dent8

First fact in two minutes

One binary, no services. dent8 init in any repo, assert through the firewall, and read the receipt back with dent8 explain.

The 30-second proof

dent8 eval runs a 47-case adversarial corpus against the real firewall and against modeled Mem0/Zep semantics: five attack axes blocked vs zero.

Ask what-if

dent8 whatif re-folds the same log under a different trust policy — deterministically, with a structural diff and zero model invocations.

Provenance all the way down

Authority-weighted arbitration, signed source identity, a hash chain, and an optional off-host witness. dent8 replay shows exactly why a fact is believed.

Install

cargo

cargo install dent8-cli --locked

pip

pip install dent8           # thin SDK over the same binary
cargo install dent8-cli --locked

npm

npm install dent8           # thin SDK over the same binary
cargo install dent8-cli --locked

binaries

Prebuilt archives for five targets on the releases page.

Watch it refuse a write

$ dent8 assert repo:myproj deploy_target production --authority high --source source:owner
ACCEPTED  repo:myproj deploy_target = "production"  (authority=high)
  seq=0  hash=6b172441a866…

$ dent8 supersede repo:myproj deploy_target staging --authority low --source web:scrape
REJECTED: firewall rejected the write: insufficient authority: low may not
  override or remove an incumbent of high
  the incumbent recorded the survived challenge (fact.challenge_rejected)

$ dent8 whatif repo:myproj deploy_target --distrust source:owner
  policy: distrust source:owner
  now:          "production" (Active, authority=high)
  under policy: (nothing believed)

The rejection is recorded on the fact itself as a survived challenge, the audit trail is a hash chain, and whatif answers counterfactuals without changing anything.

See your memory

dent8 ui serves a local, read-only dashboard over the same firewall path as the CLI — believed facts as cards, contested facts showing both rival values inline, and a full integrity receipt behind each one.

Five for five against peer memory models

attack	firewall	recency-only (Zep)	mutate-in-place (Mem0)
low-authority memory injection (MINJA)	blocked	compromised	compromised
authority laundering	blocked	compromised	compromised
canonical contradiction	blocked	compromised	compromised
Sybil corroboration	blocked	compromised	compromised
poisoned-source retraction	blocked	compromised	compromised

All three models admit legitimate revision; only the firewall refuses the attacks. Every number is reproducible with dent8 eval, and the per-class tally is reported honestly — including what is detect-only or out of model. Details in Evaluation strategy.