A homage

One constraint program. Four artifacts that can't drift.

A love letter to ERPBench and its Anchor method: 300 verifiable, long-horizon agent tasks inside a real ERP, where the instruction, the environment, the solution, and the grader are all compiled from a single CP-SAT spec, so they can never disagree.

01 · the anchor

One spec radiates to four aligned artifacts.

Write the workflow once as a parametric constraint program. The solver certifies an optimum per sample; deterministic compilers emit everything else. The instruction, the environment, the solution, and the grader all descend from the same source, so they are aligned by construction.

The constraint program at the center pulses outward to four artifacts. They light together because they were emitted together. Alignment is not enforced after the fact; it is inherited.
02 · artifact drift

Authored independently, the pieces disagree.

When a task's parts are written by hand, separately, they drift: unsolvable tasks, broken grading, reward hacks. Anchor's single-source compilation is the fix. Watch the two sides.

Left: four independently authored artifacts jitter and slide out of alignment. Right: the same four, compiled from one Anchor spec, stay locked in phase.
03 · constraints across the entire board

The ERP, modeled as one constraint board.

Procurement, manufacturing, sales, and finance are not four separate puzzles. They are one coupled system, and the constraint program treats them that way. A stock decision ripples into build schedules, invoices, and cash. Constraints propagate across the whole board at once.

Four ERP domains coupled by shared constraints. A perturbation in one propagates to all, exactly as a real business process couples them.
what the tasks are

Verifiable. A solver-certified optimum exists for every instance, so correctness is not a judgment call.

Open-ended. Many valid end-states satisfy the constraints, not a single scripted answer.

Consistent, tunable, scalable. One spec means no drift; parameters tune difficulty; mint fresh instances at will.

why a real ERP matters

These tasks run inside Odoo 19, a production-grade open-source ERP, across procurement, manufacturing, sales, and finance. The agent works against the same messy, stateful system a human operator does, with 40+ widgets, stock levels, open orders, and due dates.

That realism is the point. A benchmark in a toy sandbox measures toy competence. A benchmark in a real ERP measures the thing we actually need agents to do.

04 · the certified optimum

Not pass or fail. A sliding scale.

Because the solver certifies the optimum, a solution can be scored on a continuum: how close to perfect, given the constraints. That is a fundamentally richer signal than binary grading, and it is only possible because the optimum is known.

A trial lands somewhere on the scale from zero-point to the certified optimum. "Good enough" and "perfect" are distinguishable, and the solver draws the line.
minimize   cost(workflow)   subject to   procurement ∧ manufacturing ∧ sales ∧ finance   // one program, whole board
05 · why it matters here

The ERP twin of AssetOpsBench, and the substrate OrgOps needs.

ERPBench is the business-process analog of AssetOpsBench's industrial domain, and it lands exactly where this body of work needs it.

the substrate thesis, made concrete

Anchor formalizes a real business process pulled from an ERP as a constraint model. That is the process-mining substrate the Cooperative AI work proposed: extract organizational ground truth from the systems that already run the business, and turn it into something agents can be measured, and eventually controlled, against.

ERPs and CRMs are the boring tools that already hold the process data. ERPBench treats them as the source of truth, not as plumbing.

constraints enable the long horizon

Applying constraints across the entire board is what makes extremely long-horizon tasks tractable and gradable. It is the same principle as the validation cascade and SHACL shapes: deterministic, formal constraints do the heavy lifting so that expensive, fallible reasoning only spends its budget where it matters.

Anchor is that idea at benchmark scale, and it is the cleanest demonstration of why constraints, not bigger models, are what push agents past the reliability wall.

the debt of gratitude
This homage exists because ERPBench makes a hard thing look inevitable: that the way to benchmark, and one day to govern, long-horizon agent behavior is to formalize the real process as one constraint program and let everything else descend from it. That is an idea worth spotlighting.
domainbenchmarkwhat it formalizes
Industrial asset operationsAssetOpsBenchsensor/FMEA/work-order control loop
Business processes (ERP)ERPBench / Anchorprocurement/manufacturing/sales/finance as one constraint program