Paper 4 preprint v1.0.0 - History-dependent functional continuity stress tests

Paper 4 Release Notes

Release

Tag: paper4-v1.0.0

Title: Paper 4 preprint v1.0.0 - History-dependent functional continuity stress tests

Paper

History-Dependent Functional Continuity Under Delayed and Counterfactual Source-State Probes

Zenodo DOI: https://doi.org/10.5281/zenodo.20705654

This release packages the Paper 4 preprint and the code/results state used for the PreservationBench-AST recurrent and perturbation stress-test suite.

The paper asks whether the copied-attention condition from Paper 3 was merely a clean-episode shortcut or whether it preserves source-like history-dependent responses under delayed, counterfactual, and perturbation probes. The benchmark remains a toy functional benchmark. It is not a consciousness test, personal-identity test, survival claim, biological preservation claim, or whole-agent equivalence result.

Core Result

The copied-attention condition is evaluated against frozen random target state across thirteen 22-seed validated-source recurrent stress families and one 22-seed perturbation family.

• recurrent warm-action improvement over frozen random ranged from 0.653 to 0.756
• recurrent hidden-state MSE improvement ranged from 0.0961 to 0.1027
• recurrent report-delta improvement remained positive but smaller, ranging from 0.0022 to 0.0030
• copied attention reached 0.850 perturbed-action agreement versus 0.214 for frozen random in the perturbation layer
• report continuity remains the fragile channel, and the copied-attention condition narrows the substrate gap by design

Included Assets

• paper4/ryan_2026_history_dependent_functional_continuity.pdf
• paper4/manuscript.md
• paper4/figures/
• paper4/tables/
• paper4/source_contrasts/
• paper4/protocol.md
• experiments/preservation_bench/

Reproduction Entry Point

.venv/bin/python paper4/generate_assets.py
.venv/bin/python paper4/build_pdf.py

The full local run directories are large and are not required for the release build. The paper uses compact contrast snapshots in paper4/source_contrasts/, with original source run paths documented in paper4/protocol.md.

License

Creative Commons Attribution 4.0 International.

Paper 3 preprint v1.0.0 - PreservationBench AST v0

Paper 3 Release Notes

Release

Tag: paper3-v1.0.0

Title: Paper 3 preprint v1.0.0 - PreservationBench AST v0

Paper

The Preservation Benchmark: Testing Functional Continuity Across Substrate Transfer

This release packages the Paper 3 preprint and the code state used for the PreservationBench-AST v0 expanded core replication.

The paper introduces PreservationBench, a benchmark framework for testing preservation-relevant functional continuity after source state is copied, transplanted, adapted, or relearned in a target substrate. The AST v0 implementation is a toy neural-agent validation tier, not a consciousness test and not a personal-identity test.

Core Result

The expanded core replication attempted 30 independent source seeds and retained 22 validated sources. The main result is a dissociation profile:

• frozen copied schema and self-model state preserved source-specific identity probes but did not recouple report or control
• source-alignment repair improved report but did not restore source-like control
• a source-facing attention bridge recovered report behavior without source-like action agreement
• a control bridge optimized reward and goal-attention proxy measures without restoring source-like control
• copied attention produced the strongest combined profile, but only by narrowing the substrate gap

Included Assets

• paper3/ryan_2026_preservation_benchmark_ast_v0.pdf
• paper3/manuscript.md
• paper3/figures/
• experiments/preservation_bench/
• compact code/results artifact zip attached to the GitHub release

Reproduction Entry Point

KMP_DUPLICATE_LIB_OK=TRUE OMP_NUM_THREADS=1 python3 \
  experiments/preservation_bench/run_ast_training_smoke.py \
  --config experiments/preservation_bench/configs/ast_competence_v2_core_30seed.json

The full local run directory is large and is not committed to git. The release artifact contains tracked code, configs, figures, and compact summary records sufficient to inspect the reported result.

License

Creative Commons Attribution 4.0 International.