“Beyond your ken” — outside the range of what one can see, or know, or judge. the working definition of the problem
This is a study programme with a website attached, rather than a journal with contributors. It exists because a specific question turned out to have a rigorous literature that nobody had assembled: how do you know whether work you did not supervise was any good?
The question is usually treated as new, and it is not. It sits at the intersection of six mature fields: psychometrics, the design of experiments, the economics of delegation, the economics of organisation, the study of proxy failure, and scalable oversight. Each is formidable on its own, and they substantially do not cite one another. The programme's wager is that the intersection is empty for historical rather than intellectual reasons, and that assembling it is worth a year.
The organising claim, which everything here is arranged to test:
Your abstraction ceiling is your measurement ceiling. You cannot delegate at a higher altitude than your ability to tell whether what came back was any good, so an agent that seems unable to climb is very often an evaluator that cannot follow.
Three documents follow. The curriculum is the foundation: seven units, real primary sources, one mathematical model to be derivable in each, and one foundational experiment to be reproduced from scratch. The house standard governs any run published here, and was written from three prior runs in this repository that each failed the same way. The protocol is the experiment the programme is ultimately for, written as a registered report and left with six holes in it, because each hole is a quantity one of the units produces and publishing the holes beats filling them with guesses.
The programme as a dependency graph. Read it bottom-up: the three prior runs at the base, what was learned from them above that, then the units, each producing an output the protocol needs, and at the top the run itself. Every node is a link.
The spine, stated late because it took sixteen revisions to find. This began as a question about supervision and became a question about structure. A plan is a graph, and an agent's state is its precedents and nothing else, so position in that graph is measurable before any data exists: nine roles fall out of degree, the automorphism group says which turns may be pooled, and the ratio of what an agent holds to what it answers for is one pass over the edges. That last quantity is what this is named after and it arrived in revision fourteen, not revision one.
Unit IV is where the work went. It was marked blocked while the paper depending on it was marked active, and that survived because every check on this figure concerned geometry and links rather than the claim the colours make. There is a check now. The register carries what is settled and what is not.
Three measurement exercises have been run in this repository. All three produced working machinery and none produced a usable comparison.
In August 2026 a budget-governed agent loop issued ninety-nine work orders over about fifty-three hours, of which eighty-nine recorded a completed turn. Its stated deliverable was a curve of quality against turn count, and the number of quality scores it produced was zero. Separately, two bake-off runs gave the same brief to six harness-and-model pairs and scored every result the same way. Across those two runs, twenty-three of twenty-three entries passed every automated check.
None of this is a malfunction. Every component behaved as written. The loop specified three
quality signals; one, a liveness probe, was implemented and fired, and the judge that was to
combine them was correctly configured to withhold a score while uncalibrated, so it withheld
ninety-nine times. The bake-off gate was designed as a floor rather than a grade, and it held
as a floor should. In both bake-offs the judge panel that was supposed to break the tie
returned null.
The eighty-nine is worse than it looks. Twelve of those runs are flagged infrastructure failures, with a median duration of 61 seconds against 597 for real runs, and nine of the twelve are recorded as having passed the gate. The flag was added part-way through, so twelve is a floor rather than a count: five more of the forty-five unflagged records fall in the duration range where infra failures live. Somewhere between 13% and 19% of the run was the harness failing, scored as though it were work. See L10 and L11.
The signals that did fire are worth looking at, because they are textbook objects. The loop's probe took three distinct values across eighty-nine turns and sat at its maximum in fifty-nine of them, so two thirds of all turns are tied at the top. The bake-off reports rank entries by patch size in bytes, which was never claimed to measure quality and does not. A century-old literature explains why a measure's reliability collapses as it approaches its ceiling, and gives the coefficient that would have caught all three cases before any budget was spent.
There is a smaller failure in the first bake-off that is a different kind of instructive. Two of eleven entries recorded a zero-byte patch while their directories held a fifty-kilobyte page and a working module. The agents did the work and the patch capture returned nothing. Because the report sorts ascending by patch size, those two rows sorted to the top, and nothing in the pipeline noticed.
The temptation is to read these records for a trend anyway. The loop shows a faint downward drift in its probe and a near-doubling of gate failures between halves. Neither will bear weight: the drift is not significant at that sample size, the probe is not a quality measure, and the gate failures are confounded with task difficulty, which nothing recorded controls for. Reporting them as evidence of decline would be the error this programme exists to stop making.
What the three runs do establish is narrower and more useful. The apparatus runs. The instrument does not exist. The order of work implied by those two facts is the reverse of the intuitive one, so this volume puts a curriculum before a standard, and a standard before a single further turn.
On the bibliography. Every work cited on this site is real, and every citation is a key into a shared bibliography module that an offline test checks on every build: a dangling reference fails the build rather than rendering a broken bracket at a reader. Nothing here is cited that is not intended to be read.
This is worth a machine check rather than good intentions, because the site next door at wormhole.mino.mobi generates entire fictional literatures — plausible papers, plausible citations, plausible fields, all of it invented. The two surfaces are a matched pair. One is a demonstration of how easily an apparatus of scholarship can be counterfeited; the other had better be able to prove it is not counterfeit.
On status. Both documents are living. The curriculum will be revised as units are completed, and the protocol's blanks will be closed one at a time. Neither is a finished claim, and the dates on the revisions are the honest record of pace.