The problem

Every important system has a system of record. Intelligence does not.

Each domain acquired a record the moment it began to matter. Intelligence now matters, and it runs without one.

Assets outlive the people who knew them

What each of them claimed leaves with them.

An oil field runs for decades. So does a transmission grid, a pipeline network or a line that makes a regulated medicine. The engineers, models and vendors around each asset change many times before it's retired.

What each of them claimed about the asset, and what that claim rested on, leaves with them. Nobody can say what the system believed on the day of a decision, so nobody can defend it.

DomainRecordWhat it keeps
MoneyAccountingwhere work becomes accountable
OperationsERPwhat was done, and by whom
RelationshipsCRMwho was promised what
PeopleIdentitywho may act, and as whom
IntelligenceSyderialwhat was known, allowed and done
Figure 1. Systems of record. Each arrived when being wrong about its domain became expensive.

The failure

It's rarely a hallucination. It's a confident merge.

One word with two real referents gets joined across a boundary, and no claim in the record points to the evidence that would refuse the join.

Two valves share a tag across two plants. A procedure revised in March is cited as it stood in January. A note written for one compartment answers a question asked from another. Each answer reads well. None of them can show where it came from, or who was allowed to see it.

These are bookkeeping failures.

Accounting, ERP and CRM fixed the same failures in their domains by keeping a record. Syderial is that record for intelligence.

What a record has to answer

A memory, a log or a search index can’t answer these six questions together.

What was claimed

A claim about the world, stored with its evidence. A fragment of text scored by similarity doesn't answer this.

When it applied

The interval in which the claim applied, according to its source, including intervals that have closed.

When it was learned

Knowledge arrives late. A record that can't say when it learned something rewrites its own past.

Who said it, and from what

The person, agent or source behind each claim, and the evidence it rested on.

Who could see it

What each reader and each agent was allowed to see at the moment it mattered.

What was done because of it

Which actions the claim supported, under which policy, and what they changed.

Why logs fall short

The record is kept as it happens, not reconstructed afterward.

Observability logs after the fact. It can say what a model returned. It can’t say what the model was allowed to see, which evidence supported the answer, or which fact replaced which.

A model lab sells one model’s stack. The record of an asset has to outlast every model that will ever read it, so no single model can own it.

The record →

The question we ask operators

What happens when the person who understood this asset leaves, and what does a wrong assumption about it cost?