The Taxonomy Has to Survive Its Own Audit
the method
previously → four axes, pulled from a pile of my own repos one forced discovery at a time: input, frame, decision, consequence.
the reason this shape keeps reappearing across tools that have nothing to do with one another, and that don’t even lean on the model to the same degree, is that it was never about the domain. it’s a stance on where machine judgment belongs, across all four axes at once. choose the model’s inputs yourself, or at least verify the ones it chose were faithful, so a clean decision never runs on evidence the model quietly palmed (input). write the opinion down first and keep it yours, fixed and version-controlled, not regenerated per run (frame). put the decision in a deterministic layer you can reproduce and defend, and let the model only ever feed a checkable step (decision). and the closer an output sits to an irreversible action instead of a line a human will read, the more you have to have earned the other three, because the person who was quietly auditing your displayed outputs is not standing at the exit anymore (consequence). the domain changes every time. the shape doesn’t, because the shape is the position, and the position travels.
the tool i’m still building handed me the ending, because it holds both extremes in one unfinished codebase. its detection half is the thesis already realized → an entire subsystem where the model has zero authority, every output a pure function of rules i wrote. its trading half is the thesis still being fought for → the highest-stakes path i’ve ever built, where right now the model holds the pen and the audit has already flagged exactly which numbers i need to take back from it before it ships. same repo, both extremes. and the discipline is easiest to hold in the half where the determinism was cheap to write (detection) and hardest in the half where it’s expensive (constructing a trade), which is the whole thing in one project: the pattern was never hard to state, it’s hard to hold precisely where it’s most tempting to drop, which is exactly where the stakes run highest and the determinism is hardest to write. the reason i can tell you that without it being a confession is that i caught it with the code still open in front of me. the pattern isn’t only an autopsy you run on shipped systems. it’s a constraint you hold while you build, and the earlier you run the audit, the more it’s a design tool and the less it’s an obituary.
the taxonomy has to survive its own audit
there’s one thing i can’t leave unsaid, because the argument’s sharpest question aims straight back at the argument. to interrogate my habit of imposing an opinionated taxonomy, what did i do? i imposed an opinionated taxonomy, bounded → checked → laundered → naked and then a set of axes, wrote a scoring rubric for it (the audit prompt, with its cite-the-line-or-it-didn’t-happen rule), and pointed a model at my own repos as the auditor. the investigation was conducted in the exact grammar of the thing under investigation. i could not examine the reflex without performing it. so the frame in these posts is itself an opinion-first, model-assisted construction, and by my own frame-authority argument it does not get a pass just because it sounds rigorous. a well-organized taxonomy is exactly what a laundered one looks like from the outside.
so the frame has to eat its own dogfood. did i discover these axes, or did i improvise a frame that sprouts a fresh limb every time a new report lands? the tell that separates the two is whether the structure grows on new instances or only on new kinds. and the honest record is that four scoring pipelines in a row produced zero new axes, just more points on the ones i already had. the taxonomy declined to grow, which an accommodating container never does. then a content generator, a different kind, forced frame-authority. then a trading engine, a different kind, forced consequence-authority. then a retrieval tool, different again, forced input-authority, and only after i’d refused to grant it on symmetry alone. axes that appear only when the tool is structurally novel are tracking joints. axes that appear for every example are tracking my need to keep sorting. this structure made me wait through four repos before it moved the first time, which is the closest thing to evidence i have. and it has an outside: a program where the model holds no authority on any axis is just an ordinary program with no model in it, and a taxonomy that stops applying somewhere is carving a thing rather than swallowing everything.
i should also be plain about who authored the frame, because it’s the same confession the argument keeps pulling out of everything else. i did not derive these axes alone in a clean room. i built them the way i build the tools → opinion imposed first (the taxonomic instinct is mine, it’s the through-line of everything i make), then pressure-tested against a model used as an auditor and against the evidence of my own repos, keeping only the structure that pushed back. and when the auditor’s frame-generation improvised against reality, when it wanted to read my unfinished trading engine as a shipped confession, i overrode it. that override is the whole method running in the healthy direction: the model proposed a frame, the evidence and i disposed. delegation, but governed. author disposes. it is the good pattern performed on the argument itself, which is either the most honest thing here or the most suspicious, and i’ve decided that not being able to tell which is the correct state to be in.
and the fourth axis is the cleanest evidence i have that this frame carves joints rather than swallowing everything, precisely because of how it got seated. i noticed input-authority early and could feel it wanting in on the strength of its symmetry, which is the exact tell of accommodation, so i refused it and wrote down the test that would earn it instead → a retrieval tool where the real failure is something the other three genuinely can’t describe. then a retrieval tool ran that test and forced it, a failure the decision axis scores as clean, and i seated it without flinching. the discipline was never “refuse the fourth axis.” it was “refuse it until a tool makes refusing dishonest, then seat it the moment one does,” because clinging to a tidy count of three after the evidence lands is the same accommodation failure wearing the opposite coat. the frame earned that dimension by resisting me first and then overruling me with a tool, and that back-and-forth, me proposing and the evidence disposing, is the only procedure i know that yields a taxonomy i’m willing to trust. it is, exactly, the procedure the tools themselves are supposed to run on the model.
which retires most of my suspicion about the pipeline, though not all of it. input → frame → decision → consequence, four stations, authority able to leak at each, and now each one seated by a tool that forced it rather than by symmetry. so the shape is probably real. but “probably real” and “complete” are different claims, and the second is the one i still won’t make. four axes that each earned their seat is a strong map. “and there are exactly four, forever” is the tidiness talking again, and the whole method is a standing refusal to let tidiness talk me into anything a tool hasn’t shown me. the map is exactly as big as the tools have forced it to be, and not one axis bigger. if a structurally new kind of tool turns up tomorrow with a failure these four score as clean, the map grows again, and that will be a good day, not a broken frame.
and in a year where the loudest available move is to shove a model into the decision seat of everything and call it autonomous, “the model doesn’t get to decide” turns out to be a weirdly contrarian thing to build. i think it’s also the thing still standing in eighteen months, when everyone else is staring at a system that decided something strange, or executed it, and can’t reconstruct why. the frame held up under its own audit, which is the only credential i actually trust, and it earned four axes, each one made to prove itself before it got a seat, and not a fifth, because nothing has forced a fifth yet.