How we work
The vocabotics Standard.
We're a trust brand in a low-trust market, so we hold ourselves to a published standard. Here's what every word on this site is meant to live up to — and how you can check that it does.
1. How we write — calm, clear, candid
We write for you, not at you. Plain English, short sentences, no hype and no doom. We name the real worry, then hand you an honest way through it. If a sentence would make a skeptical shop-floor veteran roll his eyes, we cut it.
2. Drafted with AI, verified by a human
We use AI to help produce our content — and we say so plainly, because practising what we preach is the demonstration. Every AI-assisted, human-verified piece carries a badge naming the expert who checked it, their credentials, the date, and the number of sources. A named human is accountable for every published claim. “Verified by our team” isn't enough — we name the person.
3. The citation promise
Every factual or numeric claim links to a real, retrievable source — the regulation, the paper, the vendor's own docs, not a blog summarising them. Volatile facts (pricing, model capabilities, regulatory status) are date-stamped, and our living documents show when they were last reviewed. If we can't cite it, we soften it to clearly-marked opinion or cut it. We have broken this rule, and section 6 says exactly how and how often — because a standard with no record of being broken is a slogan.
4. No dead ends
Every page ends by pointing you to the genuinely useful next step — sometimes a free template, not the expensive thing. Candour beats conversion-at-all-costs, and it earns the trust anyway.
5. The doctrine behind it all
This standard serves one belief: AI should amplify people, not replace them, and humans stay accountable. That's the Pledge, and it's the reason we built any of this.
6. Where we have broken it — the counted list
Three episodes we know of, and we publish all three rather than counting zero. ONE: our own TRIAD codec results log contains a signed confession of three fabricated numbers — an inflated +14%, a +0.8% tie that never happened, and a cross-format table — written up before the benchmark run was read; our status document says the rule was broken three or four times, which we treat as the more honest figure. TWO: we published “confabulation driven from 87% to 0.0%” on eight surfaces of this site, and when we audited ourselves we could find no benchmark, corpus, sample size or baseline model behind either number; both are withdrawn rather than restated. THREE: we published a “1B WaveCore trained to loss EMA 2.69”; the 1B label was aspirational, the largest artifact we actually trained is 292.6M parameters, and the checkpoint is gone. Each one has its own page, linked below, with the numbers and the diagnosis.
7. What we do when we find one
We do not lower the number quietly. We withdraw it, we say on the page that it used to be there and why it went, and we leave the correction visible afterwards — because a site that only ever gets more impressive over time is telling you something about its editing, not its research. The reports: /lab/triad-ternary-video-codec (the fabrication confession), /lab/never-lies-retracted (the two lies our own outside-in test found), and /lab/wave-native-model-architecture (the missing checkpoint).