3 of 5 value props fully backed · 2 evidence-backed (merged PR / recorded run, not a live check) · 0 learning goals.
- An agent cannot perform an action its role doesn't own.
- vs raw Claude Code: a delegated capability, not ambient trust — overreach is a denial, not a hope.
- A privileged effect not produced by its owning actor fails verification.
- vs trust-the-signature: a valid signature isn't enough — the producer must OWN the effect, or it's an orphan/ambient effect.
- A beads work unit can travel intake → merged PR through one signed pipeline. (evidence-backed)
- vs GitHub Actions: authoring → merge with a content-addressed provenance chain, not CI orchestration.
- A derivation's identity is content-addressed: any change to its manifest changes its id.
- vs a mutable build record: the id IS the hash of the manifest, so a step can't be edited after the fact without becoming a different derivation — the chain is tamper-evident.
- You can see what each work unit cost. (evidence-backed)
- vs raw Claude Code: per-actor/unit usage is projected from the audit ledger.
(none)
See docs/value-props.md for the forcing functions + the code→value-prop map.