psikosen/ResearchSquad-Experimental-Preliminary-v0.1

ResearchSquad — EXPERIMENTAL AND PRELIMINARY

Status: early research artifacts, not a production model. Published 16 September 2026. This repository contains three separate, small checkpoints selected from local experiments. They are not a combined model, a general-purpose assistant, or a qualified autonomous researcher. These particular checkpoints use floating-point weights; they are not ternary models, despite the broader laboratory's ternary research objective.

Best observed checkpoints

Selection warning: these are the highest observed checkpoints on exposed public synthetic fixtures. Their individual scores are selection-biased estimates, not untouched final-test results. All-seed results appear below and in results.json. All 18 prior memory checkpoints and their parents were also preserved locally; 18/18 reproduced their predictions after an archive restore.

Memory experiment

The memory models are randomly initialized, two-layer causal transformers with width 128, four query heads, two KV heads, FFN width 256, vocabulary 65, and context 16. The exported model has 306,323 parameters, including unused, untrained decision heads. Those heads have no claimed task competence.

The vocabulary is a local character scheme, not an English-language tokenizer: BOS=0, separator=1, 62 alphanumeric characters at IDs 2–63, EOS=64. Inputs are BOS + six-character key + separator; generation predicts four value characters and EOS. See memory-fixture.json for the exact mapping and all public records.

A parent learned 32 random associations in 400 full-weight updates. Each of ten later tasks introduced 16 new associations with 200 rank-16 LoRA updates, alpha 32, learning rate 0.003, batch 16, AdamW betas (0.9,0.95), weight decay 0.01, gradient clipping 1, and a ten-update warmup. The optimizer resets per task. Replay uses actual earlier records; the parent-preserving version allocates half its later replay batches to starting associations. LoRA matrices merge between tasks. Training used FP32 master weights and BF16 CUDA compute on an RTX 5090.

These are memorization scores on previously presented associations, not unseen-question generalization. All methods initially acquired their new tasks at 100%; the reported measure is retention after all ten tasks. Seeds share the same associations and task order. The training origin is local synthetic data; no external pretrained model or downloaded corpus supplies these weights.

Temporal planning experiment

This is a reduced, action-conditioned world-model experiment inspired by Temporal Straightening for Latent Planning, not a reproduction of the paper's full visual system. It is unrelated to the memory model's weights or objective.

The observation encoder maps seven features through a width-32 tanh layer to eight latent dimensions. The residual dynamics model takes eight latent values and four action coordinates through a width-64 tanh layer. It was trained in FP32 on CPU for 1,200 Adam updates, batch 128, LR 0.001. The treatment adds 0.01 times displacement curvature to latent prediction MSE with stop-gradient targets. Stationary observed transitions are excluded from the angle loss.

Training uses 512 whole random-walk episodes and development uses 128 other episodes on two known 7×7 maps. Evaluation uses 32 new goal pairs per map. States can overlap; this does not measure unseen-map transfer. One map is open, the other has a wall and one-way passage. The planner optimizes soft categorical actions, then executes discrete actions in the real environment: horizon 8, 30 gradient iterations, at most 20 executed actions.

The same 64 goals recur across seeds, not 192 independent goals. The exact known-dynamics shortest-path solver succeeds on all 64. Two seeds worsen on the difficult map (13→11 and 10→9 successes); the third improves 4→13. Reduced curvature alone is not evidence of better planning. This remains an experimental branch, not a replacement for exact planning when dynamics are known.

A separate Dream-RSI-inspired numerical search test found that a plain stop rule matched the replay-selected policy on 32 additional fresh problems. No Dream-RSI-trained model is included, and no incremental advantage over that simple control was established.

Load and reproduce

These are ordinary PyTorch models with tensor-only model.pt state dictionaries. No Transformers AutoModel, hosted inference endpoint, or remote-code loader is required. Dense memory exports use the original effective matrices W + 2BA; optimizer/RNG states remain in the local frozen archive rather than this inference bundle. Original checkpoint hashes are recorded in each config.

Download all files into one directory, inspect the short modeling files, and use Python with PyTorch 2.11.0. The tested CUDA runtime is PyTorch 2.11.0+cu128; the planning model runs on CPU. Choose a PyTorch wheel appropriate to your hardware.

hf download psikosen/ResearchSquad-Experimental-Preliminary-v0.1 --local-dir researchsquad-preliminary
cd researchsquad-preliminary
python verify.py --device cpu
# On a compatible CUDA device, reproduce the reported BF16 memory path:
python verify.py --device cuda

The verifier checks checksums.json, memory recall, and saved planning trajectories. CPU FP32 memory outputs are reported separately from the reference CUDA BF16 path; exact equivalence across arbitrary devices/releases is not promised.

from verify import load_memory
model = load_memory("memory-best", device="cpu")
# model accepts integer tensors shaped [batch, sequence], with the custom vocabulary.

All inputs shipped here are locally generated public synthetic fixtures. No credentials, host paths, protected evaluator assets, or real user records are included. The package and loaders are locally checked before upload. The wider laboratory's 44-test regression suite passes; none of that grants automatic action authority or production qualification. Use these artifacts to study retention and small-world planning mechanisms, not to make general capability claims.