README.md

metadata

license: other
tags:
  - physics
  - world-model
  - research
  - safetensors

JUMP observation-only learned-z engineering bundle

This repository contains the four verified components of the current Track H engineering pipeline: an observation-only encoder, a learned future-state decoder, a latent-to-Gemma projector, and a scalar injection gate. The frozen Gemma base model is referenced but is not redistributed here.

Result status

This is an engineering artifact, not scientific evidence of a useful or causal latent representation.

• Stage C completed three seeds and was a pivot: the learned predictor did not beat persistence under the preregistered gates. G2, G3, and mechanistic-evidence flags are false.
• Stage D completed the non-textual injection and six-arm control pipeline. The structured answer for own_z matched no_z, and donor swaps produced zero movement toward donor answers. informative_z, behavioral, causal, and mechanistic claims are false.
• The learned-decoder SVGs are predictions computed from the same serialized 16-dimensional z used by the injection path. They are not ground-truth renderings.

The small validation and control sets do not support benchmark law-accuracy claims.

Components

components.json is the machine-readable jump.world-model-components/v1 manifest. Verify it and every component checksum before loading.

Role	Files	Description
encoder	encoder/model.safetensors, encoder/config.json	observation-only MLP, 96 → 64 → 16
decoder	decoder/model.safetensors, decoder/config.json	learned same-z future-state MLP, 16 → 64 → 12
future_projector	future_projector/model.safetensors, future_projector/config.json	BF16 linear projector, weight shape [3840, 16]
gemma_adapter	gemma_adapter/adapter_model.safetensors, gemma_adapter/adapter_config.json	BF16 scalar tanh gate only

The gate artifact is not LoRA and not a PEFT adapter. No Gemma base weights, optimizer state, cache, or executable model code are included. auto_map is not used.

Exact external dependencies

• Base: google/gemma-4-12B-it@707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
• Transformers source: 918dbf131d0df5b46e3f6e1d96174d62aa4d16d6
• Stage D component manifest: 04f4b4ea6c7f4e6d517cd5a27925ed948bcde1d5d744f84edcf1ee8cdbe890bb
• Stage D checkpoint: stage-d-13c3d963b9ec7171f5d138a9e737b4b6294d542d0887dfbf9a52c2efba422071

Access to the external Gemma checkpoint may require accepting its separate terms. The Hub metadata for that checkpoint and the linked Gemma terms govern the base model. This repository makes no claim that those terms are replaced by the status of these JUMP-produced component files.

Verify before loading

Pin an immutable revision when downloading:

from huggingface_hub import snapshot_download
from pathlib import Path
from jump_contracts.distribution import (
    validate_world_model_component_manifest,
    verify_world_model_component_files,
)
import json

root = Path(snapshot_download(
    "Perfect7613/jump-world-model",
    revision="<PINNED_COMMIT_SHA>",
))
manifest = validate_world_model_component_manifest(
    json.loads((root / "components.json").read_text())
)
verify_world_model_component_files(root, manifest)

For artifact-only inspection, only the encoder and decoder are needed. Gated-Gemma loading additionally requires authorized access to the exact base revision. Instantiate model classes from the JUMP source package; this repository deliberately contains no remote Python implementation.

Provenance and limitations

STATUS.json contains the fail-closed claim guards. evidence/stage-c/ records the three-seed pivot. evidence/stage-d/ records the exact six-arm matrices, tensor bindings, learned-decoder images, promoted task/run evidence, and checksums. CHECKSUMS.sha256 covers this distribution snapshot except itself.

The generator is synthetic, the current predictive task did not outperform persistence, and the Stage D intervention produced a null answer shift. Do not use this bundle as evidence that z represents hidden laws or causally controls Gemma. A future model must first beat persistence on a leakage-resistant held-out task before efficacy claims are reconsidered.

legacy/scalar-baseline/ preserves the earlier scalar dynamics smoke and is excluded from the canonical component manifest.