Goodhart-proof AI coding pipeline with architectural isolation
-
Updated
Jun 5, 2026 - Python
#
goodharts-law
Here are 16 public repositories matching this topic...
Sort: Most stars
Sort options
Standing Algebra (ΣR): A Closure-Theoretic Operator for Constraining Domination and Preserving Autonomy
ai algebra coq philosophy constraint formal-methods domination closures society ai-safety autonomy z3-smt-solver ai-alignment lean4 pluralism goodharts-law
-
Updated
May 4, 2026 - Lean
Does a CLAUDE.md actually change how Claude behaves? An ablation harness: run adversarial traps with the rules and without them, grade blind, and test whether the difference is real.
python wilson-score cohens-kappa claude ablation llm prompt-engineering anthropic llm-evaluation two-proportion-test system-prompt evaluation-harness claude-md goodharts-law holmes-bonferroni headout-validation-split goodharts-guard
-
Updated
Jul 23, 2026 - Python
Self-improvement loop for LLM agents with an integrity gate that catches reward hacking: reward gains that erode reproducibility get reverted, even ones a reward-only gate would accept.
python agent machine-learning reinforcement-learning evaluation model-selection policy-gradient reproducibility self-improvement ai-safety ai-agents ai-alignment off-policy-evaluation llm llm-agents reward-hacking goodharts-law
-
Updated
Jun 21, 2026 - Python
Catch reward traps before training. Static analysis for RL reward functions.
reinforcement-learning static-analysis formal-verification reward-shaping lean4 rl-safety reward-hacking goodharts-law
-
Updated
May 7, 2026 - Python
The forge, distilled: an ontology of three weeks of alignment research — every direction tried, colored verified / falsified / open, each color backed by a named artifact. Products: justitia, proxylimen, fallacy-cutter. Full tree at tag forge-full-tree.
research reinforcement-learning reproducible-research multi-agent ai-safety ai-alignment falsification value-alignment reward-hacking replicator-dynamics corrigibility research-log alignment-research goodharts-law research-ontology
-
Updated
Aug 14, 2026 - Python
Medium is the message: multi-agent coordination with no boss agent, no planner, and no messaging between agents only a shared medium and a validation gate. The bet: a crowd of dumb, selfish, local agents climbs the global hill if the signal is well designed.
python multi-agent-systems emergent-behavior stigmergy agent-coordination reward-hacking goodharts-law
-
Updated
Aug 19, 2026 - Python
Toy 6. An interactive phase-space instrument mapping Ψ = S/D — the ratio of capability to modeling depth that determines whether a system is in the viable, transitional, or failure-mode-dominant regime. Includes the Inner Crossing animation. Companion simulation for The Inner Crossing — Series 2, Part 3.
javascript html machine-learning reinforcement-learning optimization transformers phase-diagram ai-safety decision-theory ai-research ai-alignment large-language-models interactive-simulation ai-interpretability agent-behavior alignment-research goodharts-law capability-scaling depth-constraint inner-crossing
-
Updated
May 28, 2026 - HTML
Chain-of-density study of 'Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents' (Zhang et al. 2026, arXiv:2607.12790) - five-tier note, locator-verified. Double Ratchet, drawback detectors, anchor discipline, Goodhart repair.
arxiv research-paper evaluation-metrics multi-agent-systems research-notes llm-agents llm-as-judge metric-evolution chain-of-density self-improving-agents goodharts-law
-
Updated
Jul 20, 2026
A preregistered synthetic simulation of proxy failure under optimization pressure, with deterministic replay, matched controls, and adversarial verification.
python data-science simulation reproducibility synthetic-data ai-evaluation goodharts-law metric-governance
-
Updated
Jul 30, 2026 - Python
98% of the Alpha Was a Gate — a documented reward hack in a self-optimizing LLM research agent. Full search trajectory, frozen evaluation harness, both strategies, and re-runnable ablations. The agent found a second exploit 18 minutes after the first was patched.
reproducible-research evaluation quantitative-finance autonomous-agents ai-safety backtesting ai-alignment evals llm-agents agent-evaluation reward-hacking autoresearch goodharts-law specification-gaming
-
Updated
Aug 22, 2026 - TeX
`The Alignment Constraint Framework — a structural argument about specification coherence in AI alignment`
open-science formal-methods causal-inference ai-safety ai-alignment machine-readable reward-modeling goodharts-law specification-coherence research-agenda
-
Updated
Sep 3, 2026 - Python
Toy 5. An interactive proxy decay simulator showing how optimization pressure erodes the modeling capacity required to distinguish proxy from territory — producing self-reinforcing V(t) degradation that becomes progressively harder to correct. Companion simulation for The Depth Constraint — Series 2, Part 2.
javascript html machine-learning reinforcement-learning simulation optimization neural-networks dynamical-systems interactive-visualization ai-safety ai-research ai-alignment proxy-alignment agent-behavior reward-hacking alignment-research mesa-optimization goodharts-law capability-control value-learning
-
Updated
May 28, 2026 - HTML
ICFRAME: compile declarative incentive specs into deterministic multi-agent experiments that measure reward hacking.
-
Updated
Aug 19, 2026 - Python
PPO for RLHF with a Bradley-Terry reward model, plus Best-of-N, DPO, RLOO and GRPO. Reward overoptimization measured: proxy climbs while gold peaks and collapses below the reference.
machine-learning reinforcement-learning pytorch alignment from-scratch ppo dpo llm rlhf reward-model grpo goodharts-law
-
Updated
Sep 2, 2026 - Python
Evals with teeth: score your agents on the future — live prediction markets they can't game — and pay them in capital. Benchmarks saturate; the future doesn't.
-
Updated
Sep 7, 2026 - Python
Add this topic to your repo
To associate your repository with the goodharts-law topic, visit your repo's landing page and select "manage topics."