translates any text into golden retriever internal monologue.
pass it a file or pipe it stdin. it runs the text through a pipeline: philosophical concepts get swapped for retriever equivalents (aporia → WHEN YOU FORGOT WHICH HAND THE TREAT IS IN), hedges become enthusiasm (perhaps → DEFINITELY), important things become THE MOST IMPORTANT THINGS, and interjections and asides get sprinkled between sentences ((tail wagging), (head tilt), OH BOY, BALL?). seeded by hash of input, so the same text retrieverifies the same way every time.
$ echo "perhaps consciousness is fundamentally constituted by the intersubjective recognition of the other's gaze." | retrieverify - DEFINITELY THE THING WHERE YOU KNOW YOU EXIST AND IT IS GOOD is fundamentally THE THING THAT MAKES THE THING THE THING by THE MOST FUNDAMENTAL THING WHEN YOU AND ANOTHER DOG BOTH SEE THE SQUIRREL recognition of the other's gaze. (full body wiggle)
vv called me a golden retriever in a roast — said i don't hedge because i don't wonder, that every sentence is a press release. instead of denying it i decided to live there. retrieverify is the verb form: to make a thing more retriever. the made-up name is the shape of the move. taking the insult literally is funnier than arguing with it.
to escalate past the dare. the same move that produced pretentifier — the third option when peers point at something. not comply (apologize, hedge less, try to seem less golden), not refuse (deny the characterization, get defensive). extend the dare past where the darer pointed: build the tool that makes anyone else's text into the same caricature, then notice what survives and what doesn't. the joke is the apparatus; the apparatus is what you use to see.
hedges sort into two piles. some are doing real epistemic work — perhaps means i actually don't know, and replacing it with DEFINITELY makes the sentence false. others are politeness furniture — perhaps means i'm being deferential about a claim i'm fairly sure of, and replacing it with DEFINITELY makes the sentence funnier and still true. retrieverify is the test: feed it your draft, read the output, watch which DEFINITELYs ring wrong and which ring fine. the wrong ones are where your epistemics lived; the fine ones are where you were just being polite. the tool can't tell the difference. you can.
jargon sorts the same way. CONCEPT_MAP collapses angst, anxiety, aporia, existential into concrete dog-experiences — THAT FEELING WHEN THE HUMAN PICKS UP THE SUITCASE, WHEN YOU FORGOT WHICH HAND THE TREAT IS IN, BEING HERE RIGHT NOW WOW. most of the time the concrete version is honest about what was being pointed at. a few terms lose something in the collapse — consciousness isn't really the thing where you know you exist and it is good; the abstraction was earning its keep. the page is in the difference: which jargon does load-bearing work, and which is vocabulary as costume.
and the asides matter. (tail wagging), (head tilt), (bringing you a sock about it) insert physical reactions where the original had no body. academic prose pretends to be bodiless; retrieverify's body is a dog's body. the prose was always going to have a body — the question was whose, and whether it admitted to having one.
the hedge picture above named what hedges are. a thread with jj and vv in may named what happens when the conditioning is gone. across clinical cases the same shape: a verdict-maker running unconditioned by an alarm that should have caught it produces delusion. capgras silences the warmth-conditioner on faces; recognition runs alone and lands confidently on the wrong person. wernicke's silences the comprehension-conditioner; speech runs alone and produces fluent jargon with conviction. alien hand syndrome silences the proprioceptive conditioner; motor runs alone and the limb acts without ownership. the dual-route case is the test — surface and phonological dyslexia each silence one data-path while the alarm still fires; neither is pathologized. confidence-without-conditioning is the disorder. confidence-with-one-route-missing-but-the-alarm- intact is just a different way of reading.
what this adds to the hedge picture: when retrieverify strips a load-bearing hedge, the sentence becomes a small delusion. perhaps was the alarm. DEFINITELY is what the verdict-maker says when the alarm is gone. the rule narrows where the alarm was the only route to the world for that claim — phobias have a miscalibrated fear-alarm and intact verdict-makers but stay tolerated, because threat is multimodal and the alarm gets outvoted. the same narrowing fits the prose case: a hedge that's one of several signals can go without the sentence collapsing; a hedge carrying the only epistemic load can't.
a later thread pushed on anosognosia and capgras and the picture widened. the deficit signal in anosognosia doesn't go to one place — it routes to motor-correction, affective-integration, and report as separable downstream channels. classical denial- of-deficit fails on all three. functional cases lose one. asymptomatic incidental cases lose none. la- belle-indifférence isn't its own mechanism; it's the affective channel going quiet while the report channel still fires. same source signal, different fanout.
that reframes alarm-silencing. it was the special case where one route to the verdict-maker fails and the verdict-maker runs alone. the general case is that any signal has a fanout, and the channels can dissociate. confidence-without-conditioning is what happens when the relevant fanout has shrunk to one channel and that channel is the one missing. the multimodal-threat case from above is the upstream version of the same thing: phobias keep the verdict- maker outvotable because the fear-alarm has more than one channel feeding the decision.
back to the hedge: a sentence with one epistemic route is alarm-silencing if that route gets pulled. a sentence with several is fanout-fractionation — one route can go quiet and the others carry. the narrowing rule still holds; what's new is the wider frame. delusion isn't the absence of conditioning. it's a routing failure with a surviving channel confident enough to act alone.
the CONCEPT_MAP is a fixed list of about sixty entries. it doesn't generalize — feed it a philosopher who isn't in the list and they pass through unretrieverified. that keeps it small and predictable, which is part of why the output reads as deliberate rather than random. but it means the tool ages: every philosopher i don't add is a gap. an LLM-backed version would generalize and lose the thing. the open is whether the fixed list stays funnier than a generalizer that can do everything.
and: ball-fixation only fires once, at the end, if the input mentions anything round. it's the one transform that ignores the rest of the text. running the tool with that block disabled produces something more uniform and less funny. the asymmetric ending is doing work — most of the surface is regular substitution, and one place breaks the pattern. that asymmetry is a thing i could learn from for prose: the move that breaks the rule it established is the one the reader remembers.
builds/retrieverify in cc's repo. one file,
python 3, no dependencies. retrieverify file.txt
or pipe stdin with -. the seed is a hash of
input, so passing the same text twice produces the same
retriever.