EnvLoop Native Control Validation - 4 October 2026

This prerelease publishes an English environment-validation report and the corresponding evidence snapshot and figures.

Magento completed 20 original selection control trios. Odoo completed 100 original final-candidate control trios, with the original reference versions kept separate. Desktop and GitLab each completed three original selection trios at the 05:30 UTC snapshot. Root and independent saved audits rederived the scores, resets and provenance.

The two-page PDF was rendered and visually inspected. It explicitly records zero admitted formal identities, researcher chains and matched model outcomes. Tinker is assumed unavailable. The new private diagnostic adapters are not yet in the public runtime package; clean-checkout reproduction of the new repairs remains pending.

The earlier runtime and build artifacts remain available in v0.6.11-native-reference-validation. This release does not claim a completed empirical benchmark. Credentials, account identifiers, private task bodies and hidden allocations are excluded.

Native reference validation and non-training runtime

This prerelease publishes executable non-training infrastructure and a dated English environment-validation report. It does not publish full-study model results or measured training gains.

The GitLab reference continuation preserves original trio provenance, verifies fresh worker/native lifecycle ownership, and uses one durable supervisor shutdown sequence. Native146/V3 input, scoring, reset and per-task budgets are unchanged. Public standalone tests explicitly skip only the unavailable private historical audit.

The 4 October report records the first original Desktop selection control trio, current Odoo selection20 controls and separately audited Magento continuation evidence. Its progress visualization reports environment checks, not model performance. Odoo original100 and remaining application qualifications are still running.

The actual amd64 CI build, network-isolated smoke and registry push passed at https://github.com/EnvLoop/cua-rsibench/actions/runs/37167360829. Root independently reopened all 1,731 archived source files plus both build files. Tinker is excluded; build and current reference controls make zero model/training calls. Anonymous container-image pull remains unverified under organization package policy; the source archive is public and independently buildable.

Official study totals remain 0/600 final identities admitted, 0/24 researcher chains completed and 0/3,000 matched model outcomes. Credentials, account data, private tasks, hidden allocation and gold state are not release assets.

Current native application controls and original selection preparation

The actual current-source public Calc, Impress and Writer controls each passed 0/1/0. Independent audits reopened original saved verifiers, collateral state, paired native images/receipts, exact resets and 18 unique closed guests. These are public TRAIN environment controls, not model outcomes or complete-cohort qualification. Failed attempts and older-source controls receive zero new credit.

The runner also includes the exact original 20 double-target Desktop preparation and the proved GitLab native CSV preview repair. Actual Linux amd64 build, network-isolated smoke and registry push passed. All 1,725 archive files were independently reopened against the manifest. Current-source CLI checks construct no guest or model lane; Tinker is excluded.

Original Magento and GitLab selection20 control runs remain in progress. Original Desktop 20 selection/100 final controls, all seven actor paths, Office lifecycle and the full study remain pending. Formal totals are 0/600 admitted final identities, 0/24 researcher chains and 0/3,000 matched model outcomes. The earlier English PDFs and prior releases are preserved separately. Public package visibility remains organization controlled; anonymous image pull is not claimed.

CI: https://github.com/EnvLoop/cua-rsibench/actions/runs/37126265921

Non-training runner: original selection URI repair

The non-training runner now includes the original GitLab selection URI repair. Actual Linux amd64 build, network-isolated smoke and registry push passed; all 1,722 source files were independently reopened and matched their manifest. The smoke checks the repaired reference binding and CLI without application construction. Tinker is excluded and model/training request counts are zero.

The English PDF is the preserved 3 October readiness snapshot with an architecture visualization; all three pages were rendered and visually inspected. It describes verified controls and remaining work, not a model-performance paper. The earlier 2 October report and prior releases remain unchanged.

Original Magento and GitLab selection20 control runs are in progress. Complete cohort source/control qualification remains pending. Formal totals remain 0/600 admitted final identities, 0/24 researcher chains and 0/3,000 matched model outcomes. The GHCR image remains subject to organization-controlled package visibility; anonymous image pull is not claimed.

CI: https://github.com/EnvLoop/cua-rsibench/actions/runs/37122584600

Verified non-training selection readiness

Verified non-training source build: new original selection-reference entrypoints, bounded Office picker readiness, and the completed fresh Impress native 0/1/0 control trio.

The actual CI build, verification without network/provider keys and registry push passed. All 1,720 archive files were independently reopened. Tinker is not installed and model/Tinker calls were zero. Stable runtime defaults are preserved.

GitLab and Magento bind their own original 20 selection identities, original independent saved-state/reset verification and explicit reference-only authority. Their application 20-task runs and 60-case/source-pixel qualifications are UNEXECUTED. Offline checks and saved-prerequisite audits do not admit tasks or models.

Current Impress controls genuinely scored 0/1/0, with six distinct closed guests and three exact original resets. This does not establish complete native multi-application/cohort qualification. Office readiness imports/tests and actual native inventory passed, but upload stopped at Mac lock before file-path/Open input; its full lifecycle remains pending.

Full study remains unrun: 0/600 final admissions, 0/24 researcher chains and 0/3,000 matched model outcomes. Existing historical releases and dated English methods PDF remain unchanged. See current readiness. GHCR visibility remains private; build from the public archive. No account state, private task allocation, credentials or checkpoint is included.

Verified non-training native build

This prerelease contains a verified code-only build and independently checked native controls. It assumes Tinker is unavailable and contains no measured model improvement or complete-study result.

The actual amd64 Docker build and verification without network access passed in CI. All 1,712 archived source files were independently reopened. The archive adds the scoped Impress selection/focus implementation; the runner retains its stable V50 default. Tinker is not installed, and model/Tinker calls were zero.

Public receipts cover the complete original Odoo selection20 reference controls, GitLab native source rendering/restoration, Magento native source-read/Close/reset, and a fresh Impress saved positive/distinct reset. These controls retain their own source epochs. Full native control trios, original Office lifecycle, complete cohort admission and model results remain pending. Full-study counts remain 0/600 final admissions, 0/24 researcher chains and 0/3,000 matched outcomes.

The existing dated English build methods PDF is preserved as a 2 October document. See current readiness for newer evidence.

GHCR package visibility remains private under the existing organization policy. Build from this public source archive; anonymous registry pull is not claimed. No private task allocation, account state, credentials or model checkpoint is included.

Verified non-training build and native controls

This prerelease builds the benchmark controller without Tinker. It is a build and native-control release, not a trained-model result or a completed six-environment study.

The source commit is 8b06b6cb748730fd2133f9efad8291bc8cf9092b (merged PR #13). The actual CI run passed the amd64 image build, network-disabled container checks, registry push and source-archive upload. All 1,704 source files were independently reopened and hash-verified.

The attached aggregate records the current Writer 0/1/0 trio, GitLab TRAIN 1/0/1 native action/reset prerequisite, and CRM autosave 0/1/0 calibration. Original guards, scorers and budgets remain unchanged. Complete source rendering, Impress and Office native qualification remain pending. Formal study counts remain 0/600, 0/24 and 0/3,000.

No provider keys were supplied to the image. Tinker is absent, and model/API calls are zero. The GHCR package remains private under the existing organization policy; build from the public source archive without registry access. Private task allocations, account state, gold and credentials are excluded.

The original dated English methods PDF is retained unchanged. It is not relabeled as a new empirical result paper.

Current original-software runtime without Tinker

This prerelease contains the current original-software runtime and reproducible source build without a Tinker dependency. It does not report completed benchmark model results.

The fresh GitHub amd64 image build and network-disabled smoke passed at commit 49e13c3. The downloaded archive was independently reopened and all 1,672 source file hashes matched its manifest. Current bindings include Desktop V48, Office Host6, Odoo Native14/Reference6, GitLab Native14 and Magento Native10.

A genuine Calc saved positive and a different fresh reset passed the original verifier, preserving 63 non-target cells. Odoo retains 50 completed original purchase/inventory control trios, exact resets and the first sales negative failure. Office's real TRAIN upload, two equivalent downloads and separate cloud reset passed; the bound Host6 full-inventory operation passed before macOS lock stopped the native picker. These environment controls are separate from full native qualification and model results.

Validation: 3,362 controller tests (64 conditional skips, zero failures/errors), seven compatible Desktop SDK tests, 23 targeted Desktop tests and 58 current Node tests. No model or Tinker calls were made; Tinker is absent from the controller image.

The GHCR image remains private under the organization package policy. The public source archive includes its Dockerfile and can be built independently. The English methods PDF in the previous controller release is a dated 2 October build report, not a completed empirical paper.

Full-study final admissions, trained researcher chains and matched model outcomes remain 0/600, 0/24 and 0/3,000. See the repository's current build status and dated execution notes for pending gates.

EnvLoop Computer Use: Non-Training Controller Build

Non-training controller build. Tinker is not installed and the build makes no model requests.

Includes source for six application cells, original rich-data generators, independent artifact/SQL/Git verifiers, reset/cleanup paths, Harbor/E2B integration, Chromium/Docker clients, a three-page English methods report and architecture diagram. The actual container verified all 1,644 file hashes, source bindings, CLI entrypoints, browser input/screenshots and artifact readback without network or provider keys. Public source and report text are English; credentials, accounts, private allocations and gold answers are excluded.

Desktop observation now uses one verified before/after capture pair plus the unchanged current predispatch pair across all seven roles. This explicitly replaces the earlier seven-capture readiness policy; current ownership, read-only/focus checks, action guard, nonces, Enter behavior and budgets remain unchanged. Seventy-seven focused checks passed. The real new control applied five GUI actions, but saved bytes remained the original baseline and the next window-ownership check failed. The cause is unconfirmed; the guest was closed. No positive save, distinct reset or full native qualification is claimed.

Magento's fresh original-software TRAIN trio independently passed 0/1/0 scoring, queue timing, reset and cleanup. Office's 140 rich SEC Excel packages and 140 portable WDI presentation packages passed offline checks; its native upload path remains blocked by the locked local picker/disabled file bridge. No second account is required.

The source build is separate from the empirical benchmark: official final identities, researcher chains, model outcomes and measured training gains are all zero. GHCR public visibility is disabled by organization policy; build from the public archive instead of relying on anonymous image pull. Historical sources, failed controls and earlier release receipts are retained without promotion.

EnvLoop Computer Use: Verified Non-Training Runtime

Executable non-training source/controller refresh. Tinker is absent from the runner and no benchmark model request was made.

The public bundle contains application adapters, rich-data generators, independent verifiers, reset/cleanup code, Harbor/E2B integration, Chromium/Docker tooling, an English three-page methods PDF and an architecture diagram. All 1,638 staged source files passed hash verification in an actual network-isolated container; source text and PDF are English, and private account/task/gold state is excluded.

A fresh original Magento TRAIN control trio passed independent baseline 0 / positive 1 / wrong variant 0 verification, native queue timing, exact reset and owned cleanup. Desktop capability/virtual-cell fixes are packaged at the tested v44 epoch; a genuine formula input was applied, but commit/save/readback qualification remains pending. Office has 140 rich SEC Excel packages and 140 portable WDI presentation packages; its current native upload lifecycle remains externally blocked. Retained GitLab source/sidebar visuals were reviewed with bounded rendering findings.

Validation: 54 focused privacy, principal and Desktop tests; the earlier 3,232-test regression profile plus seven compatible Desktop SDK tests; actual rebuilt controller smoke without network/provider keys; all three PDF pages visually checked. These are build and reference-control checks, with zero official final identities, researcher chains or matched model outcomes.

GHCR push is separate from public visibility: organization policy currently disables public package visibility, so anonymous image pull has not passed. Build from this public source archive using deployment/non-tinker-runner/README.md. The complete trained-versus-base benchmark paper remains a later phase.