Relayloom

Pay for what arrived. Prove it.

Relayloom matches a supplier bill with the purchase order and with what was actually counted on the dock, holds any payment the goods do not support, and writes signed receipts you can verify with the network off. Muse Glimmer and SAM 3.1, open weight Meta models, run on your own machine.

The paperwork says 50. The dock got 46. relayloom check says HELD, and says why.

python -m pip install relayloom
relayloom doctor

Try it in a minute, no models needed

bill.json (what the bill says), po.json (what was ordered), and a count of what arrived:

relayloom check --bill-json bill.json --po po.json --counted 46
relayloom verify relayloom-out/receipts.jsonl

Bill CAL-0990: billed 50, counted 46. Result: HELD
  - CARTON-HD: billed 50 with 0 already billed, received 46
Signed receipts: relayloom-out/receipts.jsonl   Verify: relayloom verify relayloom-out/receipts.jsonl

Exit code 0 means clear to enter, 3 means held, 2 means it could not run. Add --json for scripts.

Count the goods from photos

python -m pip install "relayloom[models]"
python -m pip install git+https://github.com/ultralytics/CLIP.git   # text prompts for SAM 3.1 (a Git-only dependency, so it is not bundled)
export HF_TOKEN=...            # your own Hugging Face read token (SAM 3.1 needs the access request)
relayloom models pull          # one time; after this everything runs offline
relayloom models serve         # Muse Glimmer on 127.0.0.1, in another terminal
relayloom check --bill bill.pdf --photos dock/ --po po.json --agent

With --agent, Muse Glimmer proposes the words to look for, calls SAM 3.1 as a tool, and looks at the numbered overlay itself. Code owns the number: it is only ever the masks SAM returned, and a count is accepted only when two different words agree, or when Glimmer has looked at the overlay and no other count contradicts it. Otherwise the bill is held for a person.

What it does not do

• It does not send a bill, a photo or a number anywhere. relayloom doctor and verify make no network calls at all.
• It does not pay anything. It decides, with evidence, and a person approves what the policy says needs approval.
• It does not claim real time on a CPU. Counting a photo takes about 40 seconds and reading a bill several minutes on a laptop CPU; measured timings are on the Proof page.

Commands

Docs, the policy file and the receipt format: https://relayloom.vercel.app/developers. Source and threat model: https://github.com/Sriharsan/relayloom. MIT licensed.