Research paper · July 2026

Behavior
Before Perplexity

A behavior-first compression recipe for agentic language models under an exact physical byte ceiling.

8.39 GBPortable artifact
2,416Tensor payloads
92.2%Conditional retention
43.16 tok/sGeneration, RTX PRO 6000

The model that looked healthy was behaviorally dead.

Local tensor error was not enough. Aggregate language metrics were not enough. Several candidates retained plausible token statistics while producing malformed calls, failing to stop, or losing multi-call structure. That failure changed the promotion rule: a compressed layer advanced only when the packed artifact preserved the behavior users would actually observe.

The released file is 8,392,369,600 bytes and retains all 64 text-decoder layers. On a fresh 100-turn tool-contract gate, the full-precision teacher solved 90 turns and BTL-3 Compact retained 83 of those 90. Conditional retention was 100% for single, parallel, sequential, and abstention cases, and 30% for parallel-multiple calls.

Claim boundary

The 92.2% figure is conditional retention on one private tool-contract gate. It is not “intelligence retained,” a coding score, or evidence that every category cleared 90%. Parallel-multiple remains the measured weakness.

A selection procedure, not a magic quantizer.

The contribution is the ordered, falsifiable system used to find and repair behavioral cliffs under a fixed byte budget. Its components have prior art; the measured cookbook and deployed artifact are ours.

  1. 01

    Start from behavior

    Reject candidates that preserve perplexity or token agreement but cannot emit valid, correctly stopped tool calls.

  2. 02

    Localize the first cliff

    Replay nested compressed prefixes, then override modules at the first failing boundary instead of blaming the whole model.

  3. 03

    Allocate measured precision

    Use AVQ2 for the vector core, INT4 and BF16 only where executable behavior proves that additional precision earns its bytes.

  4. 04

    Repair the interface

    Rescue structural vocabulary rows, correct the output head, and train a bounded behavior adapter against the packed artifact.

  5. 05

    Prove the physical file

    Export one native GGUF, verify every tensor and checksum, then run the same deployed representation used by the release gate.

Vector core

AVQ2 affine-lattice codes, Hadamard-style incoherence processing, block-LDLQ reconstruction, and physical byte accounting.

Precision islands

Measured INT4 and twelve retained BF16 islands where lower precision caused cumulative behavioral failure.

Interface repair

4,096 rescued embedding rows, a rank-32 output-head residual, and a counted 32.46 MB rank-8 behavior adapter.

Native proof

A standalone GGUF with byte-verified tensors, packed CUDA and Metal execution, manifests, checksums, and no BF16 fallback.

Read and reproduce

The complete record is public.

Download the academic paper, read the chronological engineering account, or inspect the LaTeX source, figures, claim ledger, manifest, bibliography, and checksums on GitHub.