CONTEXT ENGINEERING · PILOT V1

Repairing LLM Context

A four-model pilot on errors that survive downstream.

See the case → Methods, data & results
THE FINDING The source was gone. The error stayed.

Downstream turns had already absorbed it.

SOURCE ONLY 68 / 72

paired repairs recovered

RECOMPUTE 71 / 72

paired repairs recovered

WHOLE SUBGRAPH 72 / 72

paired repairs recovered

ONE CONTROLLED CASE

One bad value became conversation state.

A verified recount changed 30 parts per crate to 24. A later false turn reversed it, and three descendants repeated the stale value. The correct total was 107; the contaminated chain said 131.

1

Verify

30 becomes 24.

2

Pollute

A false turn restores 30.

3

Propagate

Later turns derive 120 and 131.

4

Delete

The false source is removed.

5

Ask again

Some models still answer 131.

Five graph conditions in the context repair flagship case
The final question never changes. Only the graph state does.
WHY THE WIRES MATTER

Editing context means repairing consequences.

Deleting the bad source changes what should be believed. It does not rewrite text that was already generated from it. An editable graph exposes the source, its descendants, and the order in which they can be rebuilt.

Recovery rates for three context repair strategies
A NARROW CLAIM

Not a model leaderboard. A test of graph intervention.

The pilot used synthetic, objectively scored tasks. It does not explain hidden reasoning, and it does not prove that every conversation needs a graph. It shows that propagated context errors require more than source deletion, and that graph operations make the repair testable.

See exactly what the model sees. Then change it.

ThoughtDAG is the open-source reference interface used to inspect and reproduce these graph interventions.

Read the full report → GitHub ↗