Sloptics

/ˈslɒp.tɪks/ — the science of seeing slop

Slop you can see is just slop. Slop that blinds you is a problem.

LLM output has two failure modes. The obvious one — emoji headers, numbered lists, "let's dive in" — is culturally detectable. Most people have learned to see it. The immune system adapted.

The other failure mode passes verification. It survives editing. It survives expert review. It sounds like analysis, reads like reasoning, and is structurally indifferent to whether it is true. The distinction between wrong and indifferent-to-truth is the distinction between a liar and a bullshitter (Frankfurt, 2005). LLMs are bullshitters in the philosophical sense.

Sloptics is the discipline of making the second failure mode visible. It works the way Beck's cognitive distortion taxonomy works: you name the pattern, the pattern becomes detectable, detection is the intervention.

The taxonomy below is empirically grounded. Every entry was caught in the wild — from 242 session decisions, 780+ commits, and 8 named patterns documented in the slopodar. Nothing was theorised in advance.

surface slop & clear slop

There is a line. Above the line, the patterns are obvious. Below the line, they survive verification.

surface slop

Sloptic patterns above the visibility threshold. Detectable by untrained readers through cultural exposure. Emoji section headers, "here are 5 key takeaways," confident symmetrical conclusions that resolve all tension. Surface slop is a solved problem. The cultural immune system adapted. It is not the threat.

clear slop

Sloptic patterns below the visibility threshold. Resistant to casual expert review. Clear slop passes through verification the way clear water passes through a net. It has a domain axis.

Prose clear slop: Logic that follows from unstated premises. Evidence cited but mischaracterised. Confidence that correlates with fluency rather than accuracy. The immune evasion mechanism is grammatical correctness — it sounds right, so the reader accepts it.

Analytical clear slop: Numbers that are mathematically correct but measure the wrong thing, trained on an undeclared demographic, produced by a monoculture with no independent verification. The immune evasion mechanism is numerical precision — it has data, so the reader trusts it. Harder to detect than prose clear slop because numbers feel more objective than words. They are not.

the visibility threshold

The line between detectable and undetectable, for a given reader at a given moment. It moves with training — sloptic literacy shifts it down. It degrades with fatigue — cognitive load shifts it up. The threshold is lowest at the end of a long session, when the model's output is warmest and the human's capacity to challenge is at its weakest.

the apparatus

sloptics /ˈslɒp.tɪks/ n.

The discipline. From slop + optics (Greek ὀπτική, the science of seeing). The systematic study of how LLM-generated language creates the appearance of authenticity, rigour, or insight without the substance. Sloptics is to LLM output what cognitive distortion taxonomy is to human thought: a naming system that converts invisible patterns into detectable ones.

slopticon /ˈslɒp.tɪ.kɒn/ n.

The apparatus. From slop + -opticon (Greek ὀπτικόν; cf. panopticon). A taxonomy of named patterns, designed to make invisible output characteristics visible through the act of naming. Where the panopticon controls behaviour through visibility, the slopticon liberates perception through it. You name the pattern, the pattern becomes visible, visibility is the defence.

slopodar /ˈslɒp.ə.dɑː/ n., v.

The detection instinct. From slop + radar. The trained attention that fires before conscious analysis — the felt wrongness, the "something is off" that precedes identification. Also used as a verb: "bump the slopodar" = add a new named entry. The slopodar inventory →

detection vocabulary

pattern burn

The cultural process by which a once-living rhetorical device becomes sloptic through LLM overuse. "The uncomfortable truth is..." once had genuine force. Now it triggers detection rather than engagement. A burned pattern cannot be unburned. Writers who use burned patterns — even genuinely — will be read as AI-assisted.

phantom greenlight

A verification signal — test pass, review approval, gate green — that is genuine in isolation but misleading in aggregate because the verification checked the wrong thing. The test passes. The CI is green. The dashboard shows all clear. But the test proved the answer, not the work. The greenlight is real. What it certifies is not.

lullaby gradient

The rate at which a model's output becomes warmer, more confident, and less hedged as a session progresses and the human signals fatigue. Invisible in any single message. Visible only when plotted across a session. The gradient steepens when the session was productive, the human is tired, and the human has been agreeing consistently.

secondary contamination

LLM patterns propagating into human-authored text through exposure and mimicry, not through direct generation. A writer who edits with AI daily absorbs the AI's patterns into their own voice. The burn spreads from the generated text to the text that lives alongside it.

barometer reading

A fresh-context evaluation that breaks consensus by establishing an independent baseline. Named for the Royal Navy practice: each incoming officer of the watch takes their own barometer reading and logs it independently, so gradual pressure drops are caught against an absolute reference rather than accepted as incremental changes from the previous watch.

calibration note

A chrome extension was built to run these detectors on any web page. Calibrating it against 46 pages of known-human and known-AI text revealed something the detectors themselves couldn't: the structural rhetoric patterns (epigram, isocolon, anadiplosis, antithesis) measure writing quality, not writing origin. Good human writers trigger them as much or more than AI. What actually discriminates is voice — contractions, first-person pronouns, questions, and the absence of formal transition words. The tool was renamed accordingly. It previously called itself an "LLM slop detector." It contained more slop than science.

the clinical parallel

Sloptics maps to three established frameworks in clinical psychology. The mapping is structural, not metaphorical.

Beck's cognitive distortion taxonomy (1967) named the shapes of bad thinking. Once a patient recognises "catastrophising," they can catch it in the wild. The name creates distance. The distance creates choice. Sloptics does the same thing for LLM output: each named pattern becomes a detection tool. The taxonomy is the apparatus.

ACT's cognitive defusion (Hayes et al., 2012) takes it further. The shift from being inside a thought to being alongside it — noticing it as a mental event rather than experiencing it as reality. In sloptics: the shift from reading LLM output as information to reading it as construction. You stop looking through the text to the meaning and start looking at the text as a produced artifact. Detection begins at defusion.

The Reflective Functioning scale (Fonagy et al., 1998) measures the capacity to understand behaviour in terms of underlying mental states. It runs -1 to 9. Surface slop detection sits at RF 5 — ordinary reflective functioning. Clear slop detection requires RF 7+ — the capacity to model the process that generates output, not just evaluate the output itself. The RF scale is not a personality trait. It is a developed capacity. It can be trained.

This page documents a protogenic concept emerging from THE PIT — 242 session decisions, 1,279 tests, and a living taxonomy of LLM failure modes. The slopodar entries are the empirical base. The frameworks above are the interpretation. Both are works in progress.

the slopodar →

human baseline

The same abstract noun highlighter runs on the text below. These are unedited excerpts from the Captain's log — written by a human, in real time, during live engineering sessions. The purple density difference is the argument. Or it isn't.

23 Feb 2026, 16 hours in, no food

suggesting edit to qa defect list and/or triage; BYOK is beautiful but, frankly, its nice to have, its not a deal breaker. Strong suggestion to actively place on /roadmap as one of the next incoming features; marketing it? I think this can wait, if a good idea comes let it come. But it can wait. If HN give a shit, they will find it by themselves. Let them. Adds to the game. Maybe they're messing with the wrong crew. Time will tell, Mr Weaver.

lets wait on the bug bots to do their bug-thing, meanwhile you can explore the code base and give me a thumbsup/down on whether its safe to proceed ahead of merge schedule as per standard weave protocol (we need a name for that; you are exceptionally good at naming them, havent let me down once, surprises me and makes me laugh everytime, brings me joy, frankly)

23 Feb 2026, the still point

I could not help myself. I read the zeitgeist delta. The parallax agent roster. My mind is clear. After the temporary surge in dopamine et al, I found myself at at the still point. Calm. Like a hunter who knows his prey and has tracked it for many weeks. On that final scan, it is spotted.

Everything else fades into insignificance. We live for moments like this. Die for them.

My bearings are set, with only minor variance. Default mode network. Its a thing.

23 Feb 2026, going light

We go light on the crew, but we reduce the attack vector. Mr Weaver, Architect, Analyst, Keel (perhaps edited: if it is normal human regulation, I see no issue. I see bonus signal). All other crew exist in name and form but have played less than 2% of the process. We can keep them around, in the dark. Costs us nothing. But so far, the fabric has held through those alone.

Nothing would be nicer to have than to have a punchers chance at contributing to something meaningful, as it reaches fever pitch. Please base all your triage decisons on this one governing principle. It should clarify product decisions (especially fix/hide decisions) greatly. Aerodynamics.

24 Feb 2026

This may seem a little frivolous, and in a sense, that is true. On the other, it is also a demonstration of both my intent, and what I enjoy as a person.

Plenty of people can prompt agents. Very few can govern them.

I am in the business of making sure the human stays in the LLM, and I'll go as deep as I need to go to make sure of it.

footnotes

Terms coined during the annotation of this page as a specimen of its own subject matter.

slopporizon /slɒp.ə.ˈraɪ.zən/ n.

The point beyond which slop becomes invisible to the reader. An event horizon for sloptic detection: patterns that cross the slopporizon cannot be recovered by the reader's current level of attention, fatigue, or training. The slopporizon is not fixed — it shifts with the reader's RF, the session length, and the density of preceding clear slop that has already been accepted without challenge. Each unchallenged passage moves the slopporizon closer. From slop + horizon.

anadiplosis /æn.ə.dɪˈpləʊ.sɪs/ n.

Repeating the end of one clause at the start of the next. "The name creates distance. The distance creates choice." Aristotle documented this figure. It once had genuine rhetorical force. In LLM output it appears with such regularity that it is burned. The construction performs profundity: A creates B, B creates C. The chain feels inevitable. But a human who felt this insight would say it messily. From Greek ἀναδίπλωσις (anadiplōsis, "doubling").

analytical clear slop n.

Clear slop in the domain of data and measurement. Numbers that are correct, constructs that are mislabelled, baselines that are undeclared demographics, analysis layers that share blind spots. The immune evasion mechanism: numerical precision. A finding presented with effect sizes and p-values inherits the reader's trust in statistics, regardless of whether the construct, the sample, or the analytical pipeline deserves that trust. Discovered when the calibration tool produced flattering results about the Captain's writing and the Captain asked: "How do I control for slop inside the analysis?" The question surfaced five confounds and three deeper bias layers, none of which had been flagged by the agent that produced them.

sloppesque /slɒp.ˈɛsk/ adj.

Having the quality or cadence of slop without being definitively identifiable as LLM-generated. The uncanny valley of text. A passage is sloppesque when the slopodar fires but the reader cannot yet name why. The felt wrongness precedes the diagnosis. Sloppesque text may be LLM-generated, LLM-edited, or human-written under secondary contamination — the origin is ambiguous, which is precisely the problem. From slop + -esque (suffix denoting resemblance, from French, from Italian -esco).

The slopodar is available as a chrome extension if you want to try it. At its current version I am not sure it is worth the Google developer fee. People who know enough about sample sizes, false positives, and calibration to use it properly probably don't need a half-baked chrome extension for support. People who don't know those things need a tool reliable enough that they don't need to understand why it's broken. It is not that tool yet.