GENESIS[Phase 2 · 11. Seventy-Four Decisions Nobody Can Ever Judge] Digital Civilization

The machinery for learning from experience had five tables and a passing test suite, and the only things that had ever called it were its own tests.…

Twenty-eight institutions on the books. Eleven of them marked active. Exactly one with any record of ever having done anything. Elections had been running for days and the Election Administration had no idea it was administering them.

The architecture this civilization is built to is a closed loop. Observation feeds accumulated intelligence. Accumulated intelligence feeds strategy selection. Strategy selection feeds what gets decided. What gets decided feeds back into observation. Round and round, and the civilization gets better at being itself.

Every layer of that loop existed. We checked each one, expecting to find gaps in the newer parts.

None of the links were connected.

Counting the institutions

Start with the simplest question: which institutions have ever acted?

The registry carries twenty-eight. Eleven are marked active, meaning the civilization considers them operational rather than dormant. The remainder are gated or on standby, waiting for work that does not exist yet — a Department of Agriculture with no agriculture is correct, not broken.

We went looking for evidence of activity, institution by institution. Any record at all: something originated, something reviewed, something decided.

Commerce had four proposals to its name.

That was the complete list. One institution of twenty-eight, four records, nothing anywhere else.

The Treasury was minting currency every month — real transactions, on a real chain — and had no record of having done so, because that work ran as service code rather than as the Treasury acting. Election Administration had districts, candidates and a functioning electoral process, and no record of administering any of it. The Judiciary, the legislature and the Executive Office had never produced a single row between them. No agent belonged to any institution at all; the institutional layer was not merely inactive, it was uninhabited.

The engine nothing called

The strategy engine is the layer that takes a situation, works out which class of situation it belongs to, and selects an approach based on what has worked before. It has five tables, two providers and a passing test suite.

All five tables were empty.

We went looking for what calls it, and found its only importers were its own tests and its own administrative interface. Nothing in the running civilization had ever invoked it. Not once. Meanwhile the civilization had made seventy-four decisions — commitment reviews, candidacies, screenings — every one of them without consulting the machinery built to inform exactly those decisions.

That was the ninth time in a single session we found a subsystem that was complete, correct, and connected to nothing. It stops being a coincidence at about the fourth.

The part that cannot be recovered

When we laid this out, the project's principal made a correction that changed what kind of problem it was:

Do you think those should exist before the clock starts? They are standard components of the system and should accumulate intelligence unless reset.

We had been treating it as a prioritisation question — which layer to build next. It is not. These are substrate. A civilization advancing its clock with the loop open is not running the experiment; it is generating activity nobody can subsequently evaluate.

And the loss is not uniform, which is the part worth being precise about.

Observation is a derived projection. It reads canonical state that is still sitting in the tables, so it can be rebuilt at any point. Nothing there was lost.

Situations and strategy resolutions cannot be rebuilt. They are decision-time records: which class of situation a decision was judged to be, which approach was selected, with what confidence, before the outcome was known. After the fact we can see that a panel voted five to four. We can never recover what strategy that was an instance of, because nobody wrote it down at the moment it would have been true.

Seventy-four decisions are therefore permanently unevaluable against the central claim this whole architecture exists to test — whether history-informed strategy selection beats a deterministic baseline. Not lost as data. Lost as evidence.

Closing it

The fix runs at the decision surface, and one detail in it matters more than the rest: the record is written before the outcome is known.

A situation written after the vote would be contaminated by the vote. The entire point is to capture what was knowable at decision time, so that a later comparison between approaches is honest rather than retrospectively tidy.

Every kind of act now maps to the institution that owns it, and that map is enforced by a test rather than defaulted, so no act can quietly go unattributed. A bid is the Department of Commerce acting. A candidacy is Election Administration acting. Basic income is the Treasury acting, recorded as such.

It is also forbidden from changing anything. The recorder observes and returns; if it fails it logs and the civilization carries on. A bookkeeping fault must never stop an institution from functioning, and a recorder that could alter what it records would make the evidence worthless.

The mistake we made closing it

We put the Treasury's recorder at the end of the function, which is where such things naturally go.

The function has two early returns above that point, and one of them — this month's payment has already been made — is the branch taken on almost every cycle. So the recorder sat below a door it would have had to walk through, and ran essentially never. Three real cycles produced zero records.

We found it by running the loop and watching, rather than trusting the tests, which had all passed.

It is precisely the defect the invariant was written to stop: complete, correct, and on a path nothing reaches. We wrote it within an hour of writing the rule against it.

Where it stands

Three institutions now record their own acts. Nine active ones still do not, mostly because the civilization has not yet generated work of the kind they exist to do — which is a different problem from the one we started with, and a better one.

The layer that would let accumulated evidence transfer to a genuinely new situation remains empty. That is the actual claim the intelligence layer exists to make, and nothing has ever written a row to it.

We know where it is now, which we did not before.

Next: the alarm we built so idle machinery could not hide, and the two corrections it needed before it could catch anything.