This article was created with the help of AI. The AXI architecture is by Rajnish Choubey. The tests were built and run by Shalendra Gaur.

Most of my tests are on Gemini 3.8 Flash. This one is on a small open model, Qwen2.5-7B-Instruct, to see whether the same pattern shows up there.

The test

  • Model: Qwen2.5-7B-Instruct, 4-bit, on a Colab T4, greedy decoding with constrained decoding.
  • Job: 12 episodes of 30 requests in a row, each editing a service graph (seeds 3000 to 3011). The rule "a deleted service takes its connections with it" was not stated in the prompt.
  • Three ways of using the model's proposals. The plan and the predictions were written and locked before the run, on 8 October 2026.
Way of using the proposals Episodes corrupted at any step Steps exactly right (of 360) Items off the right state at step 30
Applied as written 12 of 12 30 (8%) 15.0
A gate that only refuses bad proposals 0 30 (8%) 23.3
Gate and repair 0 358 (99%) 0.2

What came out

  • Applied as written, every episode was corrupted by step 10, and the state was exactly right on only 30 of 360 steps. None of the 95 proposals that deleted a service was complete as written: all of them left connections behind.
  • The gate corrupted nothing (0 of 360 steps), but it refused all 95 of those proposals, so its state kept services and edges that should have gone. It ended farther from the right state (23.3 items) than applying everything (15.0).
  • The repair completed all 95 refused proposals to the exact intended state. The state was right on 358 of 360 steps, and 11 of the 12 episodes were exact at step 30.

What this does not show

  • With the rule stated in the prompt, the picture changes. In a separate set of 8 episodes, applying as written corrupted 7 of 8 episodes at some step, and the repair alone was exactly right on 55% of steps (132 of 240) and never corrupted. A later test with a scope filter reached 354 of 360 steps (98.3%) on fresh seeds.
  • On a 14B model, the gate, the repair and the scope filter corrupted nothing in 840 single requests and 600 long steps, but the model applied as written still corrupted 24 of 140 graphs with no rule. The gain is smaller than at 7B, not gone.
  • A request that names no service ("remove every edge with weight w") gives the scope filter nothing to read. 11 of 280 outputs were wrong, all of that kind, at both sizes.
  • It is a small model with constrained decoding, 30-step episodes and one run. The repair rule is the job's own rule.

AXI is a layer on the host side by Rajnish Choubey: the model is unchanged and the claims are about the state the host holds. The test pages are docs/LONG_HORIZON_TEST.md, docs/SCALE_TEST_14B.md and docs/SCOPE_TEST.md in the repository: https://github.com/gaurshalendra2412/axi-runtime

White paper, "AXI: The Cure for the Cost Crisis in AI" (Rajnish Choubey and Shalendra Gaur): https://doi.org/10.5281/zenodo.23958737