Six workflows live. Three phases ahead.

Hardening the baseline coordination of the ten core specialist agents under a single central Orchestrator. The Orchestrator splits business objectives into subtasks, maps dependencies, routes tasks to selected models, and monitors thread execution. This phase establishes absolute workspace isolation, live approval queues, and verification checks. Every output is hashed, audited against safety checks, and posted to the ledger for QA attestation before delivery.

Memory records what happened. The Promotion Layer decides what has earned the right to influence a future decision. When an operator overrules a screen or rejects an output with a reason, that lesson is clustered with related ones, scored on how often it recurs and whether it has been contradicted, and tested against a falsifier. What survives is promoted into a durable rule — with the source decisions linked, never discarded.

Alongside it, the Brain: a live graph of the workspace vault that lights as agents recall from it during a run. Every competitor’s memory is retrieval — store more, search better. This makes the opposite claim: most of what happened should not influence anything, and here is the mechanism that decides.

Today the gate is a real guarantee — inside our runtime. A customer has to trust that we ran it. This phase changes who has to be trusted.

A workflow is a specification: nodes, edges, gate positions, policy. Hash it and you get a stable identifier for that exact workflow at that exact version. A receipt attests that this workflow, with these inputs, was approved by this operator at this time. Anchoring the hash — never the content — lets a client verify that what ran is what they signed off on, without our cooperation and without seeing anyone else’s data.

Application-level enforcement is advisory. This is the step that makes it checkable by a third party.

The engine already handles it. Alpha research is a pipeline: read the paper, extract the hypothesis, engineer the features, backtest, validate, check regimes, decompose against known factors. That is the same shape every Synthos workflow has — find, verify, judge, gate, act. A maker-checker split across specialists, a validator that never grades its own work, and a stopping condition checked by something other than the agent’s own claim are not new requirements here. They are what the platform enforces today.

Research is where this starts. Execution follows once the accountability question is answered properly — the moment a graph routes an order, “who approved this message” becomes “who is liable for this trade”, and that is a securities question rather than an engineering one. A template we could build quickly is not the same as a template we should sell.