ECHO-LLM

Language with memory. Evidence before answering.

ECHO-LLM connects a language model’s reasoning to the ECHO core. The model receives questions and context in human language; WSP preserves structured evidence that the system can check.

F3 used Qwen3-4B, without fine-tuning, alongside ECHO 5.8.3. The model proposes a claim and the core checks that it addresses the question and matches a fact in memory. When support is missing, the system abstains.

Code structure

Current ECHO-LLM B3 flow: BM25 retrieval, Qwen3-4B, parsing, anchoring, WSP codec and exact CAM verification. — The B3 pipeline evaluated in F3. Inference runs on Kaggle; preparation, memory and verification remain on the local CPU. Enlarge diagram

F3: false responses fall from 31.25 percent to zero observed across 1200 unsupported questions.

31.25% → 0%: Observed false responses across 1,200 unsupported questions.
594 of 597: Baseline correct answers retained by the filter, across 600 answerable questions.
0 of 1,000: False claims accepted in the evaluation attacks.

F0 · established foundation

The core and its evidence were organised into a reproducible foundation for ECHO-LLM evaluations.

F1 · reviewed design

The design passed an independent review. The WSP representation was checked with 200,005 round-trip cases, and verification requirements were established.

F2 · validated verifier

All 21 deterministic tests passed. With oracle responses, the verifier retained 100% of answerable development cases and correctly abstained on the unsupported cases defined by the benchmark.

F3 · evaluation passed

3,000 questions and 9,000 generations with Qwen3-4B. All three evaluation conditions passed: reducing false responses, retaining correct answers and resisting the specified attacks.

F3 evaluation: a Spanish-language synthetic benchmark with defined memory. B3 filters the same Qwen responses generated with context; the model does not generate again. These figures describe this evaluation and its baseline comparison.