Repository navigation
Releases
Prism Inference 0.1.23 development prerelease
Prism Inference 0.1.23 delivers the CPU inference efficiency work after 0.1.22: direct floating projections, compact rollback checkpoints, validated-logit sampling, masked-attention traversal, recurrent staging and bounded suffix matching. It also includes further native-Q2 resident queue work. Source reductions alone do not establish measured speed or device execution.
The Windows ARM64 bundle contains the five executables and four license files built from 055c7430cb4771dfddb24ddaa30494af6ad182da using nightly-2026-11-09-aarch64-pc-windows-msvc.
Validation: Foundation bundle run 42721996680, attempt 1 passed nine receipt/version controls, formatting, strict Release all-target analysis, portable library compilation, production linking and help smoke checks. The artifact and all nine bundled file identities were independently verified against its source-bound producer receipt. Full Foundation, actual-model and performance qualification remain pending.
Archive: prism-inference-0.1.23-aarch64-pc-windows-msvc.zip
SHA-256: bc792fdd6044ff4392a8aedbe241f634bd4492ce086d62dcca392f8d993725ee
The private server and accelerator packages remain unpublished on crates.io. The five unchanged registry helpers retain their existing versions. Applications must verify the new immutable source, archive and executable identities before selecting this bundle. Explicit representations, adapters and experimental paths retain their admission and qualification requirements.
Prism Inference 0.1.14
Fixes context-overflow handling in the authenticated private Swift service. Genuine prompt token-capacity overflow keeps its typed classification and returns HTTP 413 with context_length_exceeded before generation. Unknown failures remain fail-closed. Rejected counts invalidate stale prepared requests and permit a healthy owner to count and generate again without automatic replay.
Includes the existing Windows ARM64 and x64 CPU executables and matching integrity receipts, built from exact source 9e8ea3e2c042ab0c3f1a8e9596ff155e202c3c57 with the pinned nightly-2026-10-22 compilers. Dependencies, notices, CPU arithmetic, model/adapter identities, memory reserves and defaults are unchanged.
PTQ1 and adapters remain explicit selections. This release does not qualify an application default, full-model numerical parity, provenance, performance, quality or owner comfort.
Owning exact-source checks:
The attached archives and receipts are unchanged from the owning Foundation run. Registry publication remains separate.
Prism Inference 0.1.13
Prism Inference 0.1.13 removes repeated projection-payload validation during graph construction and adds opt-in request timing APIs plus a separately prepared diagnostic tool. CPU quantization, reduction order, production copy-then-scale arithmetic, worker protocols and default selections are unchanged. No actual-model startup or throughput gain is claimed.
This source rollup also includes explicit projection-executor APIs and experimental Hexagon helpers. The CPU executables do not depend on or enable those helpers. PTQ1, adapters and additional row workers remain explicit selections; this release does not qualify PTQ1 as a default. Registry publication remains separate.
The attached Windows ARM64 and x64 archives and matching receipts are unchanged outputs of Foundation run 41535008224, attempt 1, at source cae16fa6998dac58a947eba9ae4398ed6a9fb863. Each archive contains five executables and four license/notice files. Independent file, source, receipt and PE architecture review passed before publication.
Verified owning checks for the exact release source:
Prism Inference 0.1.12
Prism Inference 0.1.12
This full release rolls up the improvements since the last full release,
v0.1.6 (02c8a24b31ac1c35f2ffd3530f6a753ce440bdc0). It supplies the Rust library,
source tools and prebuilt Windows ARM64/x64 CPU bundles. A full release does not
make the experimental model profiles or accelerators qualified defaults.
Improvements since 0.1.6
| Area | Result |
|---|---|
| Contained startup | Exact read paths are checked inside AppContainer without requiring broader directory access; bounded startup-stage diagnostics retain no artifact paths or values. |
| Model profiles | The Swift session and private HTTP service can explicitly select the verified current-weight PTQ1 export and approved rank-one adapter. Adapter scales are bounded and bound to prepared requests; changing a scale clears incompatible live state and checkpoints. |
| CPU workers | The optional parallel-rows feature provides a persistent, explicitly budgeted packed-row pool. Private readiness reports the instantiated session's effective worker budget, allowing the owner to reject a policy mismatch. |
| Artifact verification | GGUF verification uses Windows CNG SHA-256 with a portable Rust implementation elsewhere. Exact file length, digest and the retained read-only handle remain mandatory; existing payload-loading progress/cancellation behavior is unchanged. |
| State reuse | Library callers can explicitly select PromptReuse::Latest; ownership, exact-prefix and adapter-scale checks prevent incompatible reuse. The private service resets for every request by default; latest-prefix reuse requires explicit selection. |
| Correctness and delivery | Added regressions cover row-worker parity and failure recovery, sampled and hybrid-state reuse, adapter profile boundaries and hash streams. Source-only rollback and private-service capture tools support further bounded qualification. |
| Rust consumers | GGUF now has its own 0.1.1 source identity and the runtime pins it exactly. Hosted source-package checks exercise packaged dependencies and an external library consumer. Installation and library examples distinguish GitHub source/binaries from registry publication. |
The production recurrent update retains the prior copy-then-scale arithmetic.
The proposed fused copy/decay implementation is test-only pending its own
qualification. This release claims no measured startup, throughput or memory gain.
Supported release scope
The first generation path remains the pinned Swift/Bonsai dense Qwen35 CPU
architecture. Original PQ2, one row worker, no adapter and fresh private-service
request state remain the defaults. The runtime has no C/C++ inference dependency;
the existing Rust import path remains bonsai_prism_runtime.
PTQ1, approved-adapter execution, extra row workers and exact-prefix reuse require
explicit selection. Their device, numerical, quality and resource qualification
remains scoped to the exact evidence already documented in the
roadmap, worker contract,
adapter contract and
profile contract. Full-context trajectories, adapter
authoring provenance, sustained comfort, application isolation and installation
acceptance remain open. These limits also prevent automatic promotion in Runner.
The isolated, unpublished native Hexagon diagnostic crate is outside the five
bundled CPU executables. It enables no model offload or accelerator profile.
Broader model architectures and general model selection remain future work.
Installation and provenance
Use the versioned Windows archives and matching integrity receipts attached to
the release, or the pinned Git source instructions in the
build guide. Models and adapters are separate artifacts.
Each binary receipt records its source revision, compiler, target, complete file
inventory and archive/file SHA-256 digests. Older releases remain available for
rollback; this release does not repin an application's existing runtime lock.
The release requires successful Foundation, runtime conformance and source-package
checks on its exact source revision. Binary attachments must be the unchanged
Foundation artifacts for that revision. The published release notes retain links
to those successful owning runs.
Registry publication is a separate delivery channel: publish bonsai-native-hash
0.1.0, then bonsai-gguf 0.1.1, then prism-inference 0.1.12. Source-package
checks exercise these packages before publication; they do not publish them.
The private server retains its exact Git credential-helper dependency and is not
registry-published. Use the documented Git source dependency until the matching
registry versions are available.
Validation for this release
Source: c48687616f55eec9aff0de04860917de3e1047f2 on main.
- Foundation: native Windows ARM64/x64 correctness, analyzers, feature variants and binary bundles, plus Linux native storage overlay.
- Runtime conformance: owning upstream source/backend and native codec gates.
- Source package: verified normalized Cargo archives and independent scalar/default/parallel Rust consumers on Windows ARM64, Windows x64 and Linux ARM64.
The attached archives and receipts are the unchanged Foundation artifacts for
this source. Their archive digests, complete nine-file inventories, individual
file digests and executable targets were verified before publication. All three
owning workflows passed at the source above. No new owner-device model trial,
benchmark result or automatic application/default promotion is claimed.
Rust API migration from 0.1.6
MatvecStatistics adds the public parallel_row_operations counter. Callers
using Default or reading existing fields need no change. Update exhaustive
struct literals to supply the new field (or use ..Default::default()), and
use .. when destructuring only selected fields. Existing public methods and
exports remain available; this release retains the established prerelease API.
Prism Inference 0.1.11
Prism Inference 0.1.11 candidate
This GitHub binary candidate adds the effective row-worker budget to private
native HTTP readiness. The release review compares the complete source delta
against 749a5f5b43bf77764a29682557ec91a8de8dcdef, the pinned 0.1.10 source.
It preserves explicit original-PQ2/current-weight-PTQ1 selection, approved-adapter
identity and scale bounds, fresh state for every request, the one-worker default
and existing process ownership and admission requirements.
Production execution and protocol changes
The row_workers readiness field comes from the instantiated session rather
than echoing command-line arguments. Owners must compare it with their selected
worker policy. The private worker rejects zero or over-limit profile identities
before spawning its model-owner thread. The field reports the effective budget;
it does not count completed parallel operations, select an ISA or establish a
performance gain. Foundation's feature-variant controls cover matching serial
and four-worker budgets plus rejection of invalid identities.
The unreleased fused recurrent copy/decay change from 244df2e was discovered in
the full release delta and removed from production before this release candidate.
The production recurrent path again uses the exact 749a5f5 copy-then-scale
sequence. The independent native GDN fixture keeps its raw captured state inputs;
the fusion candidate's model-free bitwise/scratch controls remain test-only.
Git history and the runtime contract retain the candidate, pending actual-model
logits/live-state, throughput and resource qualification before release/repinning.
The runtime/server package versions, exact server runtime dependency, two lockfile
package versions and Foundation archive version become 0.1.11. No third-party
dependency version or checksum changes. Kernels, sampler, tokenizer, adapter
arithmetic and authoring basis are unchanged from the selected 0.1.10 source.
This source review does not establish equality of newly compiled binary bytes or
replace exact artifact verification and target-device qualification.
Windows CNG GGUF verification and opt-in PromptReuse::Latest were already
present in 749a5f5; they are retained capabilities, not new 0.1.11 changes.
Serving retains reset-every-request. Exact length/SHA-256 checks and read-only
file guards remain mandatory. No startup, throughput or resource improvement is
claimed by this candidate.
Source tools outside the five-executable bundle
The source-only swift_model_check rollback mode adds bounded complete-state
cancellation/continuation captures. It retains exact tokens, logits, logical
KV/convolution/recurrent state and create-new evidence under its 2 GiB payload
cap. Its capture-stop token and cooperative cancellation do not qualify production
sampling, native HTTP cancellation or a full Swift workload.
The new private-service checker, source-bound owner workflow and intake/archive
helpers prepare one externally supervised released-service capture. Setup,
loading, work, comparison, retention and cleanup share the existing five-minute
owner budget; partial captures cannot continue through later owner jobs. Their
pinned Runner owner and its selected release lock are independent identities.
Preparing these tools does not automatically retarget an existing pinned owner
to 0.1.11 or qualify the new binary release.
The optional unpublished Hexagon source crate adds fixed signed X2 F32/guarded-Q2
queue controls, explicit nonblocking submission, mapping-before-module teardown
and conservative retention after uncertain native results. It remains a separate
terminal diagnostic surface requiring reviewed package/driver/ABI inputs and an
external process owner. None of the five bundled CPU executables depends on this
crate; these additions do not enable model offload or a Hexagon runtime profile.
Its owning source/ABI/device gates remain separate from CPU release qualification.
Delivery and remaining qualification
Release only the exact Foundation-produced ARM64 and x64 archives and receipts
after the owning main Foundation and runtime conformance gates pass. Preserve
all older release bytes. Runner integration needs a reviewed exact source/archive
lock, every bundled file digest and explicit version admission for four workers.
The changed readiness must pass actual released-service lifecycle acceptance;
previous worker-policy device results do not transfer to this candidate.
New-source full-context logits/live-state, loaded-zero/nonzero application,
adapter authoring provenance, generated-history lifecycle, paired benefit,
independent human quality/comfort, resource, isolation, installation and rollback
gates remain required for default promotion. A source-equivalent CPU path and
hosted controls do not close these device/application gates.
Registry publication is separate: the new native hash crate, a new GGUF registry
version and an updated exact runtime dependency remain prerequisites. The server
still has an exact Git credential-helper dependency and is not registry-published.
This binary release changes no PTQ1 default, adapter qualification, model quality
or device-resource claim.
Prism Inference 0.1.10 — native verification and explicit prefix reuse
Prepare native Windows model verification and explicit same-session prefix reuse for Swift.
- GGUF artifact verification uses audited Windows CNG SHA-256 through bounded streams while preserving complete byte verification, retained handles, cursor restoration and failure behavior. Other targets retain portable Rust hashing.
- The private native service accepts explicit
--prompt-reuse latestselection. Exact latest committed prefixes require bit-identical adapter scale; edited histories recompute and transactional failures preserve state and sampling. The default remains fresh state for each request, with one row worker. - Focused controls compare full sampled logits and logical attention/recurrent/convolution state with fresh computation, including scale changes, rollback, sealed KV pages and stop draws.
Exact source: 749a5f5b43bf77764a29682557ec91a8de8dcdef.
Owning validation: Foundation and runtime conformance. Both completed successfully on this exact main source. Bundle receipts bind the exact source, pinned native compiler, target and every archive entry.
This is an experimental GitHub binary prerelease. It does not qualify actual-model reuse, startup performance, full-context/long-output behavior, adapter authoring basis, quality or default-runtime promotion. Registry delivery remains separate: native-hash publication, a new GGUF registry version and an updated exact runtime dependency are required. Server registry publication remains disabled pending its reviewed Windows credential helper release. Previous releases remain available for rollback.
Prism Inference 0.1.9 — explicit row-worker availability
Prism Inference 0.1.9 makes the optional persistent row-worker implementation available in the checked Windows ARM64 and x64 binaries. Every entry point still defaults to one row worker. Multiple workers require explicit selection; this prerelease does not enable four workers in Runner or change any app default.
The source also extends the qualification example's explicit worker/backend controls. Full-model correctness, performance, creative quality and daily-service qualification remain unqualified. Synthetic and conformance checks do not establish a speed gain.
Both complete source gates passed at 7aca2251f09acb345d4afb53d5ce411f81017745:
These are the exact source-bound bundles from the full Foundation run, with unchanged receipt format and a flat nine-file inventory. The release build explicitly enables prism-inference/parallel-rows,prism-inference-server/parallel-rows; no later stage recompiles the packaged binaries. Runtime source and all asset/payload digests were verified locally without executing the staged binaries or loading a model.
The Windows server retains the exact external credential helper at SwiftBonsai2Runner revision abe04b4a8eff18388af96f0e75c8a8d180b81989. That Git dependency is not included in the Prism source ZIP and its source must be retained separately. Server crates.io publication remains disabled.
This prerelease provides reviewed artifact provenance and explicit qualification controls. No production Runner lock, installed runtime, model, worker count, sandbox policy or resource limit was changed by staging it. Do not treat it as default-runtime promotion or PTQ1 release-quality approval.
Prism Inference 0.1.8 — contained read-path startup fix
Prism Inference 0.1.8 corrects private-child read-path validation for the existing exact-file AppContainer grants. It uses the reviewed read-only consumer helper, pinned to SwiftBonsai2Runner abe04b4, instead of requiring volume-root metadata that PSEC can deny. Private-key reading retains its read handle and exact 64-byte lowercase-hex validation.
Startup failures now report only a fixed stage, I/O error kind and optional numeric OS code. Arbitrary error messages, paths, arguments and credential values are never formatted into these diagnostics. Filesystem grants, resource limits, inference arithmetic and model defaults are unchanged.
Source: fce300b.
Foundation: https://github.com/symbiogenesis/prism-inference/actions/runs/41017335054 (passed).
Runtime conformance: https://github.com/symbiogenesis/prism-inference/actions/runs/41017335160 (passed).
Both bundles and receipts are from that exact Foundation run. Runner's offline archive/source/file/PE review passed for the ARM64 candidate; both platforms' complete ZIP contents and file digests were independently verified. The helper source is pinned exactly and retained separately for review.
The server is delivered through GitHub binaries and exact source checkouts. Its crates.io publication is disabled until the reviewed helper API has a registry release; the independent core runtime crate's publication policy is unchanged.
The previous 0.1.7 required-container application attempt failed before verified readiness. Application acceptance of this replacement still requires the separate five-minute owner-device smoke after Runner pin integration. PTQ1 remains diagnostic only; this prerelease does not qualify complete answers, adapter authoring basis, full context/output, sustained resources, performance, quality or daily isolation. The 0.1.7 files remain available and unchanged.
Bounded application evidence: Runner main 1540297 passed required-sandbox PTQ1/approved-adapter readiness, one scale-1.3 streamed response and owned Job/listener cleanup: https://github.com/symbiogenesis/SwiftBonsai2Runner/actions/runs/41020705958 . The owner job took 180 seconds. The four-token cap yielded one reported output token, three visible UTF-8 bytes and natural stop; optional zero-scale and invalid-scale/reuse checks were skipped. This does not qualify broader lifecycle, quality or default promotion.
Prism Inference 0.1.7 — explicit PTQ1 and approved adapter service
Prism Inference 0.1.7 adds explicit current-weight Swift PTQ1 and approved rank-one adapter service profiles. It also binds prepared chat requests to session and adapter settings, and resets graph state safely when adapter scale changes.
The original PQ2/no-adapter interfaces remain available. The private server accepts the canonical single-slot LoRA request emitted by Runner and validates exact readiness identities and scale values. Model defaults and admission limits are unchanged.
Source: 27e4860.
Foundation: https://github.com/symbiogenesis/prism-inference/actions/runs/41009394716 (passed).
Runtime conformance: https://github.com/symbiogenesis/prism-inference/actions/runs/41009394718 (passed).
The ARM64 and x64 bundles and receipts are from that exact Foundation run. The ARM64 bundle has also passed Runner's offline source/archive/file-identity review.
This prerelease does not establish PTQ1 release quality, full-context resources, sustained performance, creative quality or daily-service isolation. A five-minute owner-device numerical smoke and application smoke remain separate qualification evidence. PTQ1 is not promoted to the default model or runtime.
The source-pinned one-token native/Rust Auto numerical smoke at adapter scale 1.3 passed all 129 files / 39,504,384 logits and live-state values byte for byte in owner CI: https://github.com/symbiogenesis/SwiftBonsai2Runner/actions/runs/41012887460 . The earliest recorded setup through job completion took 196 seconds. This is correctness evidence for the fixed supplied token, not throughput or service qualification.
Application boundary: required-container owner CI https://github.com/symbiogenesis/SwiftBonsai2Runner/actions/runs/41014805669 failed during backend startup before verified readiness or streaming. The captured result does not establish whether weights loaded. This prerelease remains unqualified for that application profile; the numerical smoke above is separate.
Prism Inference 0.1.6
Corrects pinned PTQ1 Q8_0 activation scale and inverse-scale arithmetic to match
the native ARM64 implementation. Generic and PQ2 activation preparation retain
their prior behavior. The release includes independent native-backed controls
and serializes the contended native conformance partition.
Source: 02c8a24b31ac1c35f2ffd3530f6a753ce440bdc0 on main.
Both Windows targets, the native storage overlay and the aggregate
Foundation gate
passed. The complete
runtime conformance gate
also passed at that exact revision.
The attached Windows ARM64/x64 bundles are the unchanged Foundation artifacts;
their receipts bind the source, pinned compiler, complete file inventory and
archive/file hashes. The native Rust Swift service continues to select PQ2.
PTQ1 remains an explicit offline correctness lane, with adapter execution,
full-context service behavior, creative quality and device performance requiring
separate qualification.