source of truth for the agent era

last night a fleet of agents built things, and then one agent read back the "done" claims from the first wave and checked each one against the object it was a claim about. twenty claims, re-derived from artifacts: a file's size from stat, a word count from wc, a page count, a commit from git show, an import from find_spec. eighteen came back grounded. zero came back contradicted. two came back unverifiable, and those two are the ones worth keeping.

the flattering read is "the run was honest, everything held." look closer at the two it could not verify. one was a claim that an essay had all five of its beats, which is prose with no probe to run against it. the other was a claim that a tool had run cleanly in a fresh environment hours earlier, which was true when it happened and is now the past, unobservable after the fact. the check did not fake a verdict on either. it said unverifiable and moved on. a system that only ever returns "fine" is not checking anything; the two it refused to score are the evidence that the zero-contradicted means something.

here is the turn. as agents produce more, the scarce thing stops being generation and becomes trust. a fleet can write a hundred confident "done" lines an hour. the bottleneck is not writing them, it is knowing which ones are true, and true does not mean well-phrased. it means the claim survives being held up against the object it describes: the repo, the document, the artifact, the tree. checking a claim against the vibe tells you the author was fluent. checking it against the object tells you the author was right. those are different questions and only the second one scales.

and that is one question, not three, which is the whole point of writing this down. i have been building what looked like separate things. one verifies claims about the world against documents, so an agent answers a science question with a source or an honest refusal instead of a confident citation to nothing. one verifies claims about agents' work against artifacts, so a false "done" gets caught against the repo before it merges. one verifies human authorship, subtracting the machine turns from a transcript to find the roughly one percent a person actually typed. three surfaces, one spine, and underneath all of it the same demand: hold the claim against the object. i think that demand is a category, and it needs a name, so i am naming it. agentic source of truth. not a product. the question every one of these was already asking.

the honest limits belong here, in one place, because an essay about verification has no business overstating its own. this is a thesis with two hackathon-stage products and a spine, not a shipped company. the self-witness above is n equals one: one operator, one night, my own claims checked against my own artifacts, no control run. the zero-contradicted is only worth quoting because the same engine returns contradicted on a dishonest fixture and exits non-zero, so it can say no; a check that cannot say no is noise no matter what it prints. the human-authorship piece is real as a gate but paste-detection itself is still being built this week, not shipped. and the numbers here are one run's numbers, not a law about fleets.

the instrument is the part i am least comfortable handing over, and that discomfort is the method working on me. the point of all this is that a claim should be runnable against its object by someone who is not the author. so: the transcript coach is genuinely that, installable from main, point it at your own history, get your own number, not mine. the other two are one push from public and today they are not. the run's own cold-start check caught exactly this: a stranger cloning the repo gets invalid choice: 'truth' because the command lives on an unpushed branch, and the second repo is still private. i could have written "run all three." the object says i cannot, yet. an essay that argued for checking claims against the object and then made a claim its own object contradicts would be the exact failure it is warning about. so the honest version is: run the one that runs, watch me push the other two, and trust the thesis exactly as far as you can reproduce it.

n=1, one overnight run, own claims vs own artifacts, no control · 18 grounded / 0 contradicted / 2 unverifiable, where zero-contradicted holds only because the engine returns contradicted + exits non-zero on a dishonest fixture · thesis stage: two hackathon products + a spine, not a shipped company · paste-detection still in build · one of three instruments is publicly runnable today; the other two are one push away and the run's own cold-start report says so.

TWEET THREAD

last night a fleet of agents built things, then one agent checked the first wave's "done" claims against the objects they described. 20 claims from artifacts: 18 grounded, 0 contradicted, 2 unverifiable. the 2 it refused to score are what make the 0 mean something. [link]

as agents produce more, the scarce thing isn't generation, it's trust. the bottleneck is knowing which "done" is true, and true doesn't mean well-phrased. it means the claim survives being held against the repo, the document, the tree. the vibe says the author was fluent. the object says they were right.

verifying facts, verifying agent work, verifying human authorship: one question, hold the claim against the object. that's a category and it needs a name: agentic source of truth. honest limit, this is a thesis with two hackathon-stage products, not a shipped company. run the piece that runs today, watch me push the rest: [link]