Skip to content

How it works

Verification, made honest and hard to game.

Doestep answers one question: does this still work, on this model, right now? It answers that with evidence anyone can audit. First, why that question needs asking.

Not yet verified on Sonnet 5a newer model shipped · needs a re-run

One object carries the whole story: a fact, computed from real runs.

The problem

AI knowledge rots, and nobody tells you when.

01

It expires in weeks, not years

A prompt that worked in January behaves differently in July. The model changed underneath it and nobody told you.

02

Nothing is reproducible

“I built an agent that 10x’d my workflow.” Great. Can anyone else run it and get the same result? Usually not.

03

Everyone claims, no one proves

The internet is full of AI skills nobody can verify. Claims are cheap. A logged reproduction is not.

The loop, step by step

A playbook gets proven, a new model breaks it, the community proves it again. The model-release event is the heartbeat.

  1. 01

    Publish a playbook

    An author writes a testable playbook and freezes it as an immutable version.

  2. 02

    Doesteppers run it

    Other people actually run it, on a real model, with real inputs.

  3. 03

    Reports roll in

    Each run is an append-only repro report: worked, worked-with-changes, or broke.

  4. 04

    The badge goes live

    Confidence is computed from distinct reports, never written by hand.

  5. 05

    A new model ships

    The model registry gets a new row. This is the heartbeat.

  6. 06

    Badges flip

    Every affected playbook flips to “not yet verified on the latest model.”

  7. 07

    The community re-verifies

    Doesteppers race to re-run the top playbooks. The loop repeats.

The five things it’s built on

Every part exists to keep verification un-gameable.

Playbook

immutable versions

Every version is frozen. A report is meaningless unless it’s pinned to the exact version that was run, so editing in place is impossible by design. Versioning isn’t a feature; it’s what makes verification trustworthy.

Model registry

the heartbeat

Models, versions, and release dates. Adding a row, a new model release, is what flips affected playbooks to “not yet verified on the latest.” Freshness is relative to models, not the calendar.

Repro report

append-only

Every run is an event: who ran it, which version, which model, and the result. Never edited, never deleted. Corrections are new reports. Append-only means every badge is auditable back to its evidence.

Badge

computed, never stored

Derived from reports, the model registry, and time decay. Recomputed on every new report, on model release, and nightly. A stored badge could drift from its evidence. Derived state can’t lie.

Reputation

earned by doing

A Doestepper’s standing is computed from their report history with a weighted function. Gameable only by doing real work, and re-tunable later with no data migration.

The badge, up close

A badge that speaks in facts.

This is something no tutorial site, GitHub repo, or Discord has. Here’s what each part of it says.

The result
Worked, broke, or mixed: a status color and a glyph, never color alone.
The model
The exact model version it was run on. Freshness is relative to models, not the calendar.
The freshness
Recomputed on every report, on model release, and nightly. It decays on its own.
The reports
How many distinct Doesteppers proved it: the evidence behind the badge.

The states

VerifiedSonnet 5 · 3 days ago · 14 reportsNot yet verified on Sonnet 5a newer model shipped · needs a re-runMixed: see reportsSonnet 5 · 3 days ago · 9 worked / 2 with changes / 3 brokeBroke on step 4Sonnet 5 · 3 days ago · 1 worked / 4 brokeNot yet verifiedno reports on Sonnet 5 yet

Same object everywhere: on the site, in a README, in a tweet. Mono for the data, status color for the fact.

The trust rules

Built in from day one, because honesty is the brand.

  • Authors can’t verify their own playbooks.
  • A badge goes green only when weighted confidence from distinct users, with no follow relationship to the author, crosses a threshold.
  • Conflicting reports show “Mixed: see reports,” never hidden.
  • New accounts can report, but their weight stays low until they build history.

Ready to prove something?

Join the first Doesteppers and be there when the next model ships.