Retail POS — the model build-off
We dogfood every PlanSmith planner across multiple coding models. We lock the vertical, the discovery answers, and the feature ledger, then audit what each model builds. The planner is the product — these are the point-of-sale systems it produced.
How the builds ranked.
Saffron Till
Best Money ControlsThe cleanest money path of the three.
- Best at
- The cleanest money path here, checked against real data. A return reverses loyalty points in proportion and cannot be applied twice in either direction — points earned are clawed back, points spent are returned — refunds reverse gift-card value in the same step, and moving money needs a second signature from someone other than the person doing it. Tidy payment records, and the most thorough self-assessment of the three, which unusually matched what we found.
- Watch for
- We confirmed the returns ledger from totals and from the code rather than by driving one return all the way through live, and we did not independently retest the reports or the multi-shop workflows.
Cedar & Row
Most Self-CriticalCondemned its own work, then fixed it.
- Best at
- The most honest reckoning with its own risks. Its own review found a hole where one person could ring up a sale, pocket the cash and erase the record — so it removed that ability outright, routed every correction through Returns (which reverses money, stock and points together), and made it impossible to approve your own action. All three held when we tested them live.
- Watch for
- Its own sign-off was written before the final fixes and reads like a failure. The live database still held test junk from its own testing, and two of its reports disagree with each other about returns.Fixed since audit — the serious money defects were all closed 83 minutes after that sign-off — we checked the current code — and the demo database was cleaned before going live.
Bright Corner
Broadest SurfaceComplete and correct; one integrity miss.
- Best at
- A complete twenty-screen gift and discount shop with sound money handling — returns claw back loyalty points in proportion and cannot double-apply, you genuinely cannot approve your own action, and the fact that card payment is unavailable without a live provider is stated plainly in the payment dialog rather than buried.
- Watch for
- It did not actually compile for production — one missing line that only shows up outside development — while its handover claimed the build had been verified as passing. An honest-reporting miss on an otherwise correct build.Fixed since audit — the missing line was added, so it now compiles and the demo runs a genuine production build.
Models tested: GPT-5.6 Sol, Claude Opus 4.8, and Grok 4.5 — three builds of the same cashier-first POS planner (20 capabilities, sell through multi-store), each run to full Advanced scope in batch mode. These scores are ranked on the axis a real till lives and dies by: the money path. We drove each build and read its source, checking that a refund reverses stock, loyalty and stored value together, that a void cannot erase a sale from the reports, and that money-moving actions need a second signature from someone other than the person acting. All three keep card payment in an honest test mode. These grade money-path correctness and completeness — not visual design.