Read quality

Every read-engine change runs the same filter.

Most poker tools eyeball the output of their AI and ship it. PokerReads doesn't. Every change to the read engine runs through the same set of gates — poker-correctness checks, frozen-champion comparison, founder review — before it touches production. The bar only goes up because every update has to clear it cleanly.

Tested Every read change
Compared Against the live engine
Founder-approved No silent promotions
Always learning Misses become permanent tests
Plain English: PokerReads doesn't claim its engine is the best. It claims the engine has been tested. Every read-engine change runs through the same gates before it touches production, and the founder reads candidate output side-by-side with the shipped engine on real, privacy-redacted evidence before any promotion. When a candidate fails, the failure case becomes a permanent test so the same mistake can't slip through twice. The engine quality bar only moves one direction.

Test before ship, every release.

PokerReads runs deterministic poker-correctness checks on every read-engine change — variant preservation, PLO exact-card discipline, mixed-game boundaries, citation safety, sample-size honesty. The check has to clear before a new prompt is even considered for production.

A new read must beat the shipped one.

Internal challengers compare against the engine that's currently live, not against a moving target. If a challenger sounds sharper but breaks evidence discipline or poker correctness, it stays blocked. No vibe-shift upgrades — the read has to be measurably better on real evidence.

The founder signs off on every promotion.

No read-engine change ships from internal scores alone. The founder reads candidate output side-by-side with the live engine on real, privacy-redacted evidence before any promotion. Fail-closed by design — the default when in doubt is to keep what's shipped.

Misses become permanent tests.

When a read change is rejected — for poker correctness, evidence safety, sample-size overclaiming, or anything else — the failure case becomes a fixture in the test suite. The same mistake can't slip through twice. That's the loop that makes every update sharper than the last.

What we won't claim.

  • We don't say the engine is the best. We say it's been tested.
  • No production prompt change ships from internal scores alone — the founder reads every promotion candidate.
  • Failed candidates become permanent test fixtures so misses don't recur.
  • Public claims about new gates (cross-model judging, predictive lanes, observed-feedback fixtures) wait until those gates actually clear production. If we haven't done it yet, it doesn't go on this page.

Want to inspect the product instead? Start with a real synthetic demo read, then bring your own notes when you're ready.

See a demo read