docs/xor.md

xor

xor is a 35B MoE System-1 decision model. Post-trained from Qwen3.6-35B-A3B (35B total, ~3B activated per token), it extracts hidden states via ForwardMoeHidden — the MoE variant of the ForwardHidden pooling path — and reads candidate logits through the LM head. Each question runs twice (forward and reverse option order) and the probabilities are calibrated. It answers choice, score, and noul question types through /v1/systemone and supports up to 8 images.

xor runs through the shared systemone lane, so the usage guide covers the /v1/systemone request shape. This page carries the checkpoint, the forward+reverse calibration, and what has not been measured.

The checkpoint#

The Qwen3.6-35B-A3B MoE backbone loads via the existing qwen3_5_weights.cpp loader. The vision tower weights and decision head weights are included in the merged checkpoint.

Run it#

build/examples/vllm-server --model /path/to/xor --port 8000

Then send a systemone request:

curl http://localhost:8000/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{"state":"john works at google",
       "questions":{"pick":{"type":"choice","instructions":"entity type",
       "criteria":{"person":null,"organization":null}}}}'

Forward+reverse evaluation doubles inference cost per question. For a 35B MoE this is significant but inherent to the method. Prefix caching of the shared state prefix across the two passes is deferred. Confidence uses the shared kev formulas (ChoiceConfidence, ScoreConfidence); probabilities are rounded to 2 decimals.

What has not been measured#

No token gate result exists for this checkpoint yet. No byte has been fetched or hashed. The golden tests (softmax, forward+reverse calibration, confidence formulas, prompt construction) run on synthetic weights.

GPU (CUDA) build verification, GGUF k-quants, E2E parity vs the SGLang oracle, and prefix caching are owed.

Spec: .agents/specs/xor.md