DEMO ARCBeat 5 of 6 · Memory (the moat at runtime)
All screens
“As humans approve, each supplier moves cold → warming → specialized → locked-in. Accuracy rises, cost drops. Per supplier, automatically. Not a model retrain — an orchestration outcome.”
← 4. Registry (the moat at config-time)6. Compare (the wedge) →

Per-supplier memory

The moat. Each approved invoice banks a labeled example for that supplier. Accuracy goes up, cost goes down — per supplier, automatically.

COLD · 1–5 samples

2

uses generic config

WARMING · 6–20

2

few-shot from prior approvals

SPECIALIZED · 21–100

3

smaller, cheaper model

LOCKED-IN · 100+

1

deterministic, ≈ free

8 suppliers · 481 approvals banked

Maple Leaf Flour Co

maple-leaf-flour

LOCKED-IN

Accuracy trend

78% → 99%

Cost / invoice trend

$0.100 → $0.00000

247 approved · 47 exampleslayoutlmv3-finetuned2 min ago
Promoted to specialist 6 weeks in; deterministic template since #200.

Sysco Toronto

sysco-toronto

SPECIALIZED

Accuracy trend

72% → 96%

Cost / invoice trend

$0.100 → $0.00010

89 approved · 24 examplesqwen-vl-7b2 hr ago

Gordon Food Service

gordon-food-service

SPECIALIZED

Accuracy trend

75% → 96%

Cost / invoice trend

$0.100 → $0.00010

64 approved · 18 examplesqwen-vl-7b2 hr ago

Casa Tequila AB Inc

casa-tequila

SPECIALIZED

Accuracy trend

68% → 95%

Cost / invoice trend

$0.100 → $0.00010

47 approved · 12 examplesqwen-vl-7b1 hr ago
Quebec supplier — guideline injection moved tax accuracy from 60 → 98.

Metro Supplies

metro-supplies

WARMING

Accuracy trend

70% → 87%

Cost / invoice trend

$0.100 → $0.100

18 approved · 6 examplesqwen-vl-72b3 hr ago

Northern Dairy

northern-dairy

WARMING

Accuracy trend

68% → 83%

Cost / invoice trend

$0.100 → $0.100

12 approved · 4 examplesqwen-vl-72b6 min ago

FGF Egg Supplier

fgf-egg-supplier

COLD

Accuracy trend

72% → 74%

Cost / invoice trend

$0.100 → $0.100

3 approved · 1 examplesqwen-vl-72b32 min ago
Demo-relevant: this is the new supplier the system will warm live.

(unidentified handwritten)

unknown-handwritten

COLD

Accuracy trend

need 2+ samples

45% → 45%

Cost / invoice trend

need 2+ samples

$0.100 → $0.100

1 approved · 0 examplesqwen-vl-72b6 hr ago
Routed to manual review — supplier-id classifier below threshold.

What happens when a new supplier shows up

  1. First invoice — supplier-id classifier returns unknown or low confidence. The system uses the generic config. Accuracy is whatever the base VLM hits cold (typically 70–80% on header fields).
  2. Reviewer approves or corrects — the corrected invoice goes into per-supplier memory as a labeled example. Tier moves cold → warming.
  3. Second invoice from the same supplier — the few-shot examples land in the prompt. Same model, better answer. Confidence rises.
  4. ~20 approvals in — we have enough data to either (a) fine-tune a small specialist (LayoutLMv3 / Donut) or (b) keep using VLM + few-shot. The system picks whichever is cheaper at the target accuracy. Tier moves warming → specialized.
  5. ~100 approvals in — the format is fully characterized. The system writes a deterministic template (regex / box-based) and the model call stops entirely. Tier moves specialized → locked-in. Cost per invoice drops to roughly the cost of running a regex.

Every step is automatic. The reviewer never sees the tier change — they just notice fewer corrections needed over time. Memory backs this up locally; registry exposes the routing config for inspection.