Meridian
A supply-chain agent that remembers, simulates, and cannot talk past its guardrails
The most dangerous recommendation this agent could make looks completely reasonable: cancel a late order and re-buy elsewhere, for a few dollars in fees, while breaking a volume commitment signed three weeks earlier. Memory tells the agent the promise exists. Code makes sure the promise matters.
- Holdout decisions
- 14 / 14
- matched the answer key
- Trap scenarios
- 4 / 4
- bad precedent rejected
- Hidden patterns found
- 3 / 12
- probed from memory alone
- Mean decision time
- 60 s
- maximum 99 s
When a raw-material purchase order slips, a planner has to decide quickly whether to accept the delay, expedite, switch suppliers, move stock between plants, substitute a material, or cancel and re-buy. Listing those options is easy. The hard part is remembering what surrounds them: the supplier that always slips in November, or the commitment someone made in a negotiation six months ago.
Two sources of truth, kept apart
Narrative history such as notes, escalations and post-mortems lives in a Hindsight memory bank. Structured state such as purchase orders, commitments, scorecards and past decisions lives in SQLite. Memory is where the why lives, and the database is where the current facts live. A deterministic simulator prices every candidate action, a small LLM tool loop picks one, and the result is a decision card with simulated options, precedents, open commitments and a cited rationale.
Rules the model cannot argue with
- Numbers come only from the simulator. Dollar amounts and stockout days in the rationale are checked against simulator and database values, and mismatches appear as warnings on the card.
- Citations are verified. A cited document must be one the agent actually saw, dated on or before the decision date.
- Commitments are checked before acting. A recommendation that would breach an open commitment is rejected, and overridden after two rejections.
- Nothing from the future can leak in. SQL runs on as-of views, and memory hits dated after the decision are dropped.
- External actions are mocks. ERP, email and other outside calls are labelled as such and never actually made.
The application
Meridian wraps the agent in a React and FastAPI supply-chain application: a control tower, disruptions, purchase orders, production, materials, suppliers, demand, commitments and a decision log. Moving the business date replays the company exactly as it looked on that day. Accepting a recommendation writes the decision and its commitment back, so the next disruption can learn from it.
How much the numbers mean
Figures from eval/out/scorecard.md and the README in the repository. 167 offline tests were passing at the last recorded run, before the web interface was added.