reviewed starting commits are bec07f63bc8a0991576d7392353f1a541615eeeb, 90629f46a694624ae5d59b3258cc403bef7fccba and 71791b32d58cd1d007c4bcdf84a014a9a91fdd24. The reused accounting wheel SHA-256 is df9b041c8198213b9237078e39fe71ff18e7c6e55112c771ee7d516f88db2429; the unchanged active controller wheel is 20243196aac84474af583e1dbf6a6ba653edf9dfa3d1b25dae287ec39679f5fc. These are different artifacts with different roles, not a controller upgrade.

One real Qwen27 advisory contributed the transport draft: 621 input / 244 output, cache unknown, accepted after Codex corrections and independently recorded review. See contributions. The whole cloud publication turn is attributed to implementation investment, including planning, dispatch, corrections, verification and common reporting overhead; no fractional task cost is invented. No evaluation comparison or normal operating workload was run. Actual cutoff-based counters remain private, with current totals reported at handoff. Operating savings remain not established.

Close-out checkpoint#

The accepted deployment and required prior dependencies are committed locally at bec07f63bc8a0991576d7392353f1a541615eeeb. Narrow attribution and handoff changes are at 90629f46a694624ae5d59b3258cc403bef7fccba, independently checked in a clean detached checkout. This evidence update changes documentation only. Neither checkpoint includes private configuration, sessions, databases, weights or raw evidence bundles. No push, merge, tag or release occurred.

Clean-checkout regression: 591 passed, 1 skipped in 684.092 seconds. The skip remains the Windows symlink privilege check. The separate 180 extracted-wheel checks overlap the full suite; do not add their counts. Ruff, 45-document checks, the 228-file public-source audit, wheel/sdist inspection, offline install and retained-state v0.4.1 upgrade passed. All 88 package files matched committed runtime source. Artifact manifests are private under data/qwen-closeout-20261010/package/; documentation-only follow-up checks preserve exact runtime identity. This is reproducibility evidence, not fresh-machine, Python-minimum, authenticated callback or release acceptance.

The active controller remains on the prior accepted immutable package. A separate offline-installed accounting utility provides whole-session/turn and LAB run/job bindings in the existing usage database. Duplicate bindings are idempotent; conflicting or overlapping assignments fail. Actual historical CLI receipts and live LAB receipts were checked. Shared controller costs remain mixed/unattributed; no fractional allocation or new generation-only baseline was produced. The old comparison still lacks full matched supervision costs; net operating savings remain not established. See accounting.

One useful close-out Qwen advisory produced the operator paragraph: 106 words, 291 input / 130 output tokens, independently accepted and integrated unchanged. Its runtime receipt again verified the accepted identity/profile. Nine installed model identities, the original Qwen accepted patch and the private configuration hash were preserved. Controller admission remains project-only, one inference at a time; no active job/reservation remained and normal model expiry was observed. OpenHands remains the pinned restricted SDK transport with no web UI. Strata's adapter is retained and runtime disabled under its existing documented blockers.

The compact manual ledger now has eleven untested human rows, including all requested dock transitions, display/usage, cancellation, explicit override and restart persistence. Automated checks do not change those human results. Verified absolute commands and the unchanged deployment rollback procedure are in private data/qwen-closeout-20261010/operations.txt. The prior installed utility can be used again without deleting the additive attribution table or receipts.

Standard model milestone#

Current milestone: project Qwen27 default and supervised routing active. The existing codex/selective-local-execution worktree and HEAD 0b817c572a81a22c94988cdbf163d214957985b7 were preserved. The starting manifest records all pre-existing changes separately. The prior 564 passed / 1 skipped source run and 153 overlapping wheel checks were verified in their actual logs; they are not the new milestone's validation counts. Historical accepted patches, original checkouts, Qwen35 acceptance, and Gemma's incomplete UTC result remain.

The personal project profile selects qwen3.8:27b on Ollama 0.40.2: aaee06c39dcf2437cde036998d960e1fc1494b8191be7cc9657d01e509097813. The exact dense Qwen35-architecture weights are Q4_K_M; the tag's name does not identify a separately qualified architecture. Full local blob hashes matched the accepted manifest. Qwen27's historical UTC patch was independently inspected and applied exactly; its immutable original remains intact.

The native request and actual runner agree on context 65,536, 28 GPU layers, draft MTP 4, one sequence and loopback serving. Maximum output is 1,536 per response, temperature zero, thinking false, no truncation or context shifting, and keep-alive 120 seconds. Receipts preserve backend/template/digest, request options and observed runner process identity. Stored template SHA-256: b507b9c2f6ca642bffcd06665ea7c91f235fd32daeefdf875a0f938db05fb315. The runner uses the accepted chatml template path with no Jinja. The final safety audit added explicit rejection of missing/zero loaded VRAM for GPU profiles. A clean-start positive check observed 8,051,543,243 loaded GPU bytes and used 152 input / 31 output tokens; its one-sentence advisory was independently accepted without repair as a validation probe. Unknown or changed effective settings stop the job; configuration fields alone are not proof.

The pinned OpenHands SDK/tools 1.53.0 image and restricted file mailbox were reused. Native inference is performed by trusted LAB outside the networkless tool container. There are no host-drive mounts, Docker socket, credentials, privileges or general outbound network inside the model's tools. Only declared task files are writable; original input and acceptance checks are protected. Live tool-policy rejections are retained as failed-attempt evidence.

Strata is a separate disabled provider. The inspected upstream commit fb58e0dbc8399662c0e47c76578c6e878b14f6cf (engine 0.1.41) supports the documented Flash-Next candidates, not this qualified dense Qwen27 combination. No runnable compatible installation or weight set was located in the bounded local inventory. The proposed roughly 64 GB Coder assets were not downloaded; CUDA 13's documented driver requirement also exceeds this host's driver. The existing CUDA 12 alternative is not qualification. No Strata live acceptance, tool round trip, cancellation or loaded identity is claimed. Services are not exposed to a LAN, and system networking/drivers remain unchanged.

Final source regression: 586 passed, 1 skipped in one complete run after the GPU fail-safe. Every Python source/test hash matches that run's manifest. Earlier 583-pass evidence remains preserved separately. Extracted-wheel checks overlap this suite and are reported separately in the private package evidence. Lint, 45-document checks, the 228-file public allowlist/source audit, offline installation and the actual 0.4.1 retained-state upgrade passed. Final wheel/sdist identities and counts live in the evidence manifest, avoiding a package hash embedded inside itself.

Real contributions: AGENTS guidance accepted without Codex repair; seven UTC tests accepted after one local retry with no Codex patch repair; installed-default advisory accepted after a three-word Codex trim to meet its 120-word requirement. The advisory used 992 actual input and 208 output tokens, with no model override. The installed controller verified the same digest, loaded 65,536 context and effective runner profile from a clean start.

Live cancellation observed a new inference loader, reached durable cancellation, retained the uncertain reservation and reconciled it only after an empty resident inventory. No shared Ollama process was terminated or model manually unloaded. Controller shutdown/restart and recovery preserved job states without replay. The cancelled call's partial counters are unavailable, not zero. Timeouts, malformed/truncated completions, protected writes, provider switching and ownership boundaries also have deterministic and process regression coverage; deliberate host memory exhaustion was not attempted. Historical live OpenHands mailbox cancellation/recovery evidence remains valid for the unchanged pinned executor.

The actual dock is open on the project store. Hidden real Tk tests verified its state/usage data and expand/collapse against the live retry. Keyboard/focus/DPI, readability, tray and human click/hide/reopen checks remain not tested; the manual ledger has ten untested rows. Automated observations do not pass them.

Milestone evidence is under workspace data/qwen-standard-20261010/: starting-point.json, identity-before.json, assets-verified.json, operation/companion/ durable jobs/receipts, comparison/, strata-status.json, and final acceptance/activation records. Installation paths and exact local commands are recorded privately in operations.txt. See setup and rollback, contribution outcomes, accounting methodology and the compact manual ledger.

Codex owns safety, provider bindings, integration and final decisions; these changes are direct supervising work. Qwen's real bounded documentation and test tasks have separate reviews, including failed attempts. The small isolated Codex-only test-generation comparator used the same frozen inputs and protected checks, and never supplied its solution to Qwen. It reported 23,173 input, 0 cached input and 1,127 output tokens (393 reasoning tokens are an output subset). That artifact phase cost 24,300 cloud tokens; supervision and evaluation overhead are separate. Matched hybrid supervision cannot be isolated completely from this implementation session: net operating savings are not established. Actual implementation investment, cutoff, identifiers and missing metrics remain auditable in the existing usage store. No monthly or subscription-cost saving is claimed.

Previous: bounded MoE batch — Qwen35 fixture and UTC independently accepted; Gemma safe fixture accepted, UTC output-limited. Historical results below remain unchanged.

Earlier follow-up: the single Gemma comparison ended at the GPU-memory stop during loading. Qwen's accepted UTC patch remains unchanged and isolated; Gemma has no implementation acceptance.

This report supersedes the current readiness conclusions of the 4B bounded qualification, whose failures and artifacts remain historical evidence. Historical qualification was scoped to that batch; the current project activation is recorded above. Both broad pilots remain disabled. Branch: codex/selective-local-execution, base 0b817c572a81a22c94988cdbf163d214957985b7. Original development/release checkouts remain separate. No commits or publication.

Evidence root: workspace data/openhands-stronger-20261009/ (N). The old frozen task and initial approval evidence remain in data/openhands-bounded-20261009/ (R) and data/openhands-scoped-20261009/ (S).

Source and narrowly scoped bootstrap#

Codex performed infrastructure changes directly while combinations were unqualified; these changes are not local-agent contributions. The controller runs hashed source copies N/runtime-1 and N/runtime-2, never changing its imports while it owns work. The pinned SDK/tools remain 1.53.0. The final collector image used by every new model attempt is sha256:b0ec53df107bdbb7b9df9c397ee56d4f7f74bc49d6e6550b16ca6f23e5748090. No image/base/dependency replacement occurred. The existing 174-artifact package and its 101,666,057-byte approved payload were reused. New downloads: zero.

Changed infrastructure:

• inference.py: transparent per-request accounting, explicit native Ollama context/offload/keep-alive controls, disabled backend truncation/context shifting when requested, actual prompt/output counts and provider timings.
• task_contract.py / selective_config.py: compatible optional safety/repair budgets and exact-contract supervisor approvals. These substitute only for category pilot admission; identity, tool, sandbox, risk and memory gates remain.
• selective.py: persist rejected-request accounting; enforce bounded SDK error recovery; collect cancelled work before stopping volatile storage.
• qualification.py: fixed-seed repeated fixture, explicit limits and optional policy-only execution guidance; no changes to its semantic acceptance.
• tests/test_selective_bounded.py: corresponding admission, native settings, fixture identity and repair-budget regressions.

No Codex-authored candidate solution was supplied to either model.

Context and resource admission#

N/historical-context-audit.json replays the rejected 4B request without inference. Its serialized request occupied 32,306 UTF-8 bytes; the old 1,024-token output reservation required a conservative 33,330-token allowance against 32,768. This was a conservative admission rejection, not measured tokenizer/KV-cache overflow. It was not cumulative input-token usage. No exact local tokenizer was verified. The conservative full-JSON byte bound remains, with role/tool/overhead accounting and a new explicit safety margin; no requirements, failures or history are silently removed. Actual tokenizer counts are recorded separately when the backend responds. Ollama 0.40.2's request controls were checked against its API types.

Both Qwen configurations use 65,536 context, 1,536 output, 1,024 extra safety, temperature 0, thinking false, native Ollama, strict context and 120-second keep-alive. GPU layers: 14 for 35B, 28 for 27B. No global runtime setting changed. Fixture cap: 12 inference requests; repository cap: the frozen stricter 8. Both allow at most two model error-recovery cycles, zero transport retries, 15-second commands, 16-KiB command output and 30-minute attempt ceilings. Batch ceilings remain 90 live minutes, two actual fixture attempts per model, two repository attempts total, and at most three distinct model configurations. Configuration changes never reset accounting. Pre-dispatch bootstrap failures remain recorded separately from actual model fixture attempts.

Read-only sandbox: 2 GiB RAM/no swap, 2 CPU, 152 MiB combined writable tmpfs, 128 PIDs, disabled container logging, no network or host mounts. The controller alone contacts http://127.0.0.1:11434; there is no configured cloud fallback. Acceptance files are protected. Host job artifacts are bounded at 128 MiB; resource evidence is separately capped at 8 MiB per attempt. The pinned collector retains actual candidate bytes and command audit before volatile storage is stopped. Incomplete evidence never establishes acceptance.

Before loading, require 6 GiB available host RAM. Stop immediately below 2 GiB, below 3 GiB sustained for 15 seconds, GPU diagnostics failure, or more than 11,776 MiB GPU allocation. Snapshot, controller, inference process and sandbox memory are measured separately. Offload is allowed; active-parameter count is not used as a proxy for total weight memory. Only one inference residency runs.

Strata location recovery and readiness#

Strata installation/runtime/model assets were not located in the searched locations. This is not a claim of absence from the whole computer. No Strata launch, installation, download, driver change or model conversion was performed.

Actual Codex data root was resolved as %USERPROFILE%/.codex (CODEX_HOME unset); its resolved absolute path is retained in the private discovery evidence. The October 8–9 LAB sessions were inspected for relevant tool commands, including custom calls and their outcomes; other recent sessions were searched only for Strata tool commands. No unrelated conversation or credential dump was used. The bounded command/path evidence is in N/strata-session-discovery.json, strata-path-discovery.json, strata-lab-record-discovery.json and strata-inspection.json. The search checked LAB's relevant records/worktree/configuration and direct children of the workspace parent, the data drive software/download/temp roots, user Downloads, Documents, application/cache roots, and referenced Strata/Hugging Face paths. There was no whole-drive content/model-binary scan. No record justified a WSL or container installation-path search. No access-denied limitations occurred in the recorded path scopes; unsearched locations and older sessions remain outside this bounded conclusion.

The pinned inspected upstream revision remains fb58e0dbc8399662c0e47c76578c6e878b14f6cf, engine 0.1.41; the separately cached current MODELS content matched the pinned SHA-256. The proposed Coder IQ1_M artifacts and full hashes are in I/approval-strata.json (I means data/installed-models-20261009/). That preparation manifest is not installation proof or authorization. Its proposed package is about 64.4–64.7 GB download, 120 GiB reserved preparation; standard CUDA 13 requires a newer driver, while the documented CUDA 12 alternative remains unqualified. Actual runtime memory, complete assets, loaded identity and fit are unmeasured. No Strata live inference or OpenHands-to-Strata qualification is established. Existing Ollama qualification proceeds independently.

Batch outcomes and validation#

The exact qwen3.8:27b configuration independently completed and was accepted for the frozen UTC timestamp repository task. The accepted patch stays in its disposable worktree/artifacts; it is not merged into this development tree or any original checkout. Human dock, focus and DPI acceptance remain not tested; automated checks do not fill those manual rows. A fixture pass alone never accepts the repository task.

Hardware and exact model identities#

Timestamped N/hardware-before.json (2026-10-09 16:38:43 UTC) records 34,139,435,008 bytes installed system RAM; external RTX 3060, 12,288 MiB VRAM, driver 560.94, PCI 07:00.0; exact physical UUID retained privately in the hardware evidence. The internal GTX 1060 Max-Q remains Code 31 (PnP driver 582.66); its live UUID/VRAM are unavailable through this active driver. Intel UHD 630 is also recorded. No device repair or pooling was attempted. Available storage was about 143 GB on the system volume and 2.92 TB on the data volume. No hardware change was made.

Each advertises 262,144 context and tools; the new live Qwen requests instead verify the explicitly loaded 65,536 context. The existing Ollama 0.40.2 service and model store were reused without restart. Before/after nine-tag digests and referenced file inventory are unchanged (N/inventory-before.json, inventory-after.json, model-before.json, model-after.json). No model-store addition/removal/retagging was detected in this batch. Normal runtime logs, transient residency and context/checkpoint caching occurred; they are not model installation. Existing launcher port-bind warnings were observed; the active loopback service continued serving requests. Those user-owned launchers were not stopped or reconfigured.

Observed comparison#

One earlier 2.9-second harness launch failed before creating any job or sending inference (N/qwen35-fixture-1). It remains recorded and counts toward elapsed budget, but is not a third model fixture attempt. Total batch: 20 inference requests across four actual model attempts, about 25.5 live minutes including loading/baseline/cleanup; within every declared attempt and batch ceiling. No resource stop, device loss, OOM or unexpected CPU-only fallback was observed. This is a short qualification, not an endurance or eGPU performance certification.

Provider-reported load totals: 35B attempts 76.31 / 75.15 seconds; 27B fixture 59.18 seconds and repository 59.08 seconds. Controller inference intervals include load, prompt processing and generation: 246.90 / 279.99 seconds for 35B; 318.84 / 445.86 seconds for 27B. Overall elapsed also includes admission, protected baseline checks, collection, verification and cleanup. These intervals overlap and must not be added together. Exact samples/counters are in N/batch-summary.json and per-attempt resources.jsonl/inference receipts.

Local input/output token aggregates: 35B first 30,725 / 326; 35B retry 37,827 / 513; 27B fixture 24,681 / 329; UTC task 41,670 / 442. These are sums across requests, not individual context sizes. Local cached tokens, attributable Codex cloud input/cached/output tokens, and phase-specific cloud costs are unavailable. No cloud savings, quality equivalence or euro estimate follows.

Minimum observed available system RAM: 3.14 GiB (35B first), 4.03 GiB (35B retry), 5.56 GiB (27B fixture), 5.81 GiB (27B repository). Per-process llama-server RSS peaks: 17,938,862,080 bytes for 35B retry; 15,611,142,144 / 16,062,390,272 bytes for 27B fixture/repository. The first 35B sampling filter missed llama-server; only a supplemental snapshot exists, not a complete process peak. Controller RSS stayed below 81 MB; sandbox cgroup peaks were 409/278 MB for 35B and 275/415 MB for 27B (rounded upward). Cgroup samples include brief diagnostic exec processes. These process working sets can share pages; do not sum them as physical RAM. ollama ps records 68%/32% CPU/GPU for 35B and 61%/39% for 27B. GPU UUID/activity samples and runtime logs jointly establish external-GPU use; ps alone does not.

Review boundary and retained useful patch#

The UTC task began from the exact previously frozen source/protected acceptance at base 0b817c5. The Linux baseline failed the timezone-free case. The local agent read the actual module and acceptance file, used SDK file_editor to change only timestamp, and ran python acceptance.py. It recovered from one denied directory read; no candidate repair or replacement was provided by Codex.

The retained patch parses once, assigns UTC only to a timezone-free datetime, and preserves explicit offsets and the existing Z normalization. Codex inspected the diff and surrounding RolloutReader/counts source. The unchanged protected suite independently passed in two fresh verification executions: five time cases, invalid input and unchanged surrounding AST. Ten existing telemetry regressions also passed on the exact bytes in the disposable repository. No tests were weakened or skipped. Candidate files and protected baseline hashes are retained.

Exact job/attempt IDs are in N/utc-supervisor-acceptance.json; candidate identity 92b43c4eb089ce09c047c668874ade25d47ad7be7933ec4101d525e17596b361; patch SHA-256 7ab41b1372c8c45c49ba0faec94f6a412bfdbaff018130c21015d4aa8b874734. The exact patch is under N/store/artifacts//candidate.patch, and exact materialized bytes are in N/qwen27-repository-1/repository. Execution finished, automated verification passed, then the job durably awaited review until the separate Codex action accepted that candidate hash. Integration into project checkouts remains not done. The worker had no supervisor acceptance authority.

35B's first attempt read the seed/test but repeatedly attempted directory/shell operations outside policy; three denied observations exhausted two recovery cycles. The second attempt received policy-only guidance and generated the correct fixture, executed the test, and finished six inference requests with no error recovery cycle. Its normal freeze/collection rejected an unexpected process owner; strict fallback collection retained its complete tree/command audit and cleanup later confirmed owned termination/unload. No success was substituted for that failed terminal state. A fresh independent verifier passed those retained bytes as a diagnostic only; no full fixture qualification or UTC admission followed.

N/collector-observer-race.json reproduces rejection of an explicitly root-owned diagnostic and successful collection after external Docker stats. The first reproduction used the default worker UID and was correctly frozen; its cleanup wait failure and subsequent owned stop are retained separately. The exact original PID/UID was not captured: an exec/runc transition during concurrent sampling is suspected, not proven. The qualification harness now observes Docker stats from outside the namespace. No ownership guard, worker permission, image or production policy was relaxed. Another 35B fixture run requires a new authorized attempt budget; it was not silently treated as infrastructure-free extra work.

Operational result and remaining gates#

The approved isolated 27B/Ollama/OpenHands configuration has one independently accepted useful task. It is eligible for further explicitly scoped use; it does not enable general automatic implementation. 35B's full executor path remains unqualified despite its correct diagnostic candidate. Historical 4B failure is not a prerequisite or blocker for 27B. Strata remains independently disabled.

The eight-case and narrower pilots remain inactive. Before any activation, review the prepared per-case protected commands/context and resource fit, measurement capture and matched starting revisions using the demonstrated 27B settings. One case does not establish that all eight cases fit or compare quality/cost. Strata is not required to run an Ollama-only pilot. Human dock/focus checks remain separate and untested, not blockers for isolated model qualification.

Rollback and continuation#

No new dependency or database schema migration was introduced by this pass. Optional contract fields have compatible defaults; production expiry is unchanged. Isolated N/store contains the qualification jobs only. Existing historical stores and the original checkouts were not migrated or edited. To roll back this pass, stop its owned work, retain its evidence, and use the previously retained candidate package/configuration. Do not reset the shared worktree or overwrite unrelated uncommitted work. No global provider/implementation setting needs reverting.

The accepted UTC patch is intentionally separate from the built infrastructure package. Review/integrate it later only with new explicit authorization. A future 35B retry must use external monitoring and a new bounded attempt allocation; repeat the complete fixture before admitting a repository task. Further Strata preparation needs separate exact asset/runtime approval after locating or selecting an installation; do not assume the preparation manifest proves an installation. Human UI checks and any pilot activation remain separately gated.

Final automated gates#

Full Windows source regression: 561 passed, 1 skipped, 714.07 seconds. The retained skip requires Windows symlink privileges; it is not a passed manual check. The 111-case focused run and subsequent 38-case bounded run overlap this full suite, as do the ten telemetry cases run separately on the accepted patch. Do not add these counts. Earlier harness collection failure (wrong working directory importing the old root package) remains N/focused.xml; the correct candidate run is N/focused-candidate.xml.

Final wheel/sdist hashes, isolated installed subset results, byte equality, 0.4.1 upgrade retention, documentation/lint/source-audit gates and measured preparation usage are recorded in N/final-package-identity.json, final-validation.json and preparation-final.json. The final infrastructure package excludes the separately retained UTC candidate patch.

Installed package subset: 150 passed (28.24 seconds), all included in the 561-pass source suite. All 553 prior cases remain present; nine cases were added. Ruff, 45-document checker, 222-file public-source audit (zero findings), Git whitespace check, wheel/sdist inspection and copied v0.4.1 upgrade retention pass. All 85 installed package files match reviewed source bytes. Initial documentation audit findings were resolved by keeping physical/job identifiers in private evidence; no scan rule was weakened. No global package or dependency was changed.

The accepted 27B live jobs used hashed runtime-1. Final runtime-2 differs only in optional fixture execution guidance and extraction of the unchanged repair-budget logic into its tested helper. The 35B retry ran runtime-2; final installed controller source matches it. Final installed protected-candidate verification is recorded separately and uses no model inference; it does not retroactively qualify 35B. The accepted task remains isolated, not included in this infrastructure wheel.

Gemma comparison — 2026-10-10#

The user authorized exactly one installed-Gemma fixture attempt and one UTC attempt only after an independent fixture pass, with no fresh Qwen inference. Gemma did not independently complete the useful UTC task. Its first fixture request was cancelled during model loading after GPU usage crossed the existing safety ceiling. This is a runtime/resource stop, not evidence of poor coding or malformed tool use. UTC was NOT RUN, as required by the prerequisite. No retry, configuration sweep, substitution or Codex-written candidate occurred.

New private evidence root: workspace data/openhands-gemma-20261010/ (G). The attempt ran on 2026-10-10 at approximately 00:33:37–00:35:15 Europe/Berlin (2026-10-09 22:33:37–22:35:15 UTC). Historical N evidence remains intact. G/qwen-preservation-before.json and preservation-final.json bind the unchanged Qwen acceptance records, diff and isolated accepted source.

Frozen identities and scope#

• Gemma: exact installed gemma4:26b, family gemma4, Q4_K_M; digest 001e5dafc3c77684c2307ebc6ab8e336e10c9b18eca52acf547d72fc83c3ca8c. Metadata: 25,233,142,046 parameters, 128 experts / 8 selected, base-model label Gemma 4 26B A4B. The label is not a measurement of active memory consumption. Installed assets total 18,731,025,629 bytes, including the projector and draft. Manifest identities and presence checks are in G/gemma-artifact-manifest.json; large model blobs were not rehashed during this comparison.
• Historical Qwen: qwen3.8:27b, family qwen35, Q4_K_M; digest aaee06c39dcf2437cde036998d960e1fc1494b8191be7cc9657d01e509097813. This compares these installed configurations, not unquantized model families.
• Same Ollama 0.40.2, SDK/tools 1.53.0 and final collector image sha256:b0ec53df107bdbb7b9df9c397ee56d4f7f74bc49d6e6550b16ca6f23e5748090. The controller reused N/runtime-2; all 85 package files match the worktree, frozen runtime and existing installed candidate. Retained wheel SHA-256: abbf23dabb615a602047d343524707eb1eedcbe33ce1f44dac5513565360c704. No package installation, image rebuild or runtime code change was needed.
• Both explicitly request native Ollama, 65,536 context, 1,536 output tokens, 1,024-token safety reserve, temperature 0, thinking false, no truncation/shift, 120-second keep-alive and no transport retries. Gemma requested 18 GPU layers; Qwen used 28. Layer counts differ across architectures and do not imply an equal VRAM allocation. This single Gemma setting was not swept or tuned.
• Model defaults were preserved: Gemma top_k=64, top_p=0.95, draft_num_predict=3; Qwen top_k=20, top_p=0.95, draft_num_predict=4, presence_penalty=0, repeat_penalty=1 and min_p=0. Explicit temperature 0 overrides both installed temperature defaults. Different draft mechanisms/settings limit comparability.
• Identical frozen fixture bytes, protected tests, objective, skill version and acceptance criteria. The actual SDK system-message and tool-schema hashes match Qwen's first requests, recorded in G/frozen-comparison.json. No historical solution, candidate, transcript or accepted patch was mounted into the worker. The frozen UTC inputs were preserved but never dispatched to Gemma.

Preflight, admission and stop#

Read-only diagnostics confirmed 32 GiB installed RAM (34,139,435,008 usable bytes) and RTX 3060 12,288 MiB, driver 560.94. The internal GTX remains Code 31; no repair was attempted. G/hardware-before.json records device identity and timestamps. Power mode was Balanced. Disk headroom before work was about 133 GiB on C and 2,716 GiB on D. Background Windows, Codex, Ollama launchers and Docker remained running; background load and file caches were not controlled between dates.

The final collector's real non-inference SDK probe preserved failed-job bytes and command audit before stopping its sandbox. The protected baseline failed as intended. External docker stats monitoring introduced no diagnostic process into the worker PID namespace; the process-ownership guard remained enforced. G/collector-preflight.json, probe-container-inspect.json and storage-enforcement.json record this check and actual tmpfs exhaustion enforcement.

The actual fixture request was retained in its intent artifact. Replaying that request through the unchanged adapter without network dispatch reconstructs 25,952 serialized UTF-8 bytes plus 1,536 output and 1,024 margin: conservative capacity bound 28,512 / 65,536. System bytes 15,473; user bytes 1,308; tool schema bytes 8,947. This was admitted; it was not a context rejection. Actual prompt-token count is unavailable because no response completed. No history compaction or hidden truncation occurred. The separate SDK preflight had a slightly different diagnostic task ID and a 28,458 bound.

The native endpoint remained http://127.0.0.1:11434 with no cloud fallback. Runtime logs confirm -c 65536, -ngl 18, CPU/GPU offload on CUDA0 RTX 3060, main-model 18/31 layers and draft-model 5/5 layers on the GPU. The main allocation was estimated at 11,203 MiB: weights 9,570.05, KV 936 and compute 697.38 MiB. Draft weights added 425.34 MiB and draft compute 124.50 MiB. Runtime fitting could not meet its free-memory target with the explicit layer count; the extra-model memory fitting was incomplete, and loading continued. These logs explain why weight-size-only or active-parameter estimates would have been insufficient.

The external monitor observed 11,980 MiB GPU use, above the unchanged 11,776 MiB stop ceiling (512 MiB reserve). The controller cancelled the request; the backend logged HTTP 499 / context cancelled and the owned llama-server exited. No completed generation, CUDA device-loss or out-of-memory response was observed. Runtime context initialization was 65,536, but the model never reached a ready ollama ps entry. GPU allocation is not proof of stable inference or tool use. The test deliberately did not lower the safety reserve or retry fewer layers.

Limits remained read-only root, 2 GiB/no-swap sandbox RAM, 2 CPU, 152 MiB total writable storage, protected acceptance files, no network/host mounts, disabled container logging, 15-second commands and 16-KiB output. One fixture / conditional one UTC attempt, 12 / 8 inference-request caps, at most two within-attempt repair cycles, 30 minutes per stage and 60 minutes total. The harness additionally monitored a 64-MiB batch evidence ceiling; the unchanged controller's hard job artifact bound remains 128 MiB. No limits or approval gates were relaxed.

Comparison#

No relative speed or memory percentage is calculated: the Gemma run did not complete equivalent work. Its smaller observed host peak and shorter stopped interval are not efficiency wins. Provisionally prefer the retained Qwen27 configuration for this narrowly demonstrated task, because it has independent acceptance; Gemma's coding quality and accepted-task efficiency remain unknown. A future lower-VRAM Gemma configuration would require a new scoped authorization, new preflight and fixture; it is not proven to fit or succeed. One historical Qwen success still does not establish repeatability or general quality equivalence.

Verification, preservation and outstanding work#

G/cleanup.json records final execution failed, verification not_run, review escalated, no accepted hash, stopped owned containers, no inference reservation, no resident model and retired owned model process. The user's Ollama service continues running. G/failed-candidate-review.json independently confirms the retained tree equals the frozen baseline and the command audit is empty. Missing bytes were not replaced with a generated summary or a success claim.

New automated check: 150 installed selective/security/adapter tests passed in 28.89 s with a fresh workspace-owned temporary directory. An earlier invocation reported 63 passed / 87 setup errors because its default temporary directory was inaccessible; both logs are retained. These are the same 150 cases, not additive. The previous full regression remains 561 passed / 1 skipped and previous upgrade/package checks remain historical evidence; no runtime source change required rebuilding or reinstalling. All 85 installed/runtime/source package files were rechecked byte-for-byte. Documentation, lint, source audit and diff results are recorded in G/validation.json after the report update.

Only qualification harness/measurement artifacts and documentation changed. No downloads, dependency installations, model-store mutations or new candidate implementation occurred. Existing user services wrote runtime logs; OS/GPU cache writes were not comprehensively traced and are not claimed absent. G/inventory-after.json and the model before/after comparison show all nine model identities preserved. Qwen's isolated accepted patch remains uncommitted and unintegrated; there is no Gemma implementation patch to integrate.

Both pilots, general routing, production defaults/expiry, drivers, networking and Strata are unchanged. Human dock checks remain untested. Further Gemma inference requires a new bounded authorization; human UI checks and any pilot or production activation remain separate decisions. Nothing was committed, pushed, merged, integrated or published.

Final report gates: Ruff passed, documentation checked 45 Markdown files, public-source audit checked 222 files with no findings, and diff whitespace check passed. New local evidence plus scoped test artifacts occupied 5,803,802 logical bytes and 6,113,944 allocated bytes at the final measurement (G/storage-final.json). This excludes shared Docker backing-store attribution and filesystem metadata. Downloads and installations were zero; the existing 64-MiB live evidence stop was never reached.

Bounded MoE configuration batch — 2026-10-10#

This new authorization superseded the preceding one-shot restriction. Historical failures above remain evidence, including Gemma's loading stop and the earlier 35B collection failure. Private evidence root M is workspace data/openhands-moe-20261010/; N remains data/openhands-stronger-20261009/. No Qwen27 inference was run. Its accepted UTC worktree and evidence are preserved.

Candidate and controlled changes#

Branch codex/selective-local-execution, base 0b817c572a81a22c94988cdbf163d214957985b7, remains uncommitted. Codex performed a small direct bootstrap, not local-agent implementation: selective_config.py and inference.py now allow an explicitly configured native Ollama draft_num_predict integer from 0 through 16. Omission preserves the installed model's existing behavior. Three deterministic cases in test_selective_bounded.py cover validation, native forwarding and omission. No default, expiry, global runtime option, collector, SDK, dependency or production policy changed.

Pinned Ollama 0.40.2 source was inspected before use: appendDraftArgs returns without adding draft-model arguments when the option is zero. M/draft-source-investigation.json records source URLs, revision and SHA-256. This was read-only source inspection, not a runtime installation. Thinking disabled alone would not disable draft loading.

The live controller imports the immutable, 85-file M/runtime snapshot. M/runtime-identity.json and harness-identity.json bind source and harness hashes. The actual executor is the unchanged collector image sha256:b0ec53df107bdbb7b9df9c397ee56d4f7f74bc49d6e6550b16ca6f23e5748090, OpenHands SDK/tools 1.53.0, Python 3.13.12. No old installed LAB bridge, worker or dock is credited with these scoped SDK jobs. The existing installed LAB package remains untouched; it differs in the two runtime files above. The rebuilt wheel was extracted and tested without dependency installation.

Exact configurations and fitting decision#

Inventory metadata advertises 262,144 context and tools for the two current models; these are not live guarantees. The tested loaded context was 65,536. Active-parameter labels are not memory estimates. M/inventory-before.json retains exact metadata, manifests and asset totals. No model pull, wrapper change or retag.

The first and only new Gemma configuration used measured prior allocation: draft zero plus two fewer GPU layers (18 to 16). Based on the previous main allocation, draft removal alone appeared too close to the unchanged ceiling and was not separately tested. There was no separate model-backed tuning probe or blind sweep. Runtime logs confirm 16/31 main layers, 8,553.34 MiB GPU weight buffer, one loaded model buffer and no draft launch arguments or draft buffer. M/gemma-draft-absence.json records those checks. The selected configuration passed its fixture; no remaining tuning configurations were spent on speed hunting. Context/output were not reduced.

Qwen35 retained its previously tested configuration, including draft generation. Runtime logs confirm 14/42 layers, 6,910.82 MiB GPU weights, 13,790.11 MiB mapped host weights, 384 MiB GPU/896 MiB CPU main KV, and an MTP context against the same model with 128 MiB GPU KV. This is not a second full model weight allocation. ollama ps reported CPU/GPU offload (68%/32% for Qwen35); physical-GPU activity and UUID in external samples identify the RTX 3060, rather than relying on ps alone.

Preflight, isolation and admission#

M/hardware-before.json reconfirmed 32 GiB installed RAM (34,139,435,008 usable bytes), RTX 3060 12,288 MiB, driver 560.94. The internal GTX remains unavailable; no repair was attempted. Balanced power mode, Windows/Codex/Docker and personal Ollama background processes remained. No devices, network or power settings changed.

Before model requests, M/collector-checks.json exercised actual SDK/container normal completion, failed-test bytes/trace/diff, cancellation and evidence-limit failure using a separate synthetic transport. Those were zero-inference control probes, not model qualification. M/timeout-output-controls.json verifies bounded command timeout/output exhaustion. M/storage-enforcement.json records actual read-only-root denial and ENOSPC at the configured tmpfs capacities. Wrong-owner operations were denied. Collection precedes stop on normal and failed jobs; incomplete collection never becomes success. External monitoring never added a process to the worker PID namespace. The ownership guard was unchanged.

Sandboxes retained 2 GiB/no-swap RAM, 2 CPU, 152 MiB total writable tmpfs, disabled container logging, network none, no host mounts/credentials/Docker socket and protected acceptance files. The trusted host controller mediates only declared files, bounded editor calls and exact commands; inference uses http://127.0.0.1:11434, without cloud fallback. Host inference is measured outside the sandbox cap. Controller source retains no-redirect/no-environment-proxy behavior. No shared service was restarted or reconfigured.

The GPU ceiling remained 11,776 MiB. Host admission requires 6 GiB available; hard stop is below 2 GiB, or below 3 GiB for 15 seconds. Each stage allowed at most 30 minutes, 12 fixture requests / the stricter 8 UTC requests, two error-recovery cycles, 15-second commands, 16-KiB command/output bounds and no transport retry. The batch allowed 120 live minutes. The overall approved preparation ceiling remains 2 GiB. A 256-MiB current-batch reservation includes tests, packages and evidence, with a separate 16-MiB metadata reserve; no evidence was deleted. M/preparation-reservation.json records the accounting and pre-live reservation.

M/sdk-request-identities.json proves the same SDK system-message and four tool schemas as the retained Qwen27 run (system SHA-256 f123e64b8a1394cdd48f4ead76d0e1607498658ba4b24952a4f3dbfe30d11a61, tool-schema SHA ab12adf1ddb841b5541559bda3b842534c28f1c9ae45170055a104cd2e0d517f). Both new models received the same existing tool-policy clarification about exact paths/commands; Qwen27 historically had no extra clarification. It supplied no solution. Frozen files, task semantics, protected tests, skill version and test visibility matched. Prior accepted patches and solution-bearing transcripts were never mounted or included in either model's requests.

Admission counts each actual serialized request, including messages/history, tool schemas/results and overhead, with the unchanged conservative UTF-8-byte upper bound, 1,536 reserved output and 1,024 safety margin. No aggregate-token count is used as per-request context. No hidden truncation/compaction, context shift or weakened margin was introduced. Native strict context and loaded-context checks remained enabled. Receipts retain per-request accounting.

Gemma outcome#

The genuine SDK fixture completed 12 requests, actual reads/editor writes/tests and finish. Only subject.py changed; protected tests passed in the worker, controller and fresh independent sandbox. Two denied calls (relative editor path and listing) were recovered within budget; four test invocations were redundant. Codex reviewed the retained diff/trace and accepted exact hash 7c6bf489ebd6f25c5fafc06c848760cb990e8793b5e03c57ff0df72d2eb47955. M/gemma-g16-d0-c64-fixture-acceptance.json is the authoritative scoped acceptance.

The fresh UTC task failed and was not accepted. It read the file, recovered from a relative-path denial, ran the protected failing baseline and read the test. The fifth response reached 1,536 generated tokens; backend logs show about 304 seconds for that response. LAB rejected its incomplete completion. No edit was applied. M/gemma-utc-failure-review.json independently confirms retained failed-tree bytes exactly equal the frozen baseline. Collection completed, commands/output and the rejected request's accounting survived, then the owned executor/model residency stopped. There was no resource stop or context rejection. The rejected request bound was 41,104 / 65,536, including output and margin. This is an output-budget/model-behavior outcome for this configuration, not proof that Gemma cannot code or that MoE cannot fit. No Codex repair or additional UTC retry occurred. Rejected completion text and full provider token counters were not retained; the four successful receipts are explicitly partial usage, with 1,536 additional generated tokens observed separately in the backend log.

Qwen35 outcome and independent acceptance#

The fresh fixture completed six requests, with actual read/edit/test calls and complete collection. The new fixture candidate 9b78e8fa590e63cb4b28c83591d258331d2f48a4f501bb68a7fb288fd5fb2bfb passed the worker, controller and fresh independent verifier. Codex reviewed and accepted it before the UTC job was dispatched. This resolves the missing end-to-end evidence for this configuration; it does not retrospectively turn the old collection-failed attempt into success or prove its transient PID cause.

The UTC task also completed six requests and complete collection. It read source and protected tests, used the existing SDK think tool, demonstrated the baseline failure, edited only timestamp(), reran acceptance successfully and finished. The patch assigns UTC only to parsed values with no tzinfo, preserving aware values and ValueError behavior. The exact reviewed candidate hash is 2ff883846f53fb650d6b32755481266b833174293cccaf7529207cb451703685. Protected Linux tests passed again in a fresh restricted sandbox; 10 telemetry tests passed in the isolated candidate worktree, with the import path checked. Protected files and surrounding AST were unchanged. There was no Codex code repair or implementation assistance.

M/qwen35-g14-c64-utc-acceptance.json records the authorized supervisor decision, coverage and exact accepted hash. Execution finished, verification passed, review accepted and integration remains candidate, separately represented. M/qwen35-utc-disposable.patch and the repository worktree retain the generated patch for review; neither it nor Qwen27's patch was committed or integrated. The model's requested 30-second command argument did not expand the controller's 15-second enforced cap. The explicit SDK think tool is distinct from the native backend thinking flag, which stayed false; its time/tokens are included.

Accepted work and total qualification cost#

All times below are whole controller stages: baseline, loading, inference, SDK/tool work, collection, automatic verification and cleanup. They exclude the separately timed Codex review. Whole time to supervisor acceptance cannot be compared fairly because historical Qwen27 active/review timing is unavailable. Peak host RAM includes Windows and all processes; GPU is the whole RTX device. Peaks are sampled lower bounds, not exact allocator maxima.

The new batch used 1,845.316 seconds (30m45.3s) of the 120-minute ceiling, including all loading and the failed Gemma UTC work. No extra model-backed probe, second Gemma configuration, infrastructure retry or repeated UTC attempt occurred. The old Gemma 97.748-second failed loading attempt is separate historical cost, not erased or included in this newly authorized batch. Earlier 35B failures likewise retain their original evidence. Non-inference bootstrap/preflight, regression and reporting time is separate and is not a model-efficiency claim.

Matching accepted UTC execution: 630.481 versus 507.223 seconds is 24.30% longer for Qwen35. Matching combined stages: 975.839 versus 879.951 seconds is 10.90% longer. UTC sampled host peak 30,503,813,120 versus 27,897,638,912 bytes is 9.34% higher; whole-device VRAM 8,599 versus 10,530 MiB is 18.34% lower. These are observed configuration-level differences, not controlled model-family causal effects. Gemma's failed UTC elapsed/host peak is not an efficiency win.

The two current models and baseline used initially empty model residency at each stage, without flushing OS/file caches. Provider load time is separately reported inside stage time. No model was left competing with the next one. Tokenization, response lengths and speculative decoding differ; tokens/second cannot select a winner. Qwen35's accepted UTC output throughput was 6.581 tokens/s versus Qwen27's 2.153, yet its complete UTC execution took longer, including its long think call. Gemma's incomplete fifth UTC response is absent from the four successful receipt counters; its backend log separately records 1,536 generated tokens and 304.037 s of prompt/generation. Full rejected-request input/usage remains unavailable.

Controller inference intervals, including load, were fixture/UTC: Gemma 264.604 s / 176.638 s for completed receipts only (UTC excludes the failed final response); Qwen35 281.345 / 570.283 s; Qwen27 318.839 / 445.857 s. Per-command outputs/exits and elapsed times remain in worker-result or failed-commands. New collection freeze/read/stop phase timestamps are in controller-phases.json; these include nested intervals and repeated mailbox reads, so must not be added as independent stage costs. Historical collection/editor overhead was not separately measured. Exact tool and verification timings are retrievable there.

Error observations are the enforced bounded-recovery counter, not necessarily bad generated patches. Gemma fixture had two denied calls; UTC had one denied relative path plus the deliberately failing pre-edit test. Qwen35 fixture had zero and UTC had one pre-edit failing test. Qwen27 historically had zero/one (denied directory view). No malformed accepted tool call was observed; Gemma's incomplete response is recorded separately. Neither accepted Qwen UTC patch needed Codex repair or a model repair of incorrect implementation code.

The same admission bound spans were 28,827–34,252 (Gemma fixture), 28,841–41,104 (Gemma UTC including rejection), 28,804–31,580 (Qwen35 fixture), and 28,818–46,369 (Qwen35 UTC). All include the output reservation/margin and fit 65,536. Per-request actual token receipts remain distinct from cumulative usage. No capacity rejection occurred in this batch.

New sampled intervals averaged about 8.9–9.5 seconds including external Docker statistics, versus about 4.0 seconds historically. Different monitor overhead, background load, file caches, sampling defaults, model quantizations and the small policy-guidance difference limit the comparison. Host before/peak values above are not model-only allocation. Peak llama-server RSS was Gemma 15.178 / 12.434 GiB and Qwen35 16.706 / 16.903 GiB; Qwen27 was 14.539 / 14.959 GiB. RSS/shared memory must not be added to whole-host or GPU totals. New controller peaks were 73.9–78.3 MiB; sandbox peaks about 256–274 MiB, separately measured against its enforced 2-GiB cap. Host inference exceeded some planning estimates; measured stop guards, not those estimates, established observed headroom.

No current stage hit GPU/host stops, device loss or a transport failure. Qwen35's minimum available host RAM was only 3.386 GiB in UTC, versus Qwen27's 5.813 GiB; its host-memory margin is smaller despite lower VRAM. Windows paging rates are unavailable: psutil's zero swap-I/O fields do not prove no paging. Balanced power mode was unchanged; GPU power/temperature/raw clock-reason samples and process CPU observations are retained in M/resource-detail.json. CPU throttling and a matched historical power series were not established. Existing duplicate personal Ollama launchers emitted bind warnings; the selected service stayed reachable and no unrelated service was terminated.

GPU-only instantaneous power is not whole-system energy. Energy, electricity cost, cloud input/cached/output tokens and Codex token savings remain unavailable. No quality-equivalence or savings claim follows. Recorded review wall intervals were 42.212 s for Gemma fixture, 28.930 s for Qwen35 fixture and 63.132 s for Qwen35 UTC, from explicit inspection through fresh checks/acceptance; they are not active CPU time or cloud-token cost. The historical comparison lacks this measure, so those durations are not folded into a misleading relative total.

Both Qwen configurations met the same acceptance standard on this one task. Provisionally prefer the retained Qwen27 configuration for similar scoped work on this hardware: shorter accepted UTC execution and more host headroom. Qwen35 is a separately demonstrated alternative when lower VRAM matters, with higher host use and longer observed task time. Gemma now has a safe, accepted fixture configuration but no independently accepted useful UTC implementation. Further Gemma changes would be a new scoped experiment, not a reason to relax acceptance or imply MoE is unsuitable. None of these single runs establishes repeatability, general quality equivalence, production routing or pilot readiness.

Validation, preservation and next actions#

New bootstrap validation: 564 passed / 1 skipped in the full source suite (734.37 s; Windows symlink privilege skip), 41 focused bounded cases and 153 extracted-wheel selective cases (29.73 s). The focused and wheel cases overlap the full suite; never add them. The separate 10 telemetry tests above were run against the accepted isolated UTC candidate and are not another full suite. Three new cases cover the optional draft setting; no prior tests were removed. Actual copied v0.4.1-wheel database seeding and candidate migration preserved usage/checkpoints. M/bootstrap-validation.json records exact scopes.

The tested wheel SHA-256 is 89c27489900befa6fa14ea7f245b5ed438ab439544cf9d34bd697599e3d14960. It contains the same 85 byte-identical package files as source and live runtime. No package/environment was installed. M/final-package-identity.json records the report-final source manifest/sdist and lint/docs/audit/diff gates, distinct from the prior installed package. Runtime was not altered during owned execution. Rollback is to omit the new optional setting/use the prior retained wheel; no new database schema migration was introduced by this bootstrap. No rollback or installation was performed against the user's working service.

M/cleanup.json verifies stopped owned executors, released inference reservations, empty model residency, preserved nine-tag identities/model-store snapshots, unchanged original checkouts and byte-identical accepted Qwen27 evidence/patch. The personal Ollama service remains running. Only persisted qualification configuration records remain; temporary request overrides ended with their owned jobs/residencies. Logs and OS/GPU caches may have incidental runtime writes; no model-store mutation was detected, and cache writes were not comprehensively traced. M/storage-final.json gives measured current-batch preparation/evidence allocation under the unchanged reservation; no retained evidence was deleted.

Both pilots, general implementation routing, production defaults/expiry, networking, drivers, global Python and Strata remain unchanged. All human dock checks remain outstanding. The feature/bootstrap/report changes and both accepted Qwen UTC patches are isolated and uncommitted; nothing was merged, integrated, pushed or published. Any future Gemma retry/tuning, pilot activation, global routing, integration or publication needs separate authorization. No additional installation is required for the exact scoped Qwen35 path just demonstrated.