Contents
ZenithMovement
A research-grounded content engine that writes and publishes workout guides a brand can actually stand behind. It pulls real studies, sources its exercises, generates against strict anti-overclaiming rules, verifies every citation two ways, and won't publish anything that fails the check.
AI content is fast, and usually can't be trusted
Point a language model at "write me a hamstring workout guide" and you'll get fluent, confident prose in seconds. Along with invented studies, citations to papers that don't exist, and claims that sound research-backed but aren't. For most content that's a nuisance. For a site whose whole value is being grounded in real exercise science, it's fatal. One fabricated citation and the brand is worthless.
Zenith is built to make that failure impossible. The whole architecture exists to answer one question for every sentence it publishes. Can this be traced back to a real study that actually says it? If the answer is no, the sentence doesn't ship.
Six stages, each one a checkpoint
A topic and an experience level go in. A verified, publishable guide comes out, or it gets held for review. Each stage narrows what the next one is allowed to do, so the model is never free to invent the things that matter.
The research and the exercises are sourced before the model writes a word. The model assembles and explains. It doesn't get to make up the evidence or the movements. Then verification and a human gate stand between a draft and the live site.
Research, not invention
Every guide starts with real papers, pulled live from PubMed. The system builds a structured query for each topic, translating gym language like "quad workout" into the anatomical and biomechanical vocabulary the database actually indexes, then fetches and parses the real records. Nothing about the evidence is model-generated. It's all fetched data, normalized into clean objects.
Then each paper is scored: study type (a meta-analysis outranks a case report), recency, and a bonus for trusted journals. The part I like most is the relevance re-ranking. A keyword match isn't enough, because a paper can look on-topic and be about something else entirely. So the scorer boosts genuine training-science signals and actively demotes the false positives: clinical-rehab studies, supplement and ergogenic-aid trials, and adjacent-phrase traps that share vocabulary but not subject. Only the strongest, most relevant papers survive to inform a guide.
Generation on a leash
The exercises come from a vetted public-domain dataset, already sequenced into warmup, main, and cooldown. The model selects and explains the movements. It never authors the instructions. The recovery options are picked by ID from a vetted list, never invented.
The writing prompt is mostly a list of things the model is forbidden to do. Report what a study found, and never append your own conclusion. Don't generalize a result about shoulders into a claim about chest. Don't broaden a finding from one group to everyone. Don't treat a review or a proposal as a demonstrated result. Each research finding has to be a faithful report of exactly one abstract, nothing more. The brand has banned words and a fixed voice, calibrated per experience level. The model writes prose and picks IDs. Everything load-bearing is sourced.
Every citation, verified two ways
After a draft is written, every claim it cites runs through two independent checks before anyone reads it. The first is deterministic. Does the cited URL trace to a paper actually in the fetched research set? A model can't slip in a real-looking link to a study that was never pulled. The second is an auditable judgment call. The abstract and the claim are compared, and the claim is graded supported, overstated, or unsupported. On top of both, a topicality gate catches the subtle failure where a paper genuinely supports a claim but is the wrong kind of evidence. A supplement study can't be cited as proof about training.
And verification doesn't edit the draft. It reports. It produces a per-finding verdict and a pass/flag summary, and leaves the decision to a person. The system's job is to surface every problem, not to quietly paper over them.
Nothing ships unreviewed
Generated guides land as drafts in a review queue, each tagged with its status. Clean and ready, or flagged with the specific citation failures and any banned words it tripped. Review is fast because the system has already done the tedious part: tracing every source and grading every claim.
The publish step has a hard citation gate with teeth. It reads the entries about to go live and refuses to publish any with unresolved failures, whether unsupported, overstated, or off-topic. Publishing only ever includes verified content, and drafts stay private until a person clears them. The safety net isn't a guideline. It's enforced in the one script allowed to touch the live site.
Shipped, and self-maintaining
Zenith runs live. A scheduled job fills in missing topic-and-level combinations a few at a time, generating drafts until the library is complete and then going quiet on its own, since it skips anything that already exists. But it only ever makes drafts. It never publishes on its own.
When verified content is ready, the publish step exports it to the live Astro site, builds it first to catch any breakage before it goes out, then commits and pushes, and the host deploys automatically. The result is a system that does the heavy lifting unattended but keeps a human in the one place that matters: the decision to put a claim in front of readers.
What it demonstrates
The same discipline that runs through everything I build: a deterministic core, AI used only where its judgment can be audited, and an honest line between what's verified and what isn't. Zenith applies it to content. Real research, sourced facts, claims checked against their evidence, and a human gate before anything ships. It's the proof I can put a language model inside a production pipeline and still stand behind what comes out.