I tested a narrow context-management question: after an incorrect claim has propagated through several conversation turns, is deleting the original error enough?

The pilot used synthetic, objectively scored tasks represented as context DAGs. Across four model endpoints, pollution derailed 72 cases that had been answered correctly with clean context.

Repair recovered:

• 68/72 after deleting only the erroneous source;
• 71/72 after deleting the source and recomputing descendants in dependency order;
• 72/72 after removing the contaminated subgraph.

The interesting failures occurred when stale downstream turns still repeated or depended on the deleted claim. The source was gone, but its consequences were still present in the next request.

I do not interpret this as a general model ranking. Each endpoint was sampled once at temperature zero, the tasks are synthetic, and provider-default reasoning behavior was not normalized.

What I would most like to test next is whether the same pattern holds for local models and different quantization levels. Would you expect smaller or more heavily quantized models to be more sensitive to residual context?

Full report: https://chenxiachan.github.io/thoughtdag/research/context-repair-pilot-v1/

Disclosure: I designed this benchmark and maintain ThoughtDAG, the open-source reference interface used to visualize and edit the context graphs.

Disclosure: I maintain ThoughtDAG.

While using local models for research, I kept running into a simple problem: once a paper, hypothesis, or mistaken branch enters a chat, it tends to remain in later context even after I have mentally moved on.

I built ThoughtDAG to test a more explicit interaction model.

Each question, answer, and source is a node. The wires determine exactly which upstream nodes are serialized into the model's next request.

Delete a wire, regenerate the same prompt, and that branch remains visible on the canvas but disappears from the actual model input.

It currently supports Ollama and OpenAI-compatible endpoints. Canvases, documents, and API keys are stored locally.

I am less interested in general promotion than in whether this interaction is actually useful for people running local models:

• Would manual context pruning be worth the effort with smaller context windows?
• Would you prefer automatic suggestions followed by human confirmation?
• What would you need to inspect before trusting the selected context?

I am also turning these questions into a small context-control benchmark, so failure cases are especially useful.

Website: https://chenxiachan.github.io/thoughtdag/