We won't manufacture trust. Not in a result, not on this page.

The rule, extended to ourselves

No false confidence is what the product is built around. A green or non-failing result never implies more evidence, authority, calibration, comparability, or safety than the run earned — a pass is a bounded, auditable claim, not a verdict on whether your output is correct.

We hold the site to the same standard. It states the full product vision, and the four-state badge shows exactly how far the code has actually reached — vision-forward, never false-confident. A capability marked next or experimental is built in the framework but not yet in the version the plugin installs, and nothing marked so is ever shown running as if it shipped. We will not manufacture product maturity any more than Core manufactures a green result. The ahead-of-code register keeps that promise auditable.

Where each product stands

The honest availability picture, stated plainly — the same discipline we apply to a run's verdict, applied to the products themselves. Each badge is the offer axis, separate from the capability maturity of what's inside.

• EvalGlass Core open source Available today — Apache-2.0, local-first, pre-alpha (v0.2.1, tagged). Runs on your machine and in CI; sends no telemetry; needs no provider keys; has no hosted platform.
• EvalGlass Discovery private preview Finds the evaluations your AI application is missing. In private preview — not generally available.
• EvalGlass Intelligence private preview Explains why a hard failure happened and makes the evaluation system learn. In private preview — not generally available.

See the three products → · Capability roadmap →

The verb we don't ship

The strongest honesty signal isn't a sentence on this page — it's a command that doesn't exist.

You operate EvalGlass through your coding agent with a small verb surface — /evalglass setup, connect, run, view, explain, compare, baseline, ci. There is deliberately no gate, approve, certify, pass, verify or validate verb — and that absence is the identity, not an omission. The plugin is a typist and a reader, never a participant: it asks, the host validates, and a single Verdict Engine decides. So a fresh run is informational, never "passing"; activating a gate is a host YAML edit you own. A tool that could quietly make its own result pass would be exactly the false confidence EvalGlass refuses — so that verb is not written, and CI keeps it that way. check-agent-verbs Where the claim stops →

What we refuse — and how it's enforced

The product refuses false confidence with a mechanical deny-list that fails the build if a fabricated trust signal slips in. We hold our marketing to the same bar. Where a refusal can be a build gate, it is — and we name the gate. Where it's still only our word, we say so, rather than imply an enforcement we don't have.

check-name the build fails if we break it  ·  our word a promise, not yet a gate

• No fabricated ratings, reviews, or trust badges. No invented star ratings, review counts, or download numbers — not in a page, not in its structured data. check-seo
• No self-approving or overclaiming capability language. No sentence that quietly lets a tool grant itself the authority to bless its own output. check-copy
• No self-granted authority command. No gate, approve or certify verb exists in the plugin. check-agent-verbs
• No "trusted by" walls, testimonials, or superlatives where we compare. How-it-compares carries no "we win" framing, no fabricated praise, and every rival links to its own source. check-comparison
• No overclaim tokens or invented handles in our share cards. Social previews are length-budgeted and honesty-scanned; no made-up account, no overclaim word. check-social-copy
• No invented price, fabricated customer proof, or unshipped product shown as real. Pricing publishes no number it hasn't set; the commercial product pages show only clearly illustrative examples; there is no /enterprise page and no fourth product name. check-guarded-register
• No fake customer logos, invented team bios, or manufactured headshots. We show no "used by" logo wall, no borrowed faces, no invented quotes or case studies — we have none to honestly show. our word

Dated, and yours to audit

A commitment you can't audit over time isn't a commitment. Every change to this page is a public git commit, so its history is dated and inspectable — you can see when it last changed and judge whether the rest of the site kept up. The date below is checked against the page's own git-derived metadata in CI, so it can't silently rot.

Last substantive change: 21 August 2026 · how we work →

And if any claim on this site overstates reality — a feature shown as shipped that isn't, a signal the project hasn't earned — open an issue. We treat an overclaim on the website as the same class of bug as an overclaim in a scorecard.