Aperture 0.4.7: source-pinned Recipe Lab

Aperture 0.4.7 adds a source-pinned Recipe Lab that translates published inference experiments into capability-gated compatibility records for the hardware Aperture actually observes.

Run the released package without Git:

npx.cmd --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.7/bigbirdreturns-aperture-0.4.7.tgz aperture recipes --allow-scan

On Linux or macOS, use npx instead of npx.cmd. Node.js 20.11 or newer is required. Distribution remains through GitHub rather than the npm registry namespace.

What changed

aperture recipes performs the existing permissioned local hardware scan and compares it with structured experiment recipes. The first catalog is pinned to local-inference-lab/rtx6kpro commit 3023e7c2e572cd445cd62234607aaf765121da58. Aperture records selected factual observations and exact source coordinates; it does not vendor the upstream prose, scripts, images, Dockerfiles, or executable material.

The first seven records cover Qwen3.8-27B official FP8 TP1 and qualified TP4 profiles, Qwen3.5-397B-A17B NVFP4 TP4, Qwen3.5-122B-A10B NVFP4 TP2, MiniMax-M2.5 NVFP4 TP2, GLM-5 NVFP4 TP2+PP3, and Kimi K2.5 native INT4 TP8.

Each result is QUALIFIED, CANDIDATE, UNKNOWN, or BLOCKED. The classification separates hard requirements from source reference conditions. A known numerical floor such as an SM120-class native NVFP4 path can block older observed hardware. A reference 96 GB-class card does not become a fabricated minimum. Approximate source working-set figures remain reference evidence; smaller local hardware is held as UNKNOWN until exact checkpoint/context fit or another bounded canary resolves the lower boundary.

Current free VRAM is also separated from physical capability. A machine can remain a hardware candidate while its present headroom is below the source observation. Multi-GPU recipes preserve source GPU count and homogeneous-cohort assumptions rather than treating arbitrary mixed cards as an equivalent tensor-parallel system.

Hardware observation extension

When the installed NVIDIA driver exposes the fields, the scan now adds compute capability and current PCIe generation/width to the existing capacity observation. The detail probe is additive: failure does not erase the original device, VRAM, or driver inventory. GPU peer-to-peer bandwidth and NUMA locality remain explicitly unmeasured.

The recipe command accepts only --allow-scan, --home, and --out. It does not read model files, contact a model host, download weights, install a runtime, execute inference, stop another workload, or run a benchmark. A recipe classification is compatibility evidence and a next-canary decision, not reproduction of the upstream throughput or quality result.

Package identity

The package-bearing source commit is 86e0bfbe4defeff06b521fa0508c1a2c58466713, tree bb3944ed4d99f9034d06eddd465fc9c774fd05cc. Two candidate builds produced the same bigbirdreturns-aperture-0.4.7.tgz:

• Package bytes: 80,905
• Package SHA-256: 4e614815b7ce0453081b493d800bd1d5aa2f2e4e6f68f5c3886376ea88723ee7
• Source/control tests: passed
• Installed-package smoke: passed
• Candidate builds byte-identical: yes

Publication rebuilds the package twice from a successor containing those exact packaged inputs and refuses any byte or digest difference.

Retained capabilities

Release 0.4.7 retains the permissioned support receipt from 0.4.5, the native pre-download GGUF fit gate and physical-capacity split execution from 0.4.4, and exact managed-GGUF cache reuse from 0.4.6. Model inspection, acquisition, runtime installation, and execution keep their existing independent approvals.

Limits

This release establishes the recipe contract and one seven-record source catalog. It does not claim exhaustive coverage of the upstream wiki, implement vLLM or SGLang distributed serving, pool VRAM, measure GPU peer-to-peer bandwidth, infer NUMA suitability, automatically download the referenced checkpoints, reproduce upstream throughput, validate upstream model quality, or prove that hardware smaller than a reference machine will fit. QUALIFIED preserves a source qualification only when the locally observed hard/reference conditions are present; an actual Aperture load and performance canary remains separate.

Aperture 0.4.6: verified managed-GGUF cache reuse

Aperture 0.4.6 separates return use of an exact managed GGUF from permission to contact the model host or transfer weight bytes.

Run the released package without Git:

npx.cmd --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.6/bigbirdreturns-aperture-0.4.6.tgz aperture

On Linux or macOS, use npx instead of npx.cmd. Node.js 20.11 or newer is required. Distribution remains through GitHub rather than the npm registry namespace.

What changed

A saved remote GGUF now checks its exact Aperture-managed acquisition entry before model-host metadata or weight-transfer approval. The cache identity binds the selected source, immutable revision, representation, component names, declared sizes, and expected hashes. Aperture admits only the matching regular files, then completely hashes every selected component and repeats native fit before loading.

An incomplete cache, missing component, symbolic entry, size drift, or identity drift remains a cache miss and follows the existing permissioned metadata, fit, and resumable-acquisition path. --allow-download authorizes downloading or resuming missing bytes; it is not required merely to reuse a complete matching managed artifact. Runtime installation and model execution keep their separate approvals. Safetensors remain on the existing verified acquisition path and do not receive this shortcut.

Native cache-reuse passage

A separate Windows x64 host first populated Aperture's managed cache with Qwen2.5-0.5B-Instruct Q4_K_M at revision 9217f5db79a29953eb74d5343926648285ec7e67. The selected artifact was 491,400,032 bytes with SHA-256 74a4da8c9fdbcd15bd1f6d01d621410d31c6fc00986f5eb687824e7b93d7a9db; the request retained 2,048-token context and one sequence.

The executable candidate resumed the original saved remote selection without --allow-network or --allow-download. It reported no model-host request or weight transfer, repeated complete integrity and native-fit preparation, generated FINAL CACHE PASS on CUDA with 25 observed GPU layers, exited 0, released the selected GPU, and left no candidate-owned process. The attached path-free receipt excludes local paths, GPU UUIDs, and private run records.

Package identity

The package-bearing source commit is cfbc8222f932e43a6d4397cbc2ce4f1cb8131d97, tree 682c302072e5def10049f8495df565701c7cb51e. Two candidate builds produced the same bigbirdreturns-aperture-0.4.6.tgz:

• Package bytes: 73,541
• Package SHA-256: e19df7c51b6111215cbb1ca26562e405cf2c053dada49c1e340cbe8fe6670196
• Source/control tests: 169/169
• Installed-package smoke: passed

Publication rebuilds the package twice from a successor containing those exact packaged inputs and refuses any byte or digest difference.

Retained capabilities

Release 0.4.6 retains the 0.4.5 reduced hardware support receipt and the 0.4.4 physical-capacity result. The latter ran a Qwen2.5-14B Q4_K_M checkpoint whose tensor payload exceeded the selected RTX 4060's reported physical memory through one qualified CPU/GPU split at 2,048-token context.

Limits

This release establishes one exact managed-GGUF return-use path on one Windows x64 host. It does not authorize arbitrary cache imports, establish that checksum sidecars alone prove provenance, add a global cache search, add NPU numerical execution, qualify native Mac inference, implement distributed inference or pooled VRAM, broaden safetensors execution, prove model quality, or establish a general offline guarantee. Direct sources without publisher hashes retain their previously documented identity limitations.

Aperture 0.4.5: reduced hardware support receipts

Aperture 0.4.5 adds a permissioned support-intake path so a hardware, fit, or backend failure can be reported without attaching the raw machine scan.

Run the released package without Git:

npx.cmd --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.5/bigbirdreturns-aperture-0.4.5.tgz aperture

On Linux or macOS, use npx instead of npx.cmd. Node.js 20.11 or newer is required. Distribution remains through GitHub rather than the npm registry namespace.

Reduced support receipt

Create a receipt with:

aperture support --allow-scan --out aperture-support.json

The command runs the same approved hardware inventory used by planning and writes a separate aperture-support/1 JSON object. It retains CPU topology, RAM and allocation headroom, integrated and discrete graphics, NPU inventory, driver versions, storage classes and capacities, external-link observations, installed-runtime presence, provider status, and explicit unmeasured fields.

It excludes host and user names, local paths and mount labels, stable device and partition identifiers, GPU UUIDs, drive product names and serials, network adapter names and addresses, model locations, prompts, generated text, credentials, and environment variables. Dynamic provider errors are reduced to bounded classes. Labels shaped like PCI addresses, PnP identifiers, MAC addresses, or absolute paths are withheld.

The command accepts scan, destination, and output controls only. It does not inspect a model, contact a model host, acquire weights, install a runtime, execute inference, stress hardware, or upload the receipt. Existing output files are not overwritten.

Privacy boundary

The receipt is reduced, not anonymous. Hardware model names, exact capacities, driver versions, current headroom, scan timing, and the receipt timestamp can still fingerprint a distinctive machine. Inspect the JSON before sharing it and remove any field that is unnecessary for the venue.

Raw scan, answer, and run records can contain paths, hardware identifiers, prompts, and generated output. The reduced receipt does not make those raw records safe to publish.

Qualification

The package contains 162 source and control tests. The installed-package smoke path executes aperture support after a clean package installation on Node 22 and 24 across Windows, Ubuntu, and macOS. It parses the resulting JSON, verifies schema and version identity, rejects forbidden identifier fields, and rejects the runner's host name, home directory, and temporary path in all string values.

The package is built twice on the release branch and twice again immediately before publication. Publication proceeds only when the resulting byte count and SHA-256 match the reviewed manifest. The workflow refuses an existing tag or release, creates a versioned source archive and checksums, and dispatches post-publication documentation verification.

Retained numerical evidence

Aperture 0.4.5 does not change the model-loading or inference path qualified in 0.4.4. The retained physical-capacity result used Qwen2.5-14B-Instruct Q4_K_M with 8,982,142,976 tensor bytes, exceeding the selected RTX 4060's 8,585,740,288 driver-reported physical bytes by 396,402,688 bytes. A qualified CPU/GPU split preserved 2,048-token context and one sequence, returned 42, and exited successfully without selecting the separate RTX 3090 for model execution.

Limits

This release does not establish anonymity, external native inference from a newly submitted receipt, arbitrary checkpoint compatibility, native Mac inference, NPU numerical execution, distributed inference, pooled GPU memory, optimal placement, task quality, or performance leadership. A support receipt is evidence about an observed machine state, not proof that a selected model will run.

Aperture 0.4.4: bounded Windows scanning and physical-capacity split

Aperture 0.4.4 carries the hardware-to-model runner through bounded Windows discovery and one qualified CPU/GPU execution in which the selected checkpoint's tensor payload exceeds the selected GPU's physical memory.

Run the released package without Git:

npx.cmd --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.4/bigbirdreturns-aperture-0.4.4.tgz aperture

On Linux or macOS, use npx instead of npx.cmd. Node.js 20.11 or newer is required. Distribution remains through GitHub rather than the npm registry namespace.

What changed

Windows CPU, memory, display, NPU, and external-link observations now run separately from extended storage and network providers, with independent bounds and one shared PnP enumeration. A slow extended provider no longer erases completed core observations.

Generation and persistent chat now protect the admitted system-memory reserve using current physical availability. Linux also retains an admission-time cgroup-v2 ceiling. Process RSS remains diagnostic, correcting a reproduced false abort caused by memory-mapped checkpoint pages.

Automatic GGUF placement retains five percent of the observed GPU budget, capped at 256 MiB, before selecting the highest estimated fitting layer count. Explicit model, representation, context, sequence count, backend, device, and fixed-layer requests remain preserved.

Qualified physical-capacity execution

The pinned three-shard Qwen2.5-14B-Instruct Q4_K_M checkpoint has 8,982,142,976 tensor bytes. That exceeds the RTX 4060's 8,585,740,288 driver-reported physical bytes by 396,402,688 bytes.
The exact release tarball was installed through a fresh npm cache and ran that checkpoint on the isolated RTX 4060 route. Native fit selected 28 of 49 layers, retained 2,048-token context and one sequence, loaded the model, returned 42, and exited 0. The separate RTX 3090 was not selected for model execution. No Aperture-owned process remained afterward.

The execution watchdog recorded a 3,492,032,512-byte minimum available-memory floor against its 1,766,779,535-byte reserve. An independent 66-sample monitor observed at least 4,034,019,328 bytes available. These measurements establish one supported CPU/GPU split, not an optimal placement or throughput comparison.

Release identity

• Merge commit: a8133bd6acbb2e57f2da63100e19314e22a08083
• Merge tree: 6a539f4c710afde0f1bc4c484d6b84bf0bb9f275
• Package source commit: 97f9f538062d663c499df064a550b91f071acfd6
• Package: 70,601 bytes
• Package SHA-256: 0c3ddb0b987543affb096b22fa6f48deb03c475aaaca654dc4a8383312accdb4

Two independent builds from the qualified Windows worktree produced the exact package above. The attached source ZIP is a reproducible canonical Git archive of the merge tree with checkout line-ending conversion disabled. The package and source ZIP have separate recorded identities; no claim is made that repacking the ZIP produces identical npm bytes.

Verification and limits

The release head passed 156 source/control tests, clean package installation, 21 public-release contract tests, 383 documentation links, the real browser interaction and responsive-layout suite, and the hosted Node 22/24 matrix on Windows, Linux, and macOS.

The arithmetic output is a narrow smoke assertion, not broad model-quality evidence. This release does not establish arbitrary checkpoint compatibility, native Mac inference, NPU numerical execution, distributed inference, pooled GPU memory, optimal placement, or performance leadership. Full machine records, prompts, GPU UUIDs, and model weights are not included in the release.

Aperture 0.4.3: verified fresh runtime installation

Aperture 0.4.3 fixes a reproduced Windows first-use stall during native-runtime installation. Managed installation commands now receive closed stdin while stdout/stderr remain visible; the interactive controller retains ownership of user input.

A public 0.4.2 first-use run stalled before its npm child began installation. With the same open controller input, a bounded npm-version probe timed out after 7 seconds with inherited stdin and exited successfully in 101 ms with ignored stdin. Two subprocess regressions cover EOF delivery and both output streams. All 135 source/control tests pass.

A corrected candidate was then installed with a new empty application home and new empty npm cache. It installed node-llama-cpp 3.20.0 (123 packages), assessed the fixed model before acquisition, downloaded the entire Qwen2.5-0.5B-Instruct Q4_K_M checkpoint, verified its hash, reassessed after integrity checking, and generated 42 on CPU. Context 2048, one sequence, and zero GPU layers were preserved; the command exited 0. Existing runtime and model files were not reused. The existing Windows and Node installation was reused, so this is not a clean operating-system image or universal hardware qualification.

Model: 491400032 bytes; SHA-256 74a4da8c9fdbcd15bd1f6d01d621410d31c6fc00986f5eb687824e7b93d7a9db; Qwen revision 9217f5db79a29953eb74d5343926648285ec7e67. The final 0.4.3 package differs from that native-tested candidate only in package/version metadata and README. A final isolated package installation also passed.

Windows PowerShell, without Git:

npx.cmd --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.3/bigbirdreturns-aperture-0.4.3.tgz aperture

Use npx on Linux or macOS. Node.js 20.11+ is required. Metadata, runtime installation, weight acquisition and inference permissions remain separate. The pre-download native-fit assessment and existing CUDA-library discovery repair are unchanged.

Source: de2e030, tree e41bd942e66c056b6582e33701649d99822ac417, identical to tested PR #8 candidate 73c2851. All six Node 22/24 by Ubuntu/Windows/macOS test/package jobs passed in run 33980544777; documentation build/link/browser checks passed in run 33980544807.

The attached npm archive is the qualified Windows-built artifact. The source ZIP is archived from the exact merge commit. Previous release tags/assets remain unchanged. No private machine paths, device identifiers or weights are included.

The new observation is CPU first-use execution. A fresh chat invocation was blocked before starting, so earlier chat observations remain historical. Positive fixed-14B execution beyond physical RTX 4060 VRAM, NPU numerical execution, native Mac inference, arbitrary model compatibility and performance superiority remain unverified or unsupported. No such claim is added by this release.

Aperture 0.4.2: native fit before download and verified execution

Aperture 0.4.2 brings native GGUF fit assessment into the published application before the complete weight-download prompt. It preserves the selected checkpoint, representation, requested context, single-sequence requirement and explicitly chosen device. Non-fitting or unavailable estimates stop acquisition and explain the measured budget shortfall.

After complete file integrity hashing, generation and chat refresh the native assessment before loading. Every experiment trial repeats the check from the original requirements. A later pre-load refusal retains earlier trial results. Native memory checks and loaded GPU-layer/context readback remain enabled. The 0.4.1 installed CUDA-library discovery repair is retained.

Windows PowerShell, without Git:

npx.cmd --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.2/bigbirdreturns-aperture-0.4.2.tgz aperture

Use npx on Linux or macOS. Node.js 20.11+ is required; the initial package has no dependencies or installation hooks. Metadata access, model acquisition, runtime installation and execution retain separate permissions. --answer-only remains a lightweight provisional answer.

Qualification: 133 control tests and isolated package-install checks passed on Windows and in the Node 22/24 by Ubuntu/Windows/macOS matrix. Hosted documentation build, link and browser checks passed. Ordinary native Windows CPU and RTX 4060 CUDA generation returned 42 using the original Qwen2.5-0.5B Q4_K_M artifact, context 2048 and one sequence; loaded GPU layers were zero and 25 respectively. A CPU chat retained ORCHID47 across two turns and exited cleanly. The packed archive independently installed through a fresh npm cache and generated 42. Existing native runtime and model files were reused, not a fresh-machine runtime installation.

The separately selected Qwen2.5-14B Q4_K_M target at pinned revision b466e1f8c07172155743e8e1307507d8a4f91fbd had 8,982,142,976 tensor bytes. All 50 layer placements failed the then-current reserved budgets on RTX 4060 Vulkan. The ordinary CLI refused before the weight-download prompt. No full 14B checkpoint was acquired and no 14B inference occurred. Execution beyond physical VRAM, NPU numerical execution, native Mac inference, arbitrary checkpoint compatibility and performance superiority remain unverified or unsupported as documented.

Source commit: 9b903fe. Its tree matches tested candidate 54558fd exactly. PR #6 and the attached minimized verification record carry the checks. Full private paths, device UUIDs, weights and unrelated prompts are excluded. Distribution uses these GitHub assets, not an npm registry publication. Prior release tags and assets were left unchanged.

Aperture 0.4.1: verified Windows CUDA startup and chat

Aperture 0.4.1 fixes Windows CUDA startup when compatible CUDA runtime libraries are already installed but absent from the worker library search path. It extends the hardware-aware 0.4.0 runner without changing the selected model or requested context.

Launch the permissioned wizard using Node.js with npm:

npx.cmd --yes --package=github:BigBirdReturns/aperture#v0.4.1 aperture

On Linux or macOS, use npx in place of npx.cmd.

After a failed ordinary CUDA probe, Aperture checks bounded installed CUDA toolkit and Ollama prefixes, validates a complete library set with the native runtime's compatibility probe, and applies the successful library directory only to child processes. No driver, machine PATH, existing engine, or checkpoint is modified. No CUDA DLLs are downloaded or redistributed by this correction. Explicit CUDA selection still fails visibly when no compatible route is found.

Windows native verification preserved Qwen2.5-0.5B-Instruct Q4_K_M (491400032 bytes; SHA-256 74a4da8c9fdbcd15bd1f6d01d621410d31c6fc00986f5eb687824e7b93d7a9db) and context 2048. CUDA generation on RTX 3090 returned 42 with 25 GPU layers. A two-turn CUDA chat retained CEDAR83 and exited cleanly. A fresh npm package cache then acquired the public GitHub candidate and generated successfully through its ordinary CLI; the existing pinned native runtime was reused.

All 108 local control tests and the clean npm archive-install check passed. Hosted run 33942154865 passed the Node 22/24 matrix on Ubuntu, Windows and macOS. The merged tree is identical to tested candidate dd5cc696e8bb641f4c34602f28e8f8d99dd8ae51. See PR #2 and VERIFICATION.md for provenance and retained failures.

Native inference on macOS, NPU execution, execution above nominal physical VRAM capacity, arbitrary checkpoint compatibility, numerical parity and performance leadership are not established by this release. The earlier partial-offload observation exceeded currently free VRAM only. Full private machine snapshots, paths, weights and prompts are not included in the release.

Aperture 0.4.0: hardware-aware model runner

Aperture now carries hardware discovery through model selection and actual local execution. This release includes CPU, integrated/discrete graphics, neural accelerator candidates, RAM/commit, disk/volume and available link observations, with explicit device binding and non-additive shared-memory budgets.

Run the published package without Git:

npx --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.0/bigbirdreturns-aperture-0.4.0.tgz aperture

Node.js 20.11+ is required. On Windows PowerShell, npx.cmd avoids execution-policy restrictions on npm's PowerShell shim. The initial package has no dependencies or install scripts. Metadata access, weight downloads, native-runtime installation and execution require separate consent.

Native observations include Windows CPU, Intel UHD 770 Vulkan, RTX 4060 partial CPU/GPU offload, automatic RTX 3090 CUDA-to-Vulkan fallback, and Linux URL-to-download-to-runtime-to-answer execution. Public GitHub npm acquisition completed real inference on Windows and Linux. The 2,104,932,768-byte Qwen2.5-3B checkpoint completed with four GPU layers and context 2048 while driver-reported free VRAM was 1,979,711,488 bytes. This is a free-VRAM observation, not a checkpoint exceeding the GPU's nominal physical capacity or an optimal-placement benchmark.

NPU devices are inventoried; NPU numerical execution is not implemented. Native Mac inference, distributed execution, pooled VRAM and arbitrary checkpoint compatibility are not claimed. A specific legacy Qwen3.5 RoPE-layout mismatch remains visible in the verification record.

97 control tests and clean package-install checks passed in the six-job Node 22/24 by Ubuntu/Windows/macOS matrix: Actions run 33941487947. Full native observations: VERIFICATION.md.

Release commit: 463aa08a0f70cf9cf34cea92a886aaa2200d1a5b. Its tree matches the tested candidate exactly. Assets were built from that Git tree, without private model files or machine snapshots. The npm registry package is not published; the GitHub commands and attached tarball are the distribution paths.