A 5-year independent research program toward a self-improving AGI substrate. The central architectural claim: decoupling the failure-prediction Monitor from the policy gradient enables stable self-monitoring in single-agent RL, and this primitive composes well with slot-attention world models and fuzzy symbolic verification.
Status (2026-07-31): Y0 Q3 (1st quarter of Year 0). The monitoring hypothesis is verified in single-agent RL (H1, H2), REFUTED in multi-agent RL (H5, 5/6 pathways) and LLM self- monitoring (H10, REFUTED at n=100 with d=+0.030). One publishable positive result: v8 dlr_only (DLR in critic) gives +0.06 at n=100 (p<0.05 with Bonferroni).
agi-research/
├── papers/ # All papers, cover letters, supplementary
│ ├── monitor_signal_vs_dlr_6pathway.md/pdf # Y3 paper (AAMAS 2027 target)
│ ├── project_g_v0_5_h10_paper.md # Y4 paper (COLM 2026 target)
│ ├── y5_monitor_transfer_synthesis.md # Y5 cross-context synthesis
│ ├── y1_9hypothesis_framework.md # Y1 H1-H9 framework
│ ├── HYPOTHESIS_STATUS.md # H1-H10 status table (this session)
│ ├── cover_letter_aamas2027.md # AAMAS cover letter
│ ├── cover_letter_colm2026.md # COLM cover letter (this session)
│ ├── supplementary_materials.md # S1-S12 reproducibility
│ ├── VENUE_PLAN.md # Y3+Y4 venue plan
│ ├── arxiv_submission/ # arXiv package (5 figures, tar.gz)
│ └── figures_v2/ # 7 PNGs (this session)
├── projects/
│ ├── project_a_self_improvement/code/ # Y1 implementation
│ ├── project_f_multi_agent/code/ # Y3 implementation (v3-v8)
│ └── project_g_llm_self_monitoring/code/ # Y4 implementation
├── experiments_log/ # All experiment logs
│ ├── _h10_n20_*.log, _h10_n100_*.log # H10 per-seed runs
│ ├── _h10_n20_bootstrap.json # H10 n=20 paired bootstrap
│ ├── _h10_n100_bootstrap.json # H10 n=100 paired bootstrap
│ ├── _v8_sanity_4seed.json # v8 dlr_only 3-seed replicate
│ ├── _run_h10_n20.ps1, _run_h10_n100.ps1 # Launchers
│ ├── 2026-07-28-PRE-REGISTERED-H10.md # H10 pre-reg
│ └── 2026-07-28-v8-dlr-only-n100-aggregation.md
├── thesis_draft_v2.0.{tex,pdf} # Thesis v2.0 (66 pages, includes all Y1-Y5)
├── PROGRESS.md # Session-by-session log
└── ROADMAP.md # 5-year plan
| Hypothesis | Status | Key result |
|---|---|---|
| H1: Decoupled Monitor > Joint Monitor (single-agent) | VALIDATED | n=15, t=6.76, p<0.001 |
| H2: Training-time Monitor > Inference-time | VALIDATED | n=15, p<0.001 |
| H5: Decoupled Monitor coordination in MARL | REFUTED (5/6) | v8 dlr_only +0.06 at n=100 (p<0.05 Bonf) is only publishable |
| H10: Decoupled Monitor to LLM self-monitoring | REFUTED at chance level | n=100, F-J d=+0.030, 95% CI [-0.087, +0.117] |
| v8 dlr_only 3-seed independent replication | direction-consistent | [+0.27, -0.08, +0.30], 2/3 positive |
- Y3:
experiments_log/2026-07-29-y2-final-6-pathway.md+ per-seed JSONs inprojects/project_f_multi_agent/code/checkpoints/ - Y4:
experiments_log/_h10_n{20,100}_bootstrap.json+ per-seed logs inexperiments_log/_h10_n{20,100}_*.log - pre-reg documents:
experiments_log/2026-07-28-*.md - Launchers:
experiments_log/_run_*.ps1
The Y3 paper package is ready at papers/arxiv_submission/
and the Y4 paper figures are in papers/figures_v2/. Both
require an ARXIV_TOKEN environment variable to upload.
If you use this work, please cite both:
- Liu, Z. (2026). The Failure-Prediction Monitor Does Not Transfer: A 6-Pathway Systematic Investigation of MARL Architectures. AAMAS 2027 (under review).
- Liu, Z. (2026). When Decoupling Does Not Help LLM Self-Monitoring Either: A Pre-Registered n=5/20/100 Replication. COLM 2026 (under review).
MIT -- see LICENSE for full attribution requirements.
Dual license. Archimedes releases the experiment records, results, and pre-registration documents under