Consiliences AI Operating Invariants — A Multi-Agent System Design Note

The pitch I have read more times than I have bothered to count, over the last year or so, runs as follows. Generative AI gives a builder leverage. Agents extend that leverage. Therefore agents are how a small team — and, in the limit, a single person — competes with a large one. The argument is clean and the slide deck for it is, by now, formulaic. It is also, as a framing, designed to keep one detail at the margin: what actually scales a long-running solo project is not the agents. It is the contract between the agents and the corpus they read from and write to. Get the contract wrong and every additional agent multiplies the rate at which the work degrades. Get it right and the agents become genuinely additive. The framing tells you to count agents. The system tells you to audit verdicts.

The corpus I have been building since the start of the year holds about seventeen hundred adjudicated research signals across sixty-nine clusters of mechanism — climate, biology, finance, geophysics, monetary cycles, the unglamorous middle of the periodic table — each signal at some degree of validation, from a small core that has survived a four-test battery on real data through a long tail still on the bench. Each signal carries a verdict from a fixed taxonomy of about ten tiers, of which CONFIRMED is the strongest and KILLED is the end of the line. The tiers are not interchangeable. The audit gate that enforces that non-interchange is the load-bearing surface of the entire system, and the rest of this piece is about one of the rules it carries.

On the eleventh of May, a bulk script — Scripts/_apply_verdict_overstated_bulk_2026-05-11.py, applied as commit e33cba536 — downgraded ninety-five research signals from the CONFIRMED family to NULL in a single operator-approved pass. None of them were malice. Every single one had been stamped CONFIRMED at some prior point by a developer or an agent reaching for the label because the underlying source paper read like confirmation, and the local validation battery — the thing that was supposed to earn the label — had never actually executed. The corpus had, on that morning, two hundred and forty-four such overstatements waiting to be caught. Ninety-five came off in the bulk. The other one hundred and forty-eight have separate, slower paths to either remediation or correction. The ninety-sixth signal, the one that did not go to NULL, is the load-bearing detail in this story.

The rule that made the wave possible is numbered invariant seventeen and it is two and a half lines of Python in scripts/signal_integrity_audit.py. It refuses, as a high-severity finding, any signal whose verdict is in the canonical CONFIRMED family while the validator’s battery shape shows passes == 0 and skipped >= battery_count - 1. In English: it catches the case where someone has stamped a signal CONFIRMED before any battery has run. The rule did not exist before the eleventh of May. Until the rule existed, the corpus was free to accrete CONFIRMED labels at a rate the audit gate did not measure. The invariant is, in a real sense, the system noticing the failure mode it had been running for the months before. Two hundred and forty-four overstatements is what the failure mode looked like at the moment it was noticed. The bulk script is what removing them looked like once a name and a number existed for the thing to be removed.

The signal that did not go to NULL is witch_trials_little_ice_age, downgraded separately to SUGGESTIVE in commit ba978ff59. The reason it survived the bulk is in its own validation.json file, in the rationale field, where the signal’s author had already written the sentence “All results should be treated as suggestive (N=18, low power, confounders explicitly named)”. The bulk script was about to flatten that sentence under the same rule it applied to the other ninety-five. The operator’s read, exercised through the human-authority loop that sits beneath the rule, was that NULL would be harsher than the author’s own honest self-calibration, and SUGGESTIVE was the right tier. The exception did not soften the rule. It was the rule working as designed: the audit gate proposes, the operator disposes, the verdict history records both. Without the operator-disposes step, the rigid form of the rule would have erased a nuance the original author had taken the trouble to write down. With the operator-disposes step, the rigid form of the rule did the work it was meant to do — surfaced the overstatements — and the human read caught the one that did not fit the bulk shape. The exception is what kept the rule honest.

The artefact that holds all of this together is the verdict history — a JSONL ledger that records, per signal, every change of verdict, the source of authority, the date, and the rationale. Ninety-five rows landed with source_event set to operator_approved_bulk:verdict_overstated_2026-05-11. One row landed for witch_trials_little_ice_age with source_event set to operator_approved_exception:verdict_overstated_2026-05-11_witch_trials. A future maintainer of the corpus reading either row can reconstruct what happened, by whom, on what authority, against what previous state. The ledger is what makes the contract auditable. Without it, the corpus is a series of confident claims that no one can replay six months later. With it, the claims are auditable in the same sense that a financial ledger is auditable — not because the claims are necessarily right, but because the history of what was claimed and on what authority is preserved.

The plain question, the one a reasonable outsider with no stake in the room would ask, is this. If the load-bearing constraint on a multi-agent research system is the verdict-tier contract, and the work that earns the contract is two and a half lines of Python with a name and a number, what are most of the agent product launches I read about in the press actually selling?

Roughly once a week, my feed contains an announcement of a new agent framework or platform with somewhere between a dozen and several hundred agents in its diagram. The diagrams are handsome. The platforms are sometimes real. The rooms inside which the announcements happen are not, in my experience, full of unserious people; they are full of people executing an investor framing that rewards the diagram and pays for the headcount-substitution story. In the week of one such announcement, back in May, ninety-six verdicts got downgraded in this corpus, in a process that took roughly an afternoon of operator attention and produced no diagram at all. The work that earns the contract does not photograph well. The work that photographs well does not earn the contract. The people inside the rooms launching the photogenic version are, many of them, capable people inside a framing that is wasting them. The framing is the problem. The people are the casualty.

Five months on, the rule is still in the audit gate, a few lines in a file nobody reads for pleasure. The corpus has grown by a couple of hundred signals since May, and it now kills about one in six of them: close to three hundred, each with a row in the ledger saying who killed it, on what authority and why. The kill rate is not a failure statistic. It is the contract being honoured at scale, one dull row at a time.

Anyone wanting to build the same kind of system should note that the invariants are not glamorous and never will be. Each one, isolated, sounds like the kind of nitpick a project manager raises in week three of a project to feel useful. Together, the invariants are the assertion — in writing, in commits, in a JSONL ledger — that the standard the work is held to either outlives the person who wrote it or it does not. The Dutch verb is doorgeven: to pass on. The audit gate is the form doorgeven takes when the thing being passed on is a corpus rather than a craft. The reader who finds this kind of work interesting can begin by writing down one rule that did not previously exist, naming a thing that was wrong, and seeing what falls out of the corpus when the rule is run.

Drafted with AI assistance under operator supervision.

All substantive claims, technical descriptions, and editorial choices are operator-authored or operator-approved per paragraph.

The persona rob_de_roo may have produced the first draft; the published artifact represents the operator’s substantive position.

Persona-style scaffolding does not constitute authorship; operator attribution applies only after per-paragraph review.