Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
Acceptance evidence matrix
Generated by packaging/acceptance_matrix.py at 2026-10-15T06:11:47Z for revision e0951ce243511597eb243018191f194af9d5f8bc.
Every row of strata-acceptance-checklist.md appears exactly once, with an explicit result. --check fails if a row is missing, so a gap is a build error rather than an oversight.
| Result | Count |
|---|---|
| passed | 46 |
| failed | 0 |
| pending-hardware | 15 |
| pending-release | 1 |
| unsupported | 3 |
| not-implemented | 2 |
| total | 67 |
Packaging and reuse
| ID | Result | Observation | Evidence |
|---|---|---|---|
| P01 | passed | The macOS .app builds, launches, serves the supplied UI from the host's origin and passes its own 8-step self-test from inside the mounted DMG. No global Node, Python, npm, pip or compiler is required: the bundle carries a pinned Node runtime. Real generation through it is NOT proven -- no Strata pack fits this Mac (16 GiB RAM, ~21 GiB free); that leg is pending hardware. | docs/work/evidence/tauri-selftest-from-dmg.json, tauri-artifact.txt |
| P02 | pending-hardware | The NSIS per-user installer target and policy are configured and the config validates against this Tauri version, but a Windows installer must be produced on Windows. Not built here. | packaging/README.md |
| P03 | pending-hardware | The AppImage target is configured; an AppImage must be built on Linux. Not built here. | packaging/README.md |
| P04 | passed | The terminal distribution is self-contained: it carries a pinned standalone CPython 3.12.15 and its dependencies. Verified from a moved directory with a PATH whose only Python is macOS's 3.9, which the launcher correctly rejects as below the 3.10 floor. Serving, status, timeskip and tune all ran on the bundled interpreter. Only the darwin-arm64 archive is built; Linux/Windows archives need a pin plus a build there. | docs/work/evidence/terminal-dist.txt |
| P05 | unsupported | No Metal/CUDA/SYCL/HIP target row is delivered as a package yet. One engine is built and qualified as a source build on this Mac (macos-arm64-metal); the other families have recipes only. No blanket cross-platform claim is made. | packaging/README.md |
| P06 | passed | A component's cache key is the digest of its artifact bytes, not its version label, so a rebuilt artifact with an unchanged version is a different entry and a cached payload is never reused on the strength of a string. The cache is never consumed by activation (the active directory is a copy), so a compatible application update reuses the installed engine instead of downloading it again. prune_cache keeps whatever the active and rollback versions refer to. |
strata_control/components.py, tests/tuning/test_components.py |
| P07 | passed | Components are acquired by kind: the manifest declares application, runtime, agent-runtime, tuning, engine or prepared-data, each with its own platform and backend gates, and strata components plan reports whether a package may be used here without fetching anything. The backend families are held to the packet's own matrix, so a Windows SYCL or Intel-Mac Metal package is refused as unsupported rather than attempted. |
strata_control/components.py, strata components backends`` |
| P08 | passed | Desktop and Terminal read the same owner state and the same model directory; the terminal's models list reads the setup-written configs that Desktop's model surfaces use. Imported files are never copied (import is preview-only by design). |
docs/work/evidence/serve-suite.txt, terminal-suite.txt |
| P09 | not-implemented | No offline bundle format exists yet. The engine build and model preparation are documented as commands, not as a signed bundle. | docs/work/QUALIFICATION-KIT.md 3.2 |
| P10 | pending-hardware | Uninstall retention is designed in (data lives outside the install directory, and the Windows installer does not delete app data), but uninstall/reinstall cannot be exercised without an installed Windows or a signed macOS build. | packaging/README.md |
| P11 | passed | Every artifact is verified against its manifest SHA-256 and declared size before anything is unpacked, and unpacking validates every member first, so a refusal writes nothing. Refused: parent segments (including backslash and nested forms), absolute paths, Windows drive letters, UNC paths, NUL bytes, symlinks and hard links, device and FIFO members, an artifact name that is a path, and any member whose resolved destination leaves the target. Setuid/setgid bits from an archive are never trusted. Each refusal is asserted to leave the destination empty. | tests/tuning/test_components.py (TestSafeExtraction) |
| P12 | passed | An interrupted transfer keeps its partial bytes for a resume; a range request that the server ignores restarts rather than appending to a prefix; a partial larger than the declared size is discarded as stale; a completed download that is short is an error with the bytes retained; and nothing is promoted to the cache until its digest has been verified, with the .complete marker written last so an interrupted install is never mistaken for a finished one. |
tests/tuning/test_components.py (TestInterruptedTransfer) |
| P13 | passed | Activation moves the displaced version aside rather than deleting it, and the rollback record is written before the activation so a crash between the two cannot leave it pointing at the new version. rollback restores the previous version exactly and refuses when there is nothing to restore, so 'rollback succeeded' and 'the component is gone' are never the same outcome. An incompatible component is refused before anything is written or activated. |
tests/tuning/test_components.py (TestActivation) |
Runtime, models and control
| ID | Result | Observation | Evidence |
|---|---|---|---|
| R01 | passed | Serving with installed inputs invokes no downloader, package manager, git or compiler: the server's import graph is stdlib plus relative modules, and the server starts and answers from the bundled runtime with no network access attempted. | docs/work/evidence/terminal-dist.txt |
| R02 | passed | Pure serving and the read-only CLI run with no agent runtime, no tuning packages, no Textual and no UI present. Verified by running the server from the standalone bundle and by importing strata_control.management and strata_cli.cli on a bare interpreter. | docs/work/evidence/terminal-dist.txt |
| R03 | passed | One owner per installation is enforced by an O_EXCL state file plus an flock held for the host's lifetime; a second client attaches over the private channel; a stale owner is only recovered after the socket is proven dead. Fifteen transport tests cover it. | docs/work/evidence/owner-transport-suite.txt |
| R04 | unsupported | The Windows owner transport is not a named pipe with an ACL. The Python standard library cannot create one, so Windows uses a loopback-only TCP listener whose authority rests on a 256-bit control token plus a private state directory. That is weaker than the checklist's stated condition and is recorded as such. Console stop and process-tree cleanup for engine children ARE implemented (serve/winjob.py, tested). The Windows branch of the transport is exercised on macOS but has never run on Windows. | strata_control/owner_transport.py, docs/work/evidence/owner-transport-suite.txt |
| R05 | passed | Preflight refuses before load: the engine rejects a model whose metadata geometry differs from the compiled shape, naming the differing key; the GGUF preflight checks shard naming, magic, version and tokenizer identity; the tuning plan refuses a dimension whose flag the engine does not accept. | docs/work/evidence/tuning-suite.txt, docs/work/IMPLEMENTATION_STATUS.md 1.12 |
| R06 | pending-hardware | The catalog refuses unknown selections and reports a reason, and the plan refuses unsupported dimensions. Whether Qwen3.6 runs at all on this backend is answerable only with the model present: the engine documents that Qwen3.6 does not run on the Mac engine, and NVFP4 is NVIDIA-only. | docs/MAC.md, docs/work/evidence/tuning-plan.txt |
| R07 | pending-hardware | The reviewed downloader is resumable and keeps verified bytes, and import is preview-only so it never moves files. The cancel/resume behaviour cannot be exercised without a model large enough to interrupt, which does not fit here. | docs/work/QUALIFICATION-KIT.md 3.2 |
| R08 | passed | Saving settings preserves unknown fields and masks secrets (the reviewed NATIVE_FIELDS table with •••••••• round-tripping), writes atomically via a temp file and rename, and reports restartNeeded rather than applying silently. The tuning job refuses to touch the active profile at all. |
docs/work/evidence/serve-suite.txt, docs/work/IMPLEMENTATION_STATUS.md 1.12 |
| R09 | passed | Readiness requires the engine's own READY line: max_context stays 0 until it arrives, so a pre-READY 'ready' is impossible, and a timeout is reported as uncertain rather than as success. Verified against a stub engine. | docs/work/evidence/tuning-suite.txt |
| R10 | pending-hardware | Stop and unload handle active and queued work explicitly in the reviewed server, and the ten-cycle orphan check is an executable procedure. It cannot be run without a loadable model. | docs/work/QUALIFICATION-KIT.md 5 |
| R11 | passed | All three dialects are served and their documented behaviour is covered by the reviewed server suite: 801 tests pass, including streams, tool fragments, telemetry and error shapes. | docs/work/evidence/serve-suite.txt |
| R12 | passed | The agent bridge cancels per session and the host owns request cancellation for serving; the queue releases only the affected slot. The end-to-end cancellation turn assertion is inconclusive against the instant mock engine and is reported as such rather than as a pass. | docs/work/evidence/serve-suite.txt, docs/work/IMPLEMENTATION_STATUS.md 1.8 |
| R13 | pending-hardware | Group allocation, reservations and readiness-before-routing are reviewed features of the server, and multi-device behaviour cannot be exercised on a single-GPU machine. | docs/work/QUALIFICATION-KIT.md 6 |
| R14 | pending-hardware | Vision and image-generation paths exist in the reviewed server and refuse when unavailable; neither can be exercised without a real model and its mmproj. | docs/work/QUALIFICATION-KIT.md 6 |
Agent, session and permissions
| ID | Result | Observation | Evidence |
|---|---|---|---|
| A01 | passed | The supplied UI drives a turn through the pinned Harness against the local Strata API worker on loopback, with no cloud route declared in the profile and therefore no DeepSeek account required. The model is the built-in mock: this proves the route, not inference quality. | docs/work/evidence/bridge-e2e.txt, integrations/harness/profile/cordis.patch.yml |
| A02 | passed | Framing is verified at the byte level: a multi-byte character split across reads survives intact with no replacement character, an SSE line split mid-JSON is reassembled, a stream ending without a trailing newline still yields its last delta, a malformed frame is skipped without losing its neighbours, and an error frame reaches the caller in the provider's own words. Content, reasoning, usage, finish and error are all carried. | apps/host/tests/stream-framing.test.mjs |
| A03 | passed | Verified end to end, and only after four defects were cleared. A scripted Strata worker emits a bash tool call; the test's own peer opens the gateway's $events stream on /api/remote.mux and grants the escalation with allowed-once. Asserted: the turn completes, the tool ran exactly once (marker_lines == 1, so a replayed effect would fail), an approval was both asked and answered with the recorded outcome allowed-once, and the worker recorded two requests while executing nothing itself -- so the tool ran in the Harness and the transcript, not the tool, went to the API. The defects cleared to get here: the session-title request consumed the scripted tool call; the call was missing a required description; the launch-URL interactions parameter is decorative and the default is answerable; and a workspace-write command is not gated, so the call now requests a sandbox escalation, which is the only path that consults the approval gate. |
docs/work/evidence/tool-flow.txt, integrations/harness/tool_flow_e2e.py |
| A04 | passed | Verified with the denial observed, not assumed. Same scripted call as A03, but the peer answers next (do not decide, delegate), which is this vocabulary's way of saying no answerer is available. Asserted: the escalation reached the gate (approval/asked present), every recorded decision is the fail-closed unavailable, zero grants were recorded, the turn settled completed, and the marker does not exist. Each assertion fails for a different way the flow could otherwise pass while proving nothing -- the tool never reaching the gate, the gate granting, or the turn hanging with the request left pending. An earlier version of this row was marked passed and has been withdrawn: in that run the approval was never consulted at all, the tool failed argument validation for an unrelated reason, and the mechanism it relied on does not exist. |
docs/work/evidence/tool-flow.txt, integrations/harness/tool_flow_e2e.py |
| A05 | passed | Verified end to end against a live turn. The scripted worker runs with a token delay so the turn is still running when Stop is sent, and the test waits for turn/start and fails if the turn was not live -- without that it passed trivially on the first attempt, because the scripted turn finished in milliseconds and Stop arrived after turn/end. Stop is acknowledged as {accepted: true}, which is admission only; the turn then settles as aborted in the durable record, which is what actually settles it, and the tool does not run. Also asserts the marker the tool would have written does not exist, so a turn that reported cancellation while letting the tool finish would fail. |
docs/work/evidence/tool-flow.txt, integrations/harness/tool_flow_e2e.py |
| A06 | passed | Sessions are durable in the Harness format, and resume re-reads them rather than replaying effects: the interrupted-turn repair synthesises closers instead of re-running tools. Verified by reading the durable session file across its concatenated frames. | integrations/harness/bridge_e2e.py, strconv note in docs/work/IMPLEMENTATION_STATUS.md 1.8 |
| A07 | passed | The same durable session is reachable from the App Host and from a CLI-started Harness with the same DSH_HOME, sharing history, model identity and workspace. The bridge reads the identical file the UI's host would. | docs/work/evidence/bridge-e2e.txt |
| A08 | pending-hardware | Two sessions run independently in the bridge (each gets its own session), but the cross-session isolation assertions (no shared tool results, no credential leakage, no shared-runtime kill) need a real model and a tool that runs. | docs/work/QUALIFICATION-KIT.md 7 |
| A09 | pending-hardware | Subagents are a reviewed Harness feature with real depth and concurrency bounds; the UI already renders child-agent frames. Nothing has been observed end to end, and the supplied UI's agent events are not emitted by the host yet. | docs/work/QUALIFICATION-KIT.md 7 |
| A10 | pending-hardware | Bounded goals are a reviewed Harness feature (maxGoalRounds, round-limit blocking). Not exercised in this integration. | docs/work/QUALIFICATION-KIT.md 7 |
| A11 | pending-hardware | Compaction and long-context handling are reviewed Harness features. The route declares the real local window (32K) rather than inheriting a cloud default, and refuses a request that does not fit rather than truncating silently -- observed. The compaction-preserves-tool-pairing assertion needs a real model. | docs/work/IMPLEMENTATION_STATUS.md 1.8 |
| A12 | passed | The reviewed task service validates schedules with DST-correct nextRun, prevents overlapping runs, and has no catch-up burst or duplicate scheduler. Covered by the reviewed test suite and by route tests. | docs/work/evidence/serve-suite.txt |
| A13 | pending-hardware | Skills resolve against the reviewed workspace/policy roots. That prompts and inference keys cannot invoke installer or owner actions needs an adversarial test against a running agent, which needs a real model. | docs/work/QUALIFICATION-KIT.md 7 |
| A14 | passed | The Strata profile declares exactly one route, on loopback, and no cloud route at all, so a local failure cannot fall back to a hosted model -- there is nothing to fall back to. The host additionally refuses to simulate in a packaged build. | integrations/harness/profile/cordis.patch.yml, docs/work/IMPLEMENTATION_STATUS.md 1.7 |
| A15 | passed | Conversations live in the browser store by design, and browser-origin export/import is idempotent with attachments and preferences retained; historical tool records are data and are never executed. Covered by the reviewed UI suite. | docs/work/evidence/serve-suite.txt |
Time-skip and tuning
| ID | Result | Observation | Evidence |
|---|---|---|---|
| S01 | passed | Real baseline and candidate directories with content-addressed checkpoints replace the timer entirely. Every checkpoint and decision persists by SHA-256 digest of its file inventory, not by a label. | tests/time_skip/test_time_skip.py, docs/work/evidence/timeskip-harness.txt |
| S02 | passed | A previously passing deterministic check that regresses rejects the candidate and the last good checkpoint stays selected; verified end to end, where the attempt was rejected and the baseline retained. | docs/work/evidence/timeskip-harness.txt |
| S03 | passed | An unevaluable check is unknown and never a pass; a run whose checks cannot justify a change ends inconclusive with needsUserReview, and a model's opinion is declared advice and cannot gate. |
tests/time_skip/test_time_skip.py |
| S04 | passed | Open/Compare/Keep/Restore operate on the actual retained checkpoint; applying previews, keeps a rollback copy, refuses with 409 when the destination changed, and restores. | docs/work/evidence/timeskip-http.txt |
| S05 | passed | Cancellation and failure retain the last complete checkpoint, attempts are bounded, and no revision can touch the workspace, so there is no tool replay or external side effect from a rejected attempt. | tests/time_skip/test_time_skip.py |
| T01 | passed | Tuning is optional and pinned: it lives in a separate module, the CLI imports it only for its own commands, and serving, status and monitor run without it. Nothing is installed by the tuning path. | docs/work/evidence/terminal-dist.txt |
| T02 | passed | The sweep uses only flags the selected engine binary accepts, holds an exclusive device lease, fixes the workload (greedy, seed 1) for every candidate, and never writes the active profile. | docs/work/evidence/tuning-sweep.txt |
| T03 | passed | Boot failure, timeout, no-tokens and correctness failure are each recorded with a status and excluded from selection; the lease is released even when the job fails; cancellation stops new trials and keeps the evidence collected. | docs/work/evidence/tuning-suite.txt |
| T04 | passed | Two measurement modes, labelled so neither can stand in for the other. protocol reads the engine's own serve protocol; api starts the real worker from the candidate's configuration and measures a streamed chat completion over HTTP -- the full path a user waits for -- with a required expectation so a fast wrong answer cannot win. Both record peak resident memory of the process tree, so a speed win bought by spilling is visible. Repeats, a paired baseline re-measurement and drift grading are unchanged. 11 tests cover the API mode, including that a reasoning-only stream, a missing usage block and a non-matching answer are refused rather than scored. Runs against the mock worker are labelled as harness tests, never as performance results. |
strata_control/tune_api.py, tests/tuning/test_tune_api.py |
| T05 | passed | Benchmark-only levers are measured but stripped from the exported args, and verify_exported() re-measures the configuration as it would actually be written; a configuration whose advantage came from a stripped lever is withdrawn. | tests/tuning/test_tuning.py |
| T06 | passed | Identity is a content digest, so a re-quantised model with the same name is a different model. The plan records the engine digest, the model digest, the digest method, the backend, the workload, the context and the machine; a completed sweep records itself and its report states how many earlier sweeps are comparable and, for each that is not, the reason in plain words. An unknown identity is never treated as a match, and incomparable history is retained rather than deleted. Twenty-one tests cover each element of the key separately. | tests/tuning/test_tune_history.py, strata_control/tune_history.py |
| T07 | passed | strata tune apply revalidates the report hash, re-checks the sweep's artifact identity against what is installed, refuses an incomplete sweep, a no-change result, material drift or a report edited after review, snapshots the previous bytes, and writes atomically; strata tune rollback restores those exact bytes and refuses if the configuration was edited since. Nothing applies automatically. The one gap is stated by the command itself: the candidate start is not verified against a running model because no supported model fits this Mac, so --verify is wired and tested but the CLI passes None and says so. 31 tests, and a refused apply is asserted to leave the file byte-identical. |
tests/tuning/test_tune_apply.py, docs/work/evidence/tuning-apply.txt |
| T08 | unsupported | No real tuning qualification exists for any backend, because no supported model fits this Mac. Stub-engine results are labelled as stub results everywhere they appear and are never presented as a backend qualification. | docs/work/evidence/tuning-sweep.txt |
Interface, security and evidence
| ID | Result | Observation | Evidence |
|---|---|---|---|
| U01 | passed | The supplied UI is used unmodified: 54 components, 10 screens, 8 locales, the design kit's tokens and motion. The tuning panel added this round uses the kit's own primitives and tokens rather than introducing a parallel style. | docs/work/evidence/serve-suite.txt, apps/host/src/components/TunePanel.tsx |
| U02 | pending-hardware | Keyboard and focus behaviour, reduced motion, long translations and OS file/font permission flows are designed in and the UI suite covers parts of them. Verifying them in a packaged Desktop on multiple OS versions needs those hosts. | docs/work/evidence/tauri-selftest.json |
| U03 | passed | The terminal's six views render and the operator flows work at full width; the reviewed monitor suite covers compact fallbacks and the terminal is restored on exit. Verified against a live worker. | docs/work/evidence/terminal-suite.txt, serve-suite.txt |
| U04 | passed | The monitor attaches over HTTP read-only, with an allow-list of exactly four read-only endpoints, refuses redirects, caps the body and never starts or stops anything; detaching leaves work alive. | docs/work/evidence/terminal-suite.txt |
| U05 | passed | Telemetry carries a sample age and a source; a missing or stale reading renders as a dash rather than as zero, unified memory is counted once, and the packaged build refuses to fabricate readings at all. | docs/work/evidence/serve-suite.txt, docs/work/IMPLEMENTATION_STATUS.md 1.7 |
| U06 | passed | Rendered artifacts cannot reach renderer or owner privileges: the window is confined to the host origin, navigation elsewhere is handed to the system browser, permissions are denied by default, and the renderer runs sandboxed with context isolation and no Node integration. | apps/desktop/src-tauri/src/main.rs, docs/work/evidence/tauri-selftest.json |
| U07 | passed | A non-loopback bind requires an API key (the reviewed server enforces it and the test suite covers the refusal); management stays on the private channel; secrets are masked in projections and never written to logs or URLs. | docs/work/evidence/serve-suite.txt |
| U08 | passed | The two claims are separated and tested separately, because a passing runtime-only test says nothing about the strict one. runtime_only_no_download is that the server fetches nothing while serving; strict_offline is that nothing in the process tree reaches the network. Strict offline is enforced through the environment every child receives (HF_HUB_OFFLINE, TRANSFORMERS_OFFLINE, PIP_NO_INDEX, DO_NOT_TRACK, emptied proxy variables and NO_PROXY=*), applied last so a profile cannot remove it by naming the same variable; through one check_url predicate for our own fetches, which allows local schemes and loopback because talking to this machine is not a download; and through a preflight that refuses a profile declaring MCP servers, a before_load command or remote image sources, naming which. Verified against a real child process. A test asserts every child_env call site passes the policy -- which caught a real leak in the engine's own environment. limits() states what is not enforced: a child that opens a socket directly is not stopped, because that needs an OS sandbox this product does not have on all three platforms. |
tests/tuning/test_offline.py, docs/work/evidence/offline.txt |
| U09 | not-implemented | Connector operations exist in the reviewed host (Cloudflare tunnel, LAN share) but the explicit-only and never-widen-on-failure assertions are not tested. | apps/host/server/cloudflare.mjs, local-share.mjs |
| U10 | pending-hardware | A thirty-minute real soak needs a running model. The reviewed suite has shorter monitor soak evidence from the terminal handover, which is not a substitute. | docs/work/QUALIFICATION-KIT.md 5 |
| U11 | pending-release | No signing or notarization, no SBOM, no notices bundle and no font/asset redistribution evidence. Signing credentials and rights evidence must come from the owner; the artifacts produced here are explicitly labelled unsigned internal candidates. | packaging/README.md, docs/work/evidence/tauri-artifact.txt |
| U12 | passed | A packaged build has no success path that uses demo responses or demo metrics: the host refuses to simulate when STRATA_PACKAGED is set and returns a typed error instead, the monitor reports none rather than invented readings, and this matrix records every failure, skip and remaining item. |
docs/work/IMPLEMENTATION_STATUS.md 1.7, this file |
Reading this table
passedmeans the check ran on this Mac and its evidence is retained. It does not mean the feature works on a machine this is not.failedis an implementation gap. It is never recorded aspending-hardware: deferring a defect to another machine is how defects become permanent.unsupportednames a combination that is not implemented, with the reason.- Nothing here is a support claim. The macOS candidate is an unsigned internal development candidate and is real-model validation pending.