v1.0.0stdlib only$0

Verdict CLI

A Cumulative Web Inc trust-layer product

Deterministic, evidence-bound trust scoring for AI agents — the CWI Verdict Engine v1.0.0 as runnable software. A verdict with receipts, or an honest insufficient-data. It never invents a score.

Try it now — running in under 5 minutes

No install beyond Python 3.8+. Clone the repo, print a valid evidence template, and run your first score — no network, no accounts, your evidence never leaves your machine.

→ Get the code on GitHub License it: hp@cumulativeweb.com

Prefer a live cross-check? cwi-mcp-server's trust_verdict tool runs the byte-identical vendored engine v1.0.0 this CLI wraps. Store-card licensing is PENDING — price on request, never published without approval.

What it does

You hand it evidence JSON — ERC-8004 identity signals, Needle Drop verified history, First Spin published verdicts — and it returns a verdict: a score (0–1) with a published band and a full evidence trace, or a refusal when the evidence isn't there.

Admission: only verified evidence counts. claimed/pending/refuted count zero. Any disputed item refuses the whole run.
Sybil damping: issuer caps, identity clustering, self-assertion discount — applied openly, all damped items stay visible.
Determinism: no clock, no randomness, no network. Same input → byte-identical output, bound by input_sha256.
Provenance: scored verdicts are unanchored unless your input pins evidence with an anchor_ref.

Quickstart

git clone https://github.com/CumulativeWebInc/cwi-verdict-cli.git
cd cwi-verdict-cli
python3 -m verdict_cli example > evidence.json
python3 -m verdict_cli score evidence.json --out ./run-001
# run-001/report.json  (canonical engine output)
# run-001/report.md    (why the verdict is what it is)

Commands

command	what it does
`score input.json [--out DIR]`	Score evidence; write report.json + report.md
`check input.json`	One-line verdict for pipelines
`example [--kind insufficient\|disputed]`	Print a valid input template
`spec`	Print the published scoring rules
`pin`	Show the vendored engine provenance (audit me)

Score bands

score	band	meaning
0.800 – 1.000	established	Deep, multi-family verified evidence
0.600 – 0.799	emerging	Solid evidence, still building
0.400 – 0.599	thin	Real evidence, but sparse
0.200 – 0.399	weak	Minimal verified signal
0.000 – 0.199	negligible	Barely above cold start

Honesty

This tool scores evidence; it does not create trust. Garbage evidence in, garbage score out.
insufficient-data is the correct answer for new subjects — not a failure.
No "scored" demo exists on purpose. A scored example would need fabricated evidence.
No telemetry, no network, no accounts. Your evidence never leaves your machine.

The engine is vendored, not rewritten

verdict_cli/_engine.py is the canonical CWI Verdict Engine copied verbatim from CumulativeWebInc/cwi-learn (branch needs/verdict-engine, blob e34b722e…29953, pinned 2026-09-17). Weights, thresholds, and gates are spec-versioned — never edited. The test suite asserts the vendored file's sha256.