Maverick03511/causal-memory-routing-lab

Causal Memory & Routing Lab: synthetic evaluation archive

Code and protocols · Interactive explorer

Historical v0.1.0: 208,210 scalar metric rows from eight completed result records, plus nine full trace archives including the failed recurring-memory-v1 export. The earlier benchmark-v1 metric issue is retained. Archive payload totals 728,066,833 bytes. QCA synthetic arrays/receipt are in qca-synthetic.json. No human-subject, private project, peptide or web data.

Two-speed follow-up

The additive two_speed_metrics configuration contains 130,205 scalar metric rows from six independent coefficient worlds, eight stream seeds per world, eight cases and eleven policies. The two configurations contain 338,415 rows in total. The new rows add a world field; model also identifies that world. Streams/cases within a world are dependent; the six world means, not individual predictions, are the independent interval units.

Scientific qualification: FAIL. Causal controls pass; adaptation, retention, quality and the adaptive-window minimum-value gate fail. A small observed balanced advantage does not satisfy the frozen minimum-gain interval margin. Read the interpretation and frozen protocol.

two-speed-index.json supplies the exact size and SHA-256 of artifacts/two-speed-v1.tar.gz. This tenth archive contains 384 numeric NPZ streams, fitted synthetic checkpoints, frozen sources, all validation candidates, SQLite feedback and verification receipts. Load NPZ files with allow_pickle=False. Rows include exposed evaluator answers; neither the archive nor its seeds are a new hidden test set. Previous archives and v0.1.0 tags are preserved. No pretrained/peptide weights or private project data.

Dataset structure

The metrics configuration has one evaluation split. Each row contains: study (run ID), case, policy, model (initialization identifier when applicable), seed (stream seed), metric (dot-separated numeric endpoint), value (float), and record_status. Metric units depend on the endpoint: loss is natural-log nats, accuracy/error/coverage are fractions, response is feedback steps, and count fields count examples/events. None values are omitted. Rows are not independent examples: many endpoints/models share the same stream.

research-index.json retains study summaries, gates and original receipt digests. artifacts-index.json specifies exact archive sizes, hashes and URLs. Each artifacts/RUN.tar.gz contains original synthetic JSONL trace paths and representative synthetic-feedback SQLite databases when available. Frozen toy .npz checkpoints, protocols, result JSON and sanitized historical proofs are in the GitHub repository. There are no pretrained models.

Generation and intended use

Data came from the registered local generators, seeds, calibration, training, validation and test flows documented in the code repository. The permitted use is audit, reproduction and study of bounded synthetic routing/memory behavior, including negative results. The research was developed with Codex assistance and has not been peer reviewed.

Leakage and limitations

Raw traces deliberately contain evaluation labels, private rule/context IDs, switches and other scoring metadata. Those are never policy inputs. Current predictions precede feedback; SQLite retrieval requires step < cutoff. Read the field boundary before reuse. This is an exposed evaluation archive, not a fresh held-out benchmark or recommended training corpus. Training/selection on it invalidates subsequent held-out claims on these streams. New confirmation needs fresh seeds/protocols.

Do not combine incompatible studies into an overall performance score. Known experts, hand-designed observations, unequal memory/compute budgets and conditional confidence intervals limit generalization. The initial AR, matched AR, sensor primary-value and forgetting qualification failures are visible; local recall/confidence successes do not establish semantic memory, novel task learning, quantum advantage, general architecture superiority or AGI.

License and citation

Code and generated synthetic data: Apache-2.0. Cite the repository release and the primary papers listed there. NVIDIA is an external reference, not an endorser. Publication excludes workstation inventory, credentials, chat logs, old projects, environment folders and unrelated security research.