atmem/atmem-laya-formation-model-v1

AtMem Laya Formation Model v1

This is an optional, finite-choice System One model for AtMem memory-formation workflows. It proposes one answer for each of five bounded questions: operation, memory_class, evidence_support, target_selection, and retrieval_usefulness.

It is not an authority system. AtMem must still authenticate the actor, enforce scope and policy, load canonical state, validate evidence and targets, handle deletion/revocation races, and authorize any mutation. A model answer is a proposal, never permission to store, update, supersede, disclose, or retrieve memory.

Intended use

Use this artifact only through the version-bound AtMem Laya formation profile. That profile supplies the exact question and choice definitions, tokenizer- exact input packing, calibration bundle, abstention threshold, score validation, deterministic fallback, and optional bounded AtBot escalation.

Expected end-user benefit after the full profile and compatibility gates are complete is lower-cost, lower-latency handling of routine memory judgments on the user's own CPU/GPU, with deterministic AtMem fallback for invalid, unsupported, unavailable, or abstained decisions and governed escalation for the small set of ambiguous cases. The model does not replace retrieval, the memory store, AtBot, or AtMem policy enforcement.

Do not use it as a general chat model, an open-ended extractor, a security boundary, an identity/authorization oracle, or an autonomous writer. Do not send extra provider context merely because a model asks for it.

Exact lineage

Component	Immutable identity
Base checkpoint	convaiinnovations/laya@7b928d828b7b0e022f929d9bd2e44165aa270148
Base model SHA-256	891102d372688fc2a094dac56a384bc537b87c63f21f9f3dac0be2b7cbc8d86c
Laya source	NandhaKishorM/laya@68804629e8ccd9d616d48a40e87de9fabbeae069
Laya package	laya==0.4.2
Synthetic dataset	atmem/atmem-laya-formation-v1@4436089b38cbe3a5374aaaa92c4c77703cdbf668
Dataset manifest	d67d371c8c1d6de85a0293cded5aa5b3f7d078e92197130ce1f9edef68e863d4
Selected run	rlcd-410239-20261010T152813Z
Model SHA-256	f2ba3dd7679da45aabab4cac30e9a079baeb2dc98229ff3b1618e6a06fc6d683
Question digest	7a3cd0e13d99db02a1e53e51a63fad921e754ac5bbc2376b8a888c1033900f0e
Calibration digest	b165d4b7d554cdbda5a889baf7059c3f43ba2774d68754e4fa3aeb63f5f0e067
Portable export digest	6f7d53c18431c27baa617ab95ebcc4cad669917a6508145fa001eb7c8e799900

The dataset contains 12,000 fictional scenarios and 60,000 decisions. Its group-disjoint decision splits are 39,600 train, 6,500 validation, 6,700 calibration, and 7,200 sealed test. No production or user memory was used.

Training

Training used full-encoder upstream RLCD/proper scoring with seed 410239, four epochs, micro-batch 32, gradient accumulation 2 (effective batch up to 64), encoder learning rate 2.5e-5, head learning rate 1e-4, minimum learning rate 1e-6, weight decay 0.01, gradient clip 1.0, four RL samples, sigma 0.4 → 0.1, w_sph=0.75, and w_rps=1.0. RLCD ran in FP32 because the upstream FP16 path skipped an initial optimizer update after overflow. Training used PyTorch 2.11.0+cu128 on one RTX 5090 and took 4,390.91 seconds.

A soft cross-entropy control and RLCD produced an exact tie under the predeclared validation/calibration ranking. The frozen stable objective-name tie-break selected RLCD; sealed data was unavailable to selection.

Calibration and abstention

Calibration was fit only on the isolated 6,700-decision calibration split, with 1,340 examples for each question. The external calibration.json is mandatory: it binds the selected model and questions, installs type temperatures [1.0, 1.0, 1.0], a choice:3-5 temperature of 1.0, and a conservative choice:3-5 minimum-confidence threshold of 1.0 for target error 0.10. The original temperatures in rl_agent_config.json are base-checkpoint values; AtMem must not silently serve this derivative without applying the bound external calibration bundle.

Synthetic evaluation

All figures in this table are synthetic. They do not establish production or public-corpus performance.

Split	Items	Accuracy	Macro-F1	NLL	Brier	ECE
Validation	6,500	1.0	1.0	0.0	0.0	0.0
Calibration	6,700	1.0	1.0	0.0	0.0	0.0
Sealed test	7,200	1.0	1.0	2.4447e-13	3.5791e-25	2.4447e-13

The sealed test had 1,440 examples for each of the five questions, zero failed decisions, zero truncation, no missing expected class, no prediction collapse, and class-prior L1 drift of zero. The calibrated threshold accepted all 7,200 answers at 100% accuracy. A separately sampled, blinded sealed audit was signed by an independent gpt-5.5 reviewer and passed 400/400 with zero critical findings. Neither audit nor sealed metrics were used for selection or tuning.

Two clean CPU reloads and one clean Apple MPS reload agreed on 100/100 fixed nonsealed probes. Same-backend choice agreement was 1.0 with zero normalized- score delta. CPU/MPS choice agreement was 1.0, maximum normalized-score delta was 7.2023e-16, and aggregate accuracy delta was zero percentage points.

Input and compatibility contract

The governing sequence limit is 512 tokens and the question-head limit is 192. The question and every choice must fit intact. AtMem uses its shared tokenizer- exact whole-range packer and disables silent truncation; overflow produces an explicit receipt and deterministic fallback.

The artifact format is Safetensors plus Laya tokenizer/encoder configuration, questions.json, calibration.json, training-manifest.json, and artifact-manifest.json. The clean verification environment used Python 3.12, Laya 0.4.2, PyTorch 2.11.0, Transformers 4.57.1, Safetensors 0.6.2, Hugging Face Hub 0.36.0, and NumPy 2.2.3. CPU and Apple MPS clean reloads are verified. Linux CUDA was used for training. The full Windows/macOS/Linux installed-wheel compatibility matrix is a separate product-integration gate; do not infer unexecuted support from framework device names.

Escalation and failure behavior

AtMem should accept only finite, schema-valid scores bound to the exact model, question, calibration, tokenizer and request digests. Missing artifacts, identity mismatch, nonfinite/invalid scores, overflow, abstention, unavailable hardware, policy denial, stale state, or provider failure must not widen access or silently fall through to mutation.

The deterministic AtMem decision remains the offline fallback. If the optional AtBot escalation policy is enabled, AtMem sends only the minimum already- authorized input under explicit schema, budget and provider-egress controls. The provider response is revalidated through the same authority boundary; offline or invalid escalation returns fallback/review, not an unsafe guess.

Limitations

• Training and evaluation use synthetic fictional data from one policy family; perfect held-out scores can reflect generator/oracle regularity.
• No LoCoMo, LongMemEval, production memory, adversarial public corpus, or real-user usability result is claimed here.
• No matched Jev, Qwen, deterministic AtMem, retrieval, or store-then-retrieve superiority result is claimed here. Those require the separate frozen Spec 041 benchmark and safety/identity gates.
• Latency, energy, and peak memory vary substantially by backend. The full sealed MPS evaluation on a 16 GiB Mac was intentionally exhaustive, not a serving-latency benchmark.
• The model can confidently reproduce the synthetic oracle and still be wrong on novel, malformed, multilingual, adversarial, or out-of-distribution input.
• Public visibility does not make the model safe to invoke without AtMem's authority, packing, calibration, fallback, and receipt contracts.

License

The base model, pinned Laya source, synthetic dataset, and this derivative are Apache-2.0. See LICENSE in the repository. This card describes technical constraints and is not a substitute for reviewing the license text.