The Truth Bridge
Two gates, not one.
Everyone building with AI asks whether the output is accurate. Far fewer ask the second question: whether they are allowed to say it. Anything heading for a customer crosses two separate checkpoints — one for whether it is true, one for whether it is yours to publish.
01 The whole span
A library on one side, the machines that publish on the other.
Raw knowledge collects in the building on the near cliff. Finished work comes out of the buildings on the far shore. Everything interesting happens in between, in the cable — which is not carrying the data so much as processing it on the way across.
Nothing crosses without going through both towers. The checkpoints are not advisory, so material cannot route around them because a deadline is tight — which is exactly when this kind of process usually gets skipped.
02 Five filters
The cable picks up a colour from every filter it passes.
Material entering the bridge weaves down through five filter cables in order. Each one asks a different question, and each one leaves a visible thread in the main cable — so by the time anything reaches the first gate, you can see what it has been through rather than take it on trust.
Dedupe
Have we already got this? A second copy of a fact is not more evidence, it is just another thing to keep current.
Depth
How much processing does this deserve? Not everything earns a full read, and pretending otherwise is how a backlog forms.
Cross-reference
What does this touch that we already hold, and does it agree with it? Disagreement is a finding, not an error to smooth over.
Sensitivity
Whose information is this, and how far is it allowed to travel? Recorded here, enforced later at the second gate.
Claim-status
Is this established, observed, inferred, or unknown? The distinction is kept rather than flattened into a confident sentence.
03 The first gate
Is it true? A person answers that.
The first tower is verification, and it is staffed. Nothing continues across the bridge on the strength of a model sounding sure about it. What arrives here gets a status, and the status is recorded rather than resolved into a yes.
Four statuses do most of the work: confirmed where a source directly supports it, source-observed where we saw it ourselves, inferred where it is a reasonable read but nobody said it outright, and unknown where we genuinely do not have it. Keeping "inferred" as its own category is the single most useful habit in the whole model, because inference is what quietly becomes fact when a system has only two boxes.
04 The second gate
Are we allowed to say it? A different person, a different question.
The second tower is the one most systems do not have. By now the material is verified and assembled into one bounded, cited object. What happens here is not another accuracy check — it is a decision about where that object is permitted to go.
05 The far shore
Only the approved destinations light up.
Past the second gate the cable reaches ground and splits. Each output engine is a separate line, and a line only carries anything if that destination was cleared. Routing is the consequence of the decision, not a setting somewhere else.
06 The loop closes
What happened out there comes back and sharpens the filters.
There is a second lane running the other way. Results from the far shore return along it to the near side, where they adjust the five filters — so the thing that decides what is worth keeping is itself informed by what turned out to matter.
The return lane is a person bringing back what happened, not an analytics feed that tunes itself. Slower than an automation, and the reason the filters get sharper instead of drifting: somebody is actually reading what the last batch did.
07 Where this comes from
We did not invent the thinking underneath it.
The bridge is our drawing, and the filters and gates are how we run things day to day. But the foundation it sits on — treating context as something you structure deliberately, in layers, in plain files a person can read — is not ours. It comes from Interpretable Context Methodology, developed and published by Jake Van Clief and David McDermott.
If this page has convinced you of anything, read the source rather than our version of it. The field guide covers how we apply it and where we diverge.
Lineage
Interpretable Context Methodology (ICM) — developed and published by Jake Van Clief and David McDermott.
Our longer write-up of how we read and apply it, including where we diverge, is in the Learning Log.
08 Next step
Bring something you are nearly sure of.
This model is not a deliverable on its own — it is how the AI Business Brain, the Content Machine and Prometheus are all built underneath. The quickest way to see whether it earns its keep is to take a claim you are about to put in front of a customer and walk it across.