README.md

---
license: cc-by-nc-4.0
library_name: transformers
base_model: Qwen/Qwen3-4B-Instruct-2507
pipeline_tag: text-generation
tags:
  - strata
  - persistent-memory
  - structured-memory
  - neuro-symbolic
  - exact-value-copying
---

# STRATA Native LM v1

Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) for non-commercial use. Commercial use requires separate permission from the rights holders. The upstream base model and third-party dependencies retain their respective licenses.

STRATA Native LM couples a frozen Qwen3-4B-Instruct-2507 backbone to a trained 20.44M-parameter memory interface. It generates short factual responses from trusted structured observations supplied by tools, APIs, or applications. Structured writes and read addresses are supplied explicitly. The runtime executes joins; the language model verbalizes their selected results. Open-ended response generation was not evaluated.

```text
structured observation
  -> WRITE(instance, predicate, role, value)
  -> persistent structured memory
  -> memory-conditioned language model
  -> generated frame + exact value copying
```

The reader decodes and latches an address-bound value code. During response-frame generation the actual value state is replaced by a fixed neutral state, and COPY validates the selected record. Its bytes are inserted after frame completion. No previous source tokens, source-history KV-cache entries, persisted source residuals, or actual value tokens enter frame generation. Integrity checks reject invalid, stale, deleted, or corrupted references.

The 255 nonzero value codes are reused across records, not global value identifiers. Structured-service adapters assign a field index; persistent identity comes from the full instance/predicate/role address and version. The evaluated copy bank contains one runtime-selected record per request. Arbitrary application bytes remain in that record and are not encoded by the 8-bit code.

## Results

| Metric | Result |
| --- | ---: |
| Atomic writes in state history | 2,004,344 |
| Transactions | 564,312 |
| Unique base records | 4,096 |
| Evaluations across ages and packings | 61,440 |
| Persistent-memory horizon | 128 independent 8,192-token windows |
| Factual accuracy | 1.000 |
| Copied-value byte correctness | 1.000 |
| Three schemas excluded from component training | 768 unique records; 11,520 evaluations; 1.000 |
| Runtime-executed 2/3/4-hop joins | 1.000 |
| Field-specific intervention following | 1.000 |
| Integrity verification and execution replay | 1.000 |
| Structured-service retained-state reduction, including values and integrity metadata | At least 86.72% |
| Neural value-state reduction | 95.83% (128 vs. 3,072 bytes/value) |

The neural-state figure excludes application value bytes and shared address/integrity metadata. The structured-service accounting includes those bytes. The horizon is persistence across independent windows, not a single Transformer context.

STRATA and the realization-matched token-value arm both achieved 1.000 factual accuracy. The original uncached source-history timing remains in `evaluation.json` as an implementation diagnostic, not a speed claim against optimized cached serving. The three training-excluded schemas share the trained memory interface, operation inventory, and templated query format.

## Selected-Result Prompt Comparison

The same 4,096 unique selected current records were passed to the frozen base model as JSON in the current prompt, with within-request KV caching enabled. Both interfaces were evaluated twice on eight L40 GPUs without cross-request prefix caching.

| Metric | STRATA | Prompt serialization |
| --- | ---: | ---: |
| Stored UTF-8 value reproduced exactly once | 4,096/4,096 | 4,051/4,096 |
| Ordinary bindings and resolved joins | 3,328/3,328 | 3,328/3,328 |
| Mean input tokens/request | 64.52 | 136.29 |
| Batch-amortized ms/record | 37.81 | 117.51 |
| Generation-cap cases | 0 | 8 |

All 45 prompt failures were in the hard-value holdout: 36 explicit refusals, eight outputs capped at 256 generated tokens, and one other value mismatch. Prompt generation freely spells values; STRATA copies their bytes. These are measured interface tradeoffs, not an optimized-serving or retrieval-quality ranking. Both passes produced identical responses.

The `comparison/` directory contains the selected records, sealed responses, protocol, results, and executable comparison. Recompute published scores without a GPU:

```bash
python3 comparison/benchmark.py --score-published
```

After installation and base-model download, reproduce both interfaces on eight GPUs:

```bash
python3 comparison/benchmark.py --base-model "$STRATA_BASE_MODEL" --output comparison-run
```

## Install

Requires Python 3.11 or newer and a CUDA-capable PyTorch installation. A single NVIDIA L40 with 46 GB memory was used for the release inference check.

```bash
python3 -m pip install huggingface_hub
hf download nur-dev/strata-native-lm --local-dir strata-native-lm-v1
cd strata-native-lm-v1
python3 verify.py
python3 -m pip install -r requirements.txt
python3 -m pip install --no-deps wheels/strata-0.1.0.post2-py3-none-any.whl
hf download Qwen/Qwen3-4B-Instruct-2507 \
  --revision cdbee75f17c01a7cc42f958dc650907174af0554 \
  --local-dir base-model
export STRATA_BASE_MODEL="$PWD/base-model"
```

## Run

This example constructs one structured observation and reads it without putting its value in the query.

```bash
python3 load_and_answer.py \
  --event calendar:meeting_283 \
  --predicate meeting \
  --role location \
  --payload-handle 17 \
  --payload "North Campus Room 17" \
  --query "Where is the meeting?" \
  --device cuda:0
```

The JSON response contains `answer`, `frame`, `status`, `payload_handle`, `receipt`, and timing fields. Successful output contains the stored value `North Campus Room 17` verbatim; surrounding wording is generated by the model.

## Files

| Path | Contents |
| --- | --- |
| `checkpoints/` | Compact-memory reader, memory interface, and copy-action head |
| `configs/` | Inference settings |
| `evaluation.json` | Evaluation metrics, criteria, and source-artifact hashes |
| `comparison/` | Selected-result baseline data, outputs, protocol, results, and runner |
| `wheels/` | Runtime package |
| `BASE_MODEL.json` | Required upstream model and immutable revision |
| `load_and_answer.py` | Executable structured-read example |
| `verify.py`, `MANIFEST.sha256` | Local file-integrity verification |

The base model is downloaded from its upstream Hugging Face repository; its weights are not duplicated here.