KINGY.AI← Local AI planner

LOCAL AI / YOUR OWN INVOICE

Four fields.
Check every value.

Run a short plain-text USD invoice on your Mac. Compare the proposed values with your source, then save a reviewed result.

1. Set up and run

Start with an Apple-silicon Mac. The recorded recipe used an M4 Pro with 24 GiB, Qwen3-4B Q4_K_M and pinned llama.cpp running on CPU. That sample does not qualify your computer. Windows remains unqualified.

Pilot software: fresh own-input model qualification is UNESTABLISHED. Resource admission deferred before inference; the historical bundled sample remains separate. Own-input package integration is pending.

Open the public invoice self-test →

Commands after the START-HERE setup

In the extracted invoice-work-kit folder, set INVOICE_RUNTIME and INVOICE_MODEL to the pinned files as shown in START-HERE.md. Run preflight before inference. These commands create fresh attempt directories; do not create or reuse those directories yourself.

python3 --version
python3 -B verify_kit.py
python3 -B preflight.py --runtime "$INVOICE_RUNTIME" --model "$INVOICE_MODEL"
python3 -B run_local.py --runtime "$INVOICE_RUNTIME" --model "$INVOICE_MODEL" --output self-test-01
python3 -B verify_kit.py --response self-test-01/receipt.json

Offline replay expects passed: true. A newly completed public sample expects self-test-values-match. If preflight refuses, resolve the reported prerequisite before continuing. The current candidate was refused for abnormal memory pressure and less than 15 GiB free disk; no fresh candidate inference is claimed.

Create my-invoice.txt as UTF-8 plain text, then:

python3 -B run_local.py --runtime "$INVOICE_RUNTIME" --model "$INVOICE_MODEL" --input my-invoice.txt --output my-invoice-01
python3 -B verify_kit.py --response my-invoice-01/receipt.json
python3 -B review_invoice.py --attempt my-invoice-01 --output my-invoice-reviewed-01

Own-input status needs-local-review, passed: null and checker exit 2 mean a value review is still needed. The attempt retains source, request, raw response, receipt and log when reached. The review command creates JSON, formula-safe CSV and a separate review receipt after explicit source checks. You can also open the source and receipt in step 2 below. Use a new attempt/export directory each time.

Run the bundled public sample first. For your own task use readable UTF-8 text, no more than 8,000 bytes, with USD stated explicitly. PDFs, scans, OCR, tables and other currencies are outside this release. Keep every attempt in a fresh folder.

What the preflight checks mean

The read-only preflight checks the pinned runtime and model, Python, supported platform, available port, memory pressure and 15 GiB free disk reserve. The reserve is an admission guard, not a measured minimum memory requirement. It installs nothing and does not stop other processes. Fix a reported prerequisite and rerun preflight; keep existing work.

Private invoice workspace

2. Compare source and proposed values

Open your saved source.txt and runner receipt.json, or paste them. An unedited imported file keeps its exact UTF-8 source for receipt hashing and attempt export, even when displayed line endings are normalized. Any edit discards that imported binding; reopen the original source file to restore it. Bare four-field JSON is also accepted, with response completion marked unknown. This page does not run a model.

Compare with an independently prepared reference (optional)

Only use a reference established by checking the source independently, ideally before seeing this output. A matching reference does not prove how it was prepared. The same model approving itself is not independent verification.

No input checked yet.

3. Save the reviewed result

Review all four fields before exporting. A valid shape alone is not a verified invoice.

These exports contain only the four fields. The CSV quotes cells and prefixes formula-like text with an apostrophe. Missing ID, total or due date, ambiguous bank details, mismatched references and incomplete responses block a full reviewed export. A bank account correctly absent from the source stays null. Your confirmation is a self-reported review, not an observed study result. No payment or other financial action follows.

Keep the same task when something fails

Preserve the source, original receipt and acceptance check. Rerun into a new attempt folder after a supported correction. Changing the source or relaxing the check does not demonstrate recovery.

This is your declaration; no hardware or installed software is detected. Choose the pinned setup only after the kit preflight confirms it. Unknown and other setups receive a useful no-match result.

No same-task fresh-inference invoice repair is qualified. Fix Finder offers scoped related client guidance and an honest no-match result; that guidance does not establish model accuracy or reader recovery.