Cohort Latent Fabric
Cohort Latent Fabric is ColabHive's opt-in execution lane for compatible causal/autoregressive LLM endpoints. Classic inference remains the default. Encoders such as BERT, embedding and reranking models, and diffusion models are not admitted.
Start with:
- Concepts and product boundary
- REST API contract
- Run a Cohort
- Python SDK
- MCP tools
- Flags, scopes and errors
- Trust and privacy
Requests must opt in explicitly and the selected endpoint must advertise Cohort capability. Incompatible or unhealthy paths fail closed unless the caller selects an explicit Classic fallback policy.
Research boundary
Status as of 6 September 2026: R4-CEM v6 is a sealed research study of textual sender-to-reader communication. It uses 600 public-benchmark items, 11 conditions frozen before scoring, and two reader backbones through the Classic API path. It does not test whether the production hidden-state Cohort lane caused an answer improvement.
R4-M is a separate, preregistered and outcome-blind study of query-blind textual memory. Its fixed schedule covers 78 LongMemEval items — 72 answerable knowledge-update questions and 6 abstention (false-premise) questions — and 3,120 item-runs, including 624 PM/PM-shuffled reader runs. An internal validity audit (2026-09-07) found that those two kinds are not one homogeneous set: the memory policy scored 0/72 on the answerable updates and 6/6 on abstention, so any aggregate over 78 reports abstention behaviour, not memory quality. R4-M is therefore reported as a diagnostic result, and its preprint was withdrawn on 2026-09-12. R4-M memory is a research intervention assembled by the evaluation harness; it is not a user-facing memory mode and is never enabled by installing the SDK or selecting Cohort execution.
The two studies estimate different axes and are not a memory-by-multi-agent factorial. Neither supports claims about BERT, diffusion, learned-code, KV-cache, hidden-state efficacy, arbitrary graphs or universal memory gains. The research page tracks the publication boundary separately from product availability.