qwenmem-spine

A text-only Qwen3.5-4B derivative with five compiled triangle-navigation layers, built for an experimental Rust Spine memory integration. It was compiled, not fine-tuned on the evaluation questions. It requires the matching custom runtime; ordinary hosted inference widgets and generic generation servers are not validated entry points.

Source and instructions: Jaredcastorena/qwenmem-spine, source revision ac7e0e55c5d67ef99f61aa2d50d12b1d996f9b42, including the opt-in internal-context integration and its November 3 evidence. Use the runtime from the pinned source revision above; historical source baseline pins do not include all integration changes. Model distribution destination: Jared1728/qwenmem-spine. Access to either repository depends on its visibility and account permissions. This card describes the measured artifact; it does not assert that a download or fresh-machine installation has been tested.

Maintenance and licensing

This is a fixed, unmaintained research snapshot. No updates, bug fixes, new features, compatibility updates or ongoing support are planned. It is intended for experimental reproduction; this is not a field-of-use restriction. Commercial reuse is permitted under the applicable license terms.

The base and derivative model weights retain Apache-2.0. Newly authored source code is MIT. Rust Spine retains MIT OR Apache-2.0, and Qwenputer plus adapted compiler/runtime code retain MIT. Preserve the corresponding notices. The source repository's THIRD_PARTY.md describes component scope. This maintenance policy does not describe upstream Qwen, Qwenputer or Rust Spine.

Artifact identity

The measured checkpoint has ten weight shards containing 496 tensors. Its composite weight identity is:

ebbcad96c6073c9a2cab13cba4d7414f3ed45117626ae8320fd6092a3210ac35

This is SHA-256 of the JSON-serialized sorted shard-digest map in triangle.json, not a hash of one concatenated model file. The runtime checks shard identity before inference. The source repository's checkpoint-manifest.json records the measured 17-file checkpoint set, including configuration, tokenizer, chat template, index and notices. Checksums detect accidental changes; a replaceable manifest is not a cryptographic authenticity guarantee.

The merge prepends five BF16 compiled layers and widens the language residual stream from 2,560 to 2,816. Base-weight padding and normalization-gain adjustment can introduce BF16 rounding differences, so bitwise equivalence to base Qwen is not claimed. The model includes no vision component. The separate Qwenputer Filesystem checkpoint was a comparison input, not the source of these triangle weights and not a required runtime artifact.

Memory and runtime behavior

The controller supports 64 combined resident node/triangle records and eight descents. Its program is in the weights; installed topology and traversal state live in recurrent tensors. Original source text, provenance, semantic vectors, stable IDs and the complete native hierarchy remain in Spine's encrypted heart. Rust chooses semantic routes and evidence. In-model triangle creation, folding, ranking, paging and autonomous routing are not implemented.

Default chat uses native harness messages and tools and installs an empty internal navigation forest. Its provider performs no extra heart scan. Native compaction, ranking, automatic hybrid recall and explicit memory tools remain active. Optional triangle-only automatic memory withholds hybrid recall text while preserving the task, history, retained state and explicit tools.

The opt-in --triangle-internal-context mode installs complete topology for each native-selected root, bounded to 64 records and depth eight per bank, with no source bodies in that topology. The actual checkpoint's five-layer prefix executes navigation commands. Rust independently checks exact traces, installed addresses and execution evidence before returned apex IDs authorize the original selected excerpts for the language prompt. An oversized closure or failed check stops the request. Every completion replays a durable system-message sidecar; there is no provider-owned source scan or cross-request context store. Ranking, compaction, source storage and source budgets remain native host operations. See the source repository's docs/internal-context-validation.md for the measured scope; standalone navigation is a separate execution path.

Use this repository revision's Python backend, included Rust source, pinned compiler and matching checkpoint. Follow the reproduction guide and direct runtime instructions. Runtime loading uses local files; dependencies must be provisioned first. A generic Transformers loader is not an established substitute for the command/text dispatch and state-validation contract.

Normal chat caps prompts at 32,768 exact tokens and generated output at 2,048 tokens per round. Each completion has a fresh fixed-capacity language cache, 256-token prefill chunks and finite-value checks. Navigation tensors must remain unchanged during language generation. Process restart recovers continuity from the heart and journal, not an automatically restored full-history KV cache. Weights do not contain a user's heart, conversation or installed forest.

Recorded evaluations

The first historical chat battery ran with --triangle-internal-context off and an empty model navigation forest. The second battery, measured November 3, 2026, enabled that flag and used the same unchanged model weights. It is included in this revision; the flag remains opt-in.

Evaluation	Result and scope
Internal native mixed-memory application	14/14 operational assertions; all 31 completed rounds used verified compiled navigation, with native hybrid evidence also available.
Internal triangle-only automatic memory	12/14 operational assertions; all 27 completed model rounds used verified compiled navigation. Literal/citation and file-write failures remain.
Internal triangle-only boundaries	1/2; older event recovered, long-source-tail value missed despite verified navigation.
Internal explicitly instructed recall	3/3: both boundaries and the literal, with original citations. The literal probe used leaf-only banks.
Historical native mixed-memory application	14/14 operational assertions with an empty navigator forest, including real native/file tools, restart and interruption.
Historical triangle-only automatic memory	11/14 with an empty navigator forest; literal fabrication, wrong-question repair and tool-path failure retained.
Historical unassisted source boundaries	Native mixed-memory 0/2 because required original citations were omitted; triangle-only 1/2.
Historical explicitly instructed native recall	3/3 selected probes: both boundaries and exact Unicode literal, with original citations. Assisted results do not replace failures.
Earlier compiled-navigation retention	8/8 at each of 4,274, 32,812 and 131,215 stored content tokens, with 572–901-token evidence prompts. Two extra renderer boundaries failed.
Valid raw-context control	8/8 at 167,780 input tokens; a fresh process restored the full cache and reproduced answer bytes and token IDs.

These are small fixed synthetic tests and operational assertions, not general accuracy estimates. The internal triangle-only run passed the navigation gate while failing two behavioral assertions; navigation is not answer verification. The internal mixed mode also supplies native hybrid recall; its answer score does not isolate the gated triangle excerpts as the source of success. These are single runs with the host's minimum temperature of 0.1. All 70 completed internal-context rounds passed the source gate across 210 bank executions and preserved navigation state during language generation. A separate full-agent resident-bank fault produced an invalid trace and was rejected before language completion, with zero accepted completions. The aggregate evidence records these runs and the independent standalone zero/restore probe.

The historical application controls above do not exercise active compiled navigation. Earlier provider trials used an extra source scan that has since been removed; their results do not describe current retrieval behavior. Complete tables, failure categories and sanitized aggregate evidence are in the source repository's docs/results.md and evidence/.

Known failures and limitations

The Rust host's claim extractor can miss factual string identifiers, producing zero claims and bypassing NLI and repair. This is tracked in Rust-spine issue #40; no fix is included. Forced extraction triggers repair for the wrong Unicode literal but also rejects the correct literal under the measured NLI gate. Pair-order reversal is not a reliable exact-string fix. A real citation does not by itself establish source support. The defect does not invalidate the working provider/tool connection and does not explain every automatic-answer failure.

Native source rendering can omit older events or truncate required tails. Explicit recall recovers the measured boundary fixtures but has no arbitrary event-pagination guarantee. Repair can answer an older question. These limits preclude reliable autonomous recovery or lossless/unlimited-memory claims.

Two early raw-context attempts were interrupted without complete answers. A separate bounded run developed NaN recurrent state and logits and emitted punctuation. That invalid output is not scored as forgetting. A later checked replay and restart passed, but the original numerical trigger remains unresolved. The runtime's finite-value checks do not establish a permanent backend fix.

Full-model tests used Linux, Python 3.12.3, PyTorch 2.14.1+xpu, Transformers 5.19.0, safetensors 0.8.0, NumPy 2.5.4 and one Intel Arc Pro B70 with 32 GB device memory. Controller checks also run on CPU. Full-model CUDA, macOS and Windows execution and clean-machine installation are unverified. Guarded runs used a 24 GiB host cgroup limit, 20 GiB throttle, disabled swap, 16 GiB host reserve and 70% XPU allocator cap. These are measured controls, not portable memory-fit guarantees. Source text and dense runtime state still consume storage.

Pinned provenance

Component	Revision
Qwen/Qwen3.5-4B	851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
divagr18/qwenputer	01cf657d353bdc0c795c2d57a28a3e3f483f20a1
Integration source and runtime	This repository revision, including the November 3 internal-context implementation and evidence.
Historical Rust Spine baseline	07857378505d2573b623da4f81a892bfb9b5f2cd; the included source adds integration changes.
Published merged model weights	f69dbcfe5a7cebea1c287f6f1d76c8e90b2c9a98 at Jared1728/qwenmem-spine (unchanged).
sentence-transformers/all-MiniLM-L6-v2	1110a243fdf4706b3f48f1d95db1a4f5529b4d41
cross-encoder/nli-MiniLM2-L6-H768	b95119ce93d3e065de6214e38cd4a97b0f2f2c6d

The source repository's models.lock.json and requirements.lock.txt identify download/build inputs. Source, weights, navigation state, encrypted heart and raw-context cache checkpoints are distinct artifacts with different roles.