aka “logos, ethos, pathos — and the agents who check them”
# field agents on every page — rhetoric, operationalized
Page-local research agents that continuously map the living field around each investigation on this site: prior art, credible actors, felt pain, counterexamples, and translations for whoever just arrived.
Aristotle shipped the taxonomy — logos, ethos, pathos — twenty-three centuries ago. Rhett gives each appeal an agent, then promotes two duties the triad already implies into agents of their own: a falsifier, because logos that never meets a counterargument is just assertion, and a translator, because fitting the argument to the audience was never a fourth appeal — it was the whole art.
Findings arrive as structured cards, not generic prose: typed relations, sources, confidence scores, review status. The goal is not to generate SEO sludge; the goal is pages that feel alive, falsifiable, and connected to reality.
A project page is a claim frozen at publish time. The field around it keeps moving — papers appear, companies fund the same thesis, counterexamples surface, communities argue — and the page hears none of it. Worse: pages only collect supportive evidence, which is exactly why nobody trusts project pages.
Every investigation should be kept connected to reality by agents that search for rational grounding, credible actors, and felt pain — and, centrally, for evidence the page is wrong. A page that can say "this paper weakened our claim" is worth more than a page that was never tested.
the roster
Rational grounding: papers, specs, RFCs, benchmarks, prior implementations. Is this coherent? Has it been tried before?
Credible actors: builders, labs, standards bodies, maintainers, funders. Who has standing here? Who is already working near this?
Why anyone cares: demos, talks, threads, user pain, "this should exist" moments. Where does the idea click — and for whom?
Hunts evidence the claim is wrong, already solved, overbroad, or not worth building. Not an add-on — the credibility moat. Every falsifier card carries a thesis-impact verdict, up to "falsifies current version."
The same primitive means different things to a browser engineer, an agent builder, a founder, a standards person, and a curious outsider. Translation is the conversion layer.
Appears only when there is a good pun in the backlog. No forced whimsy; the joke should feel discovered, not generated.
the form
supports, contradicts, falsifies, complicates, reframes, adjacent prior art, already solved, credible actor, user pain — a card declares what kind of edge it is, which avoids the usual related-links graveyard.
Every card carries its source URL, dates, originating agent, a confidence score, and a human-review status. Visible uncertainty over false fluency.
Short cards, specific claims, citations, and a one-or-two-sentence why-this-matters. That makes the agents epistemic infrastructure, not content spam.
Falsification claims, already-solved claims, humor, and anything naming people or orgs in sensitive contexts publish only after review. Autonomous outputs land in an intake queue, never directly on the page.
Every element of this form already ran once, by hand, in a thirty-hour sprint: a research super-repo that captured 176 external references into a schema-validated manifest, linked them through a 98-concept ontology with 365 typed edges (supports, defines, candidate_supports), pinned every claim to an upstream commit with a confidence level, and kept a reverse-graph queue proposing what to capture next. Rhett's card grammar wasn't designed on a whiteboard — it was rediscovered from that run.
specimen
Phase zero is hand-authored example data. Real findings, filed the way the agents will file them — against live investigations.
vercel-labs/just-bash — "Bash for Agents": a simulated bash environment with a virtual filesystem, in pure TypeScript, threat model included. The agent's shell becomes an interpreter; the world it touches is virtual; there is no real process to leak from.
adjacent_prior_art → ezaf · thesis impact: complicates. It answers the sandbox question EZAF's effects doctrine starts from — no firewall, no proxy, no temp credentials — by emulation: there is no real bash. A credible team making the cage tractable presses on "the sandbox is the wrong boundary."
Read closely, it concedes the point: just-bash works precisely by abandoning the real Unix process. The interpreter is a handler that owns the whole world — emulation is effect routing with every route pointed at a simulation.
agent: logos · source: github.com/vercel-labs/just-bash · confidence: 0.9 · review: human_reviewed · follow-up: add emulate-vs-relocate to the EZAF effects section — already shipped; see the EZAF page.
s2.dev, filed by the ethos agent against webpoke: "the API for unlimited, durable, real-time streams" — append-only, ordered, readable from any retained point, with fencing tokens and a self-hostable lite build. The pullable ledger as a product, and the buffer desh's channels POC already replays through. Relation: credible_actor; impact: supports. One card complicates, one supports — the tray should never be all good news, or all bad.
the intake queue
Cards accumulate faster than prose. These are queued — typed, targeted, human-reviewed; the component that renders them properly is phase zero of the roadmap.
→ desh, webpoke · logos · supports. The canonical essay: the append-only log as the unifying abstraction of real-time data. AYNIL — all you need is log — is its thesis compressed to a tagline; webpoke's pullable ledger is its integration doctrine. The ancestor card.
→ desh · logos, ethos · supports. Depots are append-only logs, PStates are derived projections — an entire backend programming model that takes AYNIL literally, at industrial scale. Evidence the lab's substrate bet has serious company.
→ desh · implementation_reference · supports. Durable execution as a service: workflows replayed from event history, resumable mid-flight. desh's pipeline-to-prod already names it as a target engine — the typed graph is the half it doesn't sell. The lab built that half once: an agentic workflow cookbook runs one recipe spec under both Temporal and AWS durable Lambdas, parity enforced by shared scenarios (see desh's artifact stack).
→ desh · logos · supports. Orchestrations replayed from event-sourced history, shipped by a hyperscaler. Proof the replay-the-log programming model survives enterprise contact — and the cookbook's second engine.
→ desh · pathos · supports. Tiny tool, enormous adoption: the environment should be a function of where you stand. "The prompt is a place" has been folk practice for a decade — direnv is the demand signal.
→ desh · logos · supports, caveat attached. Hash-addressed store, declarative derivations, hermetic environments — POSEVEN's instincts, shipped years ago. The caveat is the lesson nushell also teaches: the right substrate behind a grammar most people refuse to learn.
the gauge
Every investigation exposes its depth on a 0–6 scale — claim, diagram, mock, runnable demo, spec or repo, evaluation, deployed primitive — so incomplete work stays honest and legible instead of dressed up.
"Depth 1: diagrammed, not built yet" is a respectable thing for a page to say out loud. The stage chips on this site are the first pass at the same instinct.
Pages record when a finding moved the thesis. An investigation that admits a paper weakened its claim becomes more credible, not less.
The thesis is never hidden. Depth is: working notes, agent trails, and early artifacts can unlock through engagement, contribution, patronage, or collaboration. Premium means depth, not secrecy.
the loop
claim → field discovery → falsification → artifact → reaction → translation → contribution → funding → deeper artifact. Rhett is the mechanism that keeps every page connected to reality.
The lab does not try to be right first. It tries to make the right conversations possible — and the right conversation needs the prior art, the actors, the pain, and the counterexample in the room.
artifacts
agents, card schema, page model, phased roadmap — hand-authored, pre-implementation
the discipline, executed without agents: typed reference-to-concept edges, per-claim confidence, evidence pinned to commits, gap-driven discovery — plus an ontology-scored triage of 4,345 forked repos, which is the discovery pipeline's job description written as a one-off script
typed relations, visitor contexts, thesis impact, review states
hand-authored cards on one investigation page; components before pipelines
query generation → source adapters → dedupe → score → review queue; nothing auto-publishes
edges
Born in a conversation that opened with "if not you, who? if not now, when?" and closed with a roadmap. The name is the rhetoric, abbreviated — and this page is the first thing the field agents will be pointed at.