The thesis (owner's): slop is born from insufficient or poor context given to the model — treat it as an input bug, not a model property.
The mechanism: a model's prior is the corpus average (ready-made-concepts: post-RLHF, the ready-made categories are the mean of the species' inventory). An under-conditioned generation samples near the prior's center — and the prior's center is, by definition, the text that belongs everywhere and nowhere: the hedged phrasing, the listicle skeleton, the generic register that anti-slop detectors fingerprint. Slop is the posterior collapsing to the prior for lack of evidence. Context — the actual thread being answered, the product's real mechanics, numbers with provenance, the writer's own stance corpus — is what conditions generation away from the center to a specific, situated position. Sufficiently conditioned output isn't "slop that evades detection"; it is not slop.
Three corollaries:
- The anti-slop control is input-side. In an agent production line, the load-bearing component is the context pack assembled per artifact (target thread/query + product facts + voice corpus + the arena's SOP); the output lint is the backstop, not the control (from-sop-to-operation §4b).
- Slop debugging runs upstream: when a draft slops, diff what context was missing — falsifiable per artifact, fixable per pipeline.
- This is StandMeet's theorem wearing marketing clothes: a persona grounded in a curated corpus answers in a voice because the corpus conditions it off the prior — content production and audience-tailored introduction are the same problem, and the vault is the context factory for both.
- The voice corpus must be sourced, not assumed. Corollary 1 says the L3 voice layer does the work — but where does it come from for an owner who hasn't already written one? A description of the voice (a digest) is inert; only real prose conditions off the prior. So a general machine needs a bootstrap ingestion loop (ask the owner → mine their traces → else collect a peer corpus for tone), and where authentic aliveness can't be sourced it must route to the human, never fake it (corpus-ingestion).
- A statistical machine-feel detector is a soft tripwire, never a judge. The output lint (corollary 1's backstop) can be augmented with an AI-text detector (e.g. Binoculars, run locally) that scores how prior-center a draft reads and, above a threshold, routes it to human review. But detectors are unreliable and biased against non-native / idiosyncratic prose (~61% false-positive on non-native TOEFL essays in one study) — so the owner's authentic voice will false-positive. Therefore it is soft, advisory, human-overridable, and it flags machine-feel without fixing it (the fix stays corpus + human). A hard gate on a biased detector would suppress the real voice the machine exists to produce — the exact inversion of the goal.
- Slop has a second axis — mid — that context/register can't reach. Off-center-in-voice is necessary but not sufficient: a draft can read human and still be the safe median take (obvious, no stance, could be said by anyone). That is a stance/risk deficit, not a register deficit — no statistical detector catches it, only a judgment does. It is this note's "taste residual" made a first-class target, handled by an anti-mid strategy + auditor (anti-mid): take a falsifiable position, say the non-obvious thing, and sharpen-or-route rather than ship the median.
The honest bound on "always": long generations drift back toward the prior mid-stream, and the taste residual (jenny-hoyos-craft-is-scrapeable's admission — rules don't cover it) is not purely a context variable. As an engineering default the thesis holds: suspect the context first, the model last.
The controlled experiment (2026-07-16) — confirmed, with a refinement
A 3-arm blind test: the same FlexMesh Reddit-reply task at three context thicknesses (A thin/bare task · B +product-facts · C +full manifest: room culture, SOP rules, voice, the maker-not-user disclosure doctrine, and a "give the free manual method first" instruction). Generators were blind to the experiment; three independent blind judges scored each draft 0–10 (0 = generic slop, 10 = grounded/room-native). Mean slop by arm: A 5.83 · B 6.00 · C 8.17.
- Directionally confirmed: thick context roughly halved the distance to the ceiling (5.83 → 8.17).
- The refinement — not all context is equal; product facts are nearly inert. B ≈ A: adding product facts did almost nothing, because the base model already holds category facts (they're in the prior). The entire lift came from the L3-ish layer, and the judges' own reasons name what worked: a specific situated move the prior doesn't surface (the FSA-presort free method — "real depot knowledge nobody fakes"), the honest disclosure structure ("I'm the dev not a driver," give the free win first), and register (no marketing cadence). So: slop is conditioned away by norms/voice + specific situated knowledge, not by facts. Generic facts are inert; the specific, the situated, and the register do the work. This sharpens corollary 1 — a context pack heavy on facts and light on voice/norms/specific-moves will still slop.
Honest bounds on the experiment (it is a directional pilot, not proof): n = 6 drafts / 3 judges, underpowered; the judges are the same model family as the generators (shared blind spots; "AI-judged slop" ≠ a real courier's read); every arm still had the rich L0 (per-execution) trigger thread, so none was truly prior-center — the test probed the top of the context range, making the effect size a lower bound; and the thick arm was seeded a specific content move (the free method), so part of C's win is "a better idea supplied," not pure voice — fair as a test of the whole pack, not a clean isolation of register alone.
Prior evidence: the executor acceptance test (2026-07-16) — a clean-context agent given rich SOPs + product docs produced simulated Reddit replies with zero slop register, grounded in actual product mechanics; the same model under a thin prompt produces the fingerprint every detector catches.