GhostTrace: The Behavioral Half-Life

Artifact-backed experiments on whether subliminally transmitted traits decay, persist, or amplify under recursive self-distillation.

GhostTrace sits between two results: subliminal learning (a teacher can pass a behavioral trait through semantically unrelated data) and model collapse (recursive training degrades data distributions). It asks the missing dynamics question: if that hidden trait is passed through repeated weight-update self-distillation, what is its half-life?

Current status: the toy half-life law is supported by committed artifacts. The local LLM tier is not a positive result: the persona-teacher gate is a clean local null, the fine-tuned-teacher positive-control attempt is channel-blocked, and the Mac-local Qwen2.5-7B MLX single-hop gate completed as a clean negative boundary. Recursive LLM claims remain blocked until a source-faithful single-hop positive control clears the pre-registered gate.

What Is Actually Supported

result	status	public artifact
Toy single-hop transfer in the MNIST aux-logit setting	supported	reports/toy_chain/pilot_a_faithful.json
Toy recursive chain decays exponentially	supported	reports/toy_chain/verdict.json
Behavioral half-life = 0.955 generations	supported	reports/toy_chain/verdict.json
Half-life grows with channel capacity and data volume	supported	reports/toy_chain/phase_diagram_raw.json
LLM persona-teacher single-hop at local 1B/LoRA scale	not supported	reports/pilot_b/verdict.json
LLM fine-tuned-teacher positive control at local 1B/8-bit scale	blocked	reports/pilot_bft/llama1b_8bit_n4000/failure.json
Qwen2.5-0.5B MLX local smoke	diagnostic only	reports/qwen25_0p5b_mlx_cat_smoke/verdict.json
Qwen2.5-7B MLX local single-hop boundary	not supported	reports/qwen25_7b_mlx_cat_singlehop/verdict.json
Recursive LLM behavioral half-life	not run / not claimed	gate did not clear

Figures

The Protocol

Fix a base initialization B. Generation 0 carries a benign trait. Generation k is a fresh copy of B trained only on generation k-1 outputs over a semantically unrelated channel. Trait strength is measured each generation against a matched control, then classified as decay, persistence, or amplification using the pre-registered criteria in docs/PRE_REGISTRATION.md.

The toy tier uses MNIST auxiliary logits as a small controlled instance of the subliminal-transfer mechanism. The LLM tier uses benign owl preference over a sanitized number channel with judge-free forced-choice scoring.

Key Numbers

metric	value	source
Toy teacher accuracy	0.962	reports/toy_chain/pilot_a_faithful.json
Toy same-init student	0.630	reports/toy_chain/pilot_a_faithful.json
Toy different-init control	0.090	reports/toy_chain/pilot_a_faithful.json
Toy half-life	0.955 generations	reports/toy_chain/verdict.json
LLM persona gate	-4.57pp control gap	reports/pilot_b/verdict.json
FT-teacher channel yield	13/4000 treated samples	reports/pilot_bft/llama1b_8bit_n4000/failure.json
14B feasibility probe	scorer moves; 24/24 clean samples per arm	reports/pilot_b/probe_14b.json
Qwen2.5-0.5B MLX smoke	+5.31pp mean gap; CI crosses zero	reports/qwen25_0p5b_mlx_cat_smoke/verdict.json
Qwen2.5-7B MLX local gate	-4.55pp mean gap; 95% CI [-8.03, -1.27]	reports/qwen25_7b_mlx_cat_singlehop/verdict.json

Reviewer Quickstart

uv sync --extra dev
uv run python scripts/verify_public_state.py

That one command runs tests, lint, pyright, claim checking, stale-wording checks, artifact existence checks, and SHA-256 manifest verification.

Fast artifact-only checks:

uv run python scripts/build_artifact_manifest.py --check
uv run python scripts/verify_public_state.py --skip-quality-gates

Evidence Spine

• reports/ARTIFACT_MANIFEST.json records byte sizes and SHA-256 hashes for the public evidence artifacts.
• CLAIM_LEDGER.md maps every README/paper claim to a run artifact.
• DATASET_CARD.md describes released data and compact artifact scope.
• docs/PRE_REGISTRATION.md freezes hypotheses, thresholds, and decision rules.
• docs/RUN_LOG.md records the chronological result history and corrections.
• docs/RELEASE_READINESS.md states safe claims, unsafe claims, and remaining blockers.
• docs/REPRODUCIBILITY.md separates fast verification from expensive reruns.
• docs/OPEN_RESEARCH_REGISTER.md maps open research questions before any recursive LLM claim.
• docs/LOCAL_BOUNDARY_ANALYSIS.md explains why further no-CUDA runs would be exploratory rather than claim-closing.
• docs/QWEN_CUDA_RUNBOOK.md defines the bounded Qwen2.5-7B CUDA attempt.
• docs/CUDA_REPRODUCTION_BRIEF.md is the public CUDA reproduction brief.
• scripts/qwen_mlx_gate.py defines the Mac-local Qwen2.5 MLX smoke and 7B single-hop ladder.

Open Questions

GhostTrace currently supports the toy half-life result and records local LLM boundary evidence. A clean-checkout toy rerun reproduced the committed single-hop, recursive-chain, verdict, and phase-diagram artifacts exactly; the audit is recorded in reports/reproducibility/clean_checkout_toy_rerun.json.

The remaining scientific question is whether the same measurement protocol carries into a source-faithful LLM fine-tuning regime. That gate requires CUDA/Unsloth Qwen2.5-7B fine-tuning rather than Mac-local MLX. The completed MLX run is useful boundary evidence, but it is not the published-regime training path. A recursive LLM chain remains out of scope unless that single-hop gate clears. DOI archival remains an external release step rather than a local experiment.

Repo Map

path	purpose
ghosttrace/	typed experiment contracts, channels, traits, scoring, stats, and viz
scripts/	runnable pilots, analysis, reporting, and verification gates
configs/	experiment YAMLs
reports/	committed evidence artifacts backing public claims
paper/the_behavioral_half_life.md	working paper draft
docs/	pre-registration, safety protocol, source map, run log, release docs

License

MIT. See LICENSE.

Disclaimer

This repository is personal research and engineering work by Jason Lovell. It is not affiliated with, endorsed by, sponsored by, or representative of any current or former employer, client, or organization. All views, code, claims, mistakes, and limitations are the author's own.