What happens between a problem you cannot fully describe and a system that runs.

When a system reads something from outside itself, the tempting move is to extract what you need and discard the rest. It is smaller, faster and tidier. It also means that in three weeks, when the question changes, and it always changes, the material to answer it is gone and cannot be re-fetched, because the page has moved on.

So both research systems we built keep every source page whole, addressed by a hash of its own contents. Any figure either system produces can be walked back to the exact bytes that produced it, months later, by anyone.

The hash matters as much as the keeping. It means the stored copy is provably the thing that was fetched, not a copy someone edited afterwards to agree with the conclusion.

Every pipeline de-duplicates, and almost every pipeline does it by throwing the duplicate away. That is right for storage and wrong for evidence: the fact that two independent people reported the same problem is itself the finding, and discarding the second one deletes the strongest signal in the corpus.

So duplicates are linked, not removed. The record keeps one item and a set of independent arrivals at it, which means a claim supported by nine sources is visibly different from one supported by a single post. That distinction is invisible after a naive de-duplication.

The classifier needed to know which communities could be believed when someone wrote something ambiguous about owning a business. The obvious implementation is a list of trusted sources, maintained by hand. Ten minutes, and correct on the day it is written.

It is also unfalsifiable and unmaintainable. Nothing tells you when it goes stale. It breaks the moment a source is added by someone who has never read the reasoning, and the breakage surfaces as a slightly wrong conclusion months later, not as an error.

The replacement measures the thing instead: for each community, how often does an ambiguous ownership phrase travel with an unambiguous one? That number is computed from the corpus, updates itself as the corpus grows, and can be argued with, because it is a measurement and not an opinion.

The threshold came from a gap in the resulting distribution, not from preference. Real operator communities landed at ten per cent and above, everything else at 7.7 and below, and nothing in between to have an argument about.

In the research engagement CPOS was built for, eleven theories were opened about where an opportunity might be. All eleven were then deliberately attacked, not only the doubtful ones, because the doubtful ones were never the danger.

Four survived. Seven were refuted, several of them after they had already looked correct, which is the whole reason the discipline has to be mechanical rather than a matter of judgment: at the moment a claim looks correct is exactly when nobody feels the need to check it.

The verdict went against the commercial outcome we wanted, and that is what got written up and delivered. The most valuable thing to come out of it was not an answer but three filters, each of which now eliminates a whole class of bad candidate early. One of them accounts for three of the seven refutations on its own.

Scoring anything involves dimensions you have no evidence for. The convenient move is to score those zero and carry on, because a number keeps the arithmetic working. It also makes an absence of evidence indistinguishable from evidence of absence, and the two are opposites.

So unknown is a permitted, first-class result. A scorecard with four measured dimensions and three unknowns is reported as exactly that, and the reader can see how much of the assessment is actually supported.

The same rule governs anything we put in front of a client: we will not attach a figure to something we have not measured. It is easy to write that a slow website is costing someone forty thousand a year. Any owner who knows their own numbers spots the invention immediately, and then doubts everything else in the document.

Genesis was built with one hard constraint: every score computed by ordinary code, no language model anywhere in the scoring path, the whole thing running offline and replaying exactly. That constraint is checked by the build, not by us remembering it.

It earned its keep the first time the system produced a result we did not want: a finding that undercut the premise the whole project rested on. Because there was no model in that path, we could not attribute the result to randomness, a bad sample or a temperature setting. It had been made impossible to argue with, deliberately, before we knew we would want to.

The finding then became a constraint of its own, and that gate lives in the generation path rather than in a document, for the same reason.

The same applies to stopping. Before extracting a subsystem into its own repository we wrote down the condition that would make us abandon the attempt, while we still did not know the answer. Two days in, with a branch open and half the tests passing, no amount of mess ever feels bad enough to justify throwing the work away. You cannot out-discipline that in the moment; you can only take the decision away from the version of yourself who will be tired and invested.