PortentRequest access
Trust & safety

The whole point is not being confidently wrong.

Portent is built like a security product, not a demo. Every claim is sourced, every number is provenance-stamped, and every irreversible action stops for a human. Below is how each guarantee is enforced, plus an honest account of what we have not yet proven.

Proof · live countsupdated on every promote
  • 342
    canonical lessons
  • ~10,700
    candidates in review
  • 109
    detect → fix patterns
  • 994
    tests green
  • 11
    layers
  • 100%
    retrievable · actionable · attributed

Five guarantees, each with how it's enforced

  • 01

    No hallucinated numbers

    Benchmarks come from real measured runs with full provenance: commit, image digest, command, timestamp. A number without provenance never ships.

    How it's enforced

    The formatter can only emit a value that carries a provenance record. A number missing commit + image + command + timestamp fails the output gate before it reaches you.

  • 02

    Human-gated for anything irreversible

    The agent stops for spending money, provisioning capacity, destructive operations, credentials, or evicting someone else's work. It prepares the action and hands you one command.

    How it's enforced

    Every planned action is classified. Anything touching spend, capacity, destruction, credentials, or eviction routes to a stop-and-confirm handoff, where the agent stages the exact command and waits.

  • 03

    Everything is attributed

    Each learned lesson links to the real event it came from. Unattributed folklore is explicitly rejected.

    How it's enforced

    The promote step rejects any lesson whose source-event link doesn't resolve. Attribution is a first-class axis of the coverage ratchet, checked on every promote.

  • 04

    Sensitive-data scrubber

    Account IDs, private IPs, internal hostnames, customer names, and keys are stripped before anything is stored or shown.

    How it's enforced

    A scanner runs before storage and again before display. Matches are redacted, and a regression suite guards against the scrubber ever regressing on a known pattern.

  • 05

    Reproducible

    Same inputs → same outputs. The determinism is testable and tested.

    How it's enforced

    Tools are pure functions. A determinism suite re-runs identical inputs and byte-compares the outputs in CI, part of the 994 tests kept green.

The provenance card

Every metric ships stamped with where it came from

Every benchmark takes this shape: a value plus the exact commit, image digest, command, and timestamp that produced it. The cards below show the format. The numbers are placeholders, marked as such, because we don't publish a benchmark until a real run backs it.

all-reduce bandwidthprovenance-stamped
—GB/s
commit
0000000
image
substrate/bench@sha256:…
cmd
nccl-tests all_reduce_perf -b 8 -e 8G
at
pending first live run
note
placeholder · not a real run
kernel speedup vs referenceprovenance-stamped
—×
commit
0000000
image
substrate/codesign@sha256:…
cmd
substrate bench --layer L0 --reps 30
at
pending first live run
note
placeholder · not a real run
determinism suiteprovenance-stamped
994tests green
cmd
substrate test --determinism --byte-compare
note
real count from brief §4; commit/image redacted pending public run

An unproven metric shows as a dash, never a guessed number. When a live run lands, the value and its full provenance replace the placeholder together.

What “not yet done” honestly means

We don't declare “done.” It's earned.

A capability is “done” only when three things are true: a live run reproduced it end-to-end, the closed learning loop captured what it taught, and the numbers are real and sourced. Until all three hold, we say so plainly. The candor is deliberate: a tool whose promise is “no fabricated numbers” cannot round its own status up.

  • ● reproduced live run
  • ● closed learning loop
  • ● real measured numbers

Public benchmark numbers on this site are placeholders until the first reproduced live run. Corpus counts above are real.