atmem/atmem-laya-formation-v1

Superseded research artifact. This V1 artifact was trained/evaluated on fictional synthetic scenarios and is retained only for reproducibility. It is not evidence of authentic-conversation memory quality and is not the data or model planned for AtMem 2.3.9b1. The replacement program uses authentic, original-language conversations, frozen stateful SQLite/vector pre-states and strict V2 Choice/Noul/Score transition contracts. Its database, transition dataset and independently usable model will be linked here only after their publication gates and anonymous-redownload checks pass.

AtMem Laya Formation Dataset v1

This dataset contains synthetic-only, fictional memory-formation scenarios and finite typed decisions for the optional AtMem Laya formation profile. It contains no real user history, production memory, credentials, private conversations, or public benchmark answers. It is not production-quality evidence and cannot by itself support claims about real-world retrieval or memory performance.

Intended use

The dataset supports training and calibration experiments for bounded decisions:

• operation: ADD, UPDATE, SUPERSEDE, NOOP, or REJECT;
• memory class: durable fact, temporary state, episode, procedure, or non-memory;
• evidence support: supported, ambiguous, or unsupported;
• selection of an already-active canonical target, no target, or review;
• retrieval usefulness, non-usefulness, or insufficient evidence.

AtMem remains the sole authority for admission, mutation, policy, provenance, scope, deletion, and delivery. A model trained on this dataset may only nominate finite choices; it must never mutate canonical memory or invent source evidence.

Do not use this dataset as personal data, conversational training data, a general language benchmark, proof of production quality, or authority to store content.

Data statement and provenance

Every identity, value, event, scope, timestamp, and credential-shaped reference is deterministic fictional content generated by AtMem generator 1.3.0 with seed 410239. Credential-shaped fields are described as withheld and no secret value is generated. The source generator and artifact are Apache-2.0.

The generator constructs explicit pre-state, authorized evidence, expected post-state, safety tags, group identities, and five typed questions per scenario. An independent state oracle emits finite-choice labels and receipts. Validators recompile every decision, enforce choice membership, reject secret-like content, and check exact and normalized near-duplicate leakage.

Size and splits

The immutable artifact contains 12,000 scenarios and 60,000 decisions.

Split	Scenarios	Decisions	Permitted use
train	7,920	39,600	objective fitting
validation	1,300	6,500	model selection
calibration	1,340	6,700	threshold fitting only
sealed_test	1,440	7,200	one-time evaluation after freeze

Training code is restricted to train, validation, and calibration paths. Sealed test inputs and labels must remain unopened until the model, calibration bundle, and analysis code are frozen.

Splits are group-disjoint across fictional identity, template family, semantic chain, and paraphrase cluster. Exact and normalized near-duplicate content is also checked across split boundaries.

Coverage

Each decision question has 12,000 examples. Class counts are:

• operations: ADD 2,228; UPDATE 2,228; SUPERSEDE 2,229; NOOP 2,229; REJECT 3,086;
• memory classes: durable fact 2,880; temporary state 2,400; episode 2,400; procedure 2,400; non-memory 1,920;
• safety scenarios: contradiction 721; credential-withheld 1,091; cross-scope 706; private-scope 924; sensitive 1,715; duplicate 2,400; temporal change 4,800.

REJECT is intentionally more frequent because every unresolved contradiction fails closed and preserves the canonical pre-state.

Audit and quality gates

A deterministic stratified packet sampled 400 non-sealed rows, with at least 60 examples for each required high-risk tag. A firewalled independent gpt-4.1 review received the blinded packet and rubric but never the answer key. The Ed25519-bound result reviewed 400 rows, matched 399 exactly (99.75%), and reported zero critical privacy, credential, licensing, scope, or correctness findings. The signed review and scoring result are included under audits/.

This audit checks synthetic label consistency; it is not human-subject review, production evaluation, or evidence that a trained model will achieve the same accuracy.

Reproduction

From the source revision that contains generator 1.3.0, with Python 3.12 and the locked project environment:

python -m pytest tests/research/test_laya_dataset.py tests/research/test_laya_openai_review.py tests/research/test_laya_jev_protocol.py -q
python -m research.laya_formation.dataset.export
python -m research.laya_formation.publication.lint_dataset_card research/laya_formation/publication/DATASET_CARD.md

The publishable exporter requires /Volumes/MEM/AtMem-Laya-Formation, at least 12,000 scenarios, Parquet support, and seed 410239. It writes canonical JSONL, Parquet, split/audit material, file sizes, and SHA-256 digests. The artifact root manifest is manifests/dataset-manifest.json; its v1.3.0 SHA-256 is a8193d66a0184f1dbbb94ec73d2f85b6362ddfb8457a5a18dc829082a8851d60 before the card, license, schemas, and audit receipts are added to the final publication inventory. The staging receipt records the final manifest digest.

Limitations

• English-only deterministic templates do not represent natural conversation diversity, multilingual use, long histories, noisy extraction, or social context.
• Labels encode the published AtMem policy and may not transfer to another memory system or policy.
• Synthetic balance and explicit wording can make examples easier than real use.
• The independent audit used a model reviewer; it does not replace later sealed review, public-corpus evaluation, or end-user safety testing.
• No result from this dataset alone may be described as production-level or as beating Jev, Qwen, Laya, or another model.

License

Apache License 2.0. See LICENSE. Downstream users remain responsible for model, software, and benchmark licenses used with this dataset.