Sloth Ledger — the receipt for your time

A ledger of the minutes you get back. You describe a chore in plain English, an open-weight model estimates what it costs by hand and whether it is worth automating, and every time an automation actually runs, the reclaimed minutes land in a real time-series database. The dashboard answers one question honestly: how much time did this actually save me?

Open-source AI is not decoration here. The judgement comes from an open-weight model (qwen/qwen3.8-27b, served by Groq's OpenAI-compatible API), the arithmetic comes from the database, and the deterministic rules floor still answers with no key and no network — so the ledger never stops working.

Live demo: https://sloth-ledger.onrender.com — a free Render web service that sleeps when idle, backed by Tiger Cloud. It is a shared ledger, so writes are rate-limited per visitor and the page labels itself a public demo.

Why the numbers can be trusted

The model is allowed an opinion, never a figure:

• If your description contains a countable quantity (40 receipts), the minutes come from count × per-item rate, a table you can read and edit.
• If you measured the chore yourself, your number wins.
• Only when there is nothing countable does the model's estimate stand, and it is clamped to 1–480 minutes.
• Totals and weekly rollups come from SQL, never from generated text.

In the demo below the model proposed 20 minutes for renamed 40 receipts; the ledger booked 13, because 40 × 20 seconds is arithmetic and arithmetic belongs to the database. Each card prints the sum that produced its figure.

Run it

python app.py            # http://127.0.0.1:8787

Python 3.11+, standard library, no pip install. With no .env the app uses the rules floor for judgement and a local SQLite file for storage — it runs on a plane, in a cabin, or with the Wi-Fi off. Add keys to switch the lanes on.

Deploy your own

render.yaml is a working Render blueprint: free plan, Python 3.12, health check on /healthz, one dependency (pg8000), and two secrets (DATABASE_URL, MODEL_API_KEY) left as sync: false so they are pasted into the dashboard and never committed. Point Render at this repo, apply the blueprint, and it builds itself. Step-by-step notes are in docs/deploy.md; the service as it actually runs — region, plan, environment, verification transcript — is recorded in docs/DEPLOYED.md.

The free instance sleeps after fifteen minutes of inactivity. That is survivable here because the reclaimed minutes live in Tiger Cloud, not on the instance's ephemeral disk: the web process is disposable, the ledger is not.

Lanes

Copy .env.example to .env and fill in what you have.

Lane	Setting	What it buys
Open-weight model API	MODEL_BASE_URL, MODEL_API_KEY, MODEL_NAME	Judgement from qwen/qwen3.8-27b on Groq (any OpenAI-compatible endpoint works, including a local server)
Rules floor	nothing	Same output shape, no key, no network, always available
Tiger Cloud / TimescaleDB	DATABASE_URL	Hypertable + continuous aggregate rollups

With no DATABASE_URL the store is one SQLite file. With it, reclaimed minutes go into a ledger.reclaimed hypertable and the top-automations panel reads a continuous aggregate the database maintains itself.

The schema

See docs/SCHEMA.sql. Two tables and one continuous aggregate:

• ledger.chore — what you did, what it cost, and the verdict (automate, delegate, or ignore forever).
• ledger.reclaimed — a hypertable of minutes actually handed back, one row per automation run.
• ledger.reclaimed_daily — a time_bucket('1 day', ...) continuous aggregate, refreshed by policy, so the dashboard query stays small as history grows.

The app connects as a least-privilege role (sloth_app), not the database owner. docs/SCHEMA.sql includes the grants it needs.

Demo data, honestly

The eight entries shipped in tools/seed_demo.py are a labelled synthetic corpus for a freelance designer — illustrative chores, not measurements of a real person's week. Replace them with your own before claiming any savings. Minutes booked in the screenshots come from those entries.

The hosted demo is a shared ledger, so its totals move as visitors add chores and run automations. The screenshots and the figures in the DEV post are the snapshot taken at publication; the live counters are allowed to disagree with them.

Tests

python tests/test_ledger.py     # estimate reconciliation, verdicts, idempotency
python tools/smoke_http.py      # walks the real HTTP surface (server running)

Both use the standard library only. SLOTH_FORCE_RULES=1 pins the deterministic floor so a model upgrade cannot change expected verdicts mid-suite.

Licence

MIT. See LICENSE.