ActTruth ACTTRUTH

Your AI stops lying about what it did — and stops acting without a budget.

ActTruth is a thin layer for tool-using agents: an act ledger (including quiet timeouts) and an unprompted spend allowance. The more autonomous the agent, the more you need receipts — not vibes.

What we prove SOVERYN house

Two steps

  1. See wrongness — quiet tool failures become durable FAIL rows, not silence.
  2. Learn from wrongness — the same FAIL pattern repeating arms a soft LESSON so agents stop blind retry loops.

Does not require a memory graph. Does not replace your model — bolts on beside tools.

Free core · paid house

ActTruth Tier 1 is free (Apache-2.0): ledger, budget, soft lessons, wrappers, local proof. Trust has to be runnable — not behind a license key.

SOVERYN is how you get a full crew under that truth — Command Center, HITL, memory, installs — and how we get paid (house + wiring/consulting).

Money sentence: ActTruth is free so trust compounds. SOVERYN and the people who wire it are how we get paid.

Proof is the product

Claims are locked in an open pytest suite — not marketing copy. Dogfood receipts stay on the operator’s machine until they choose to share. 

pip install -e packages/soveryn-acttruth   # from the monorepo today
acttruth proof --skip-pytest
python -m pytest tests/test_acttruth.py -v

We publish what is tested publicly. Live house fail rates and tool names are not scraped onto this site.