> **Reproducibility** > Three flagship research claims, re-derived by CI on a stated schedule in the source repos — not just disclosed as simulated, but continuously re-verified against the real code. > > Source: https://jasonstiltner.com/corpus/reproducibility/ --- # Reproducibility Every simulated result on this site already carries a “Simulation” tag — that's a disclosure. This page is the next step: the statistics below are re-derived by CI, in the public repo, on a schedule stated for each one. Not a promise that a number can't change — a standing invitation to check whether it still does. The number shown here is whatever each source repo's CI most recently published as of this site's last build/deploy — not a live, between-deploys feed. Site build last pulled fresh CI data at Sep 21, 2026, 3:58 PM UTC. ## The Punishment Paradox [View on project page →](https://jasonstiltner.com/projects/grounded-commitment-learning/#punishment-paradox) Published, full-scale: r = -0.972, p < 0.001 — increasing consequences decreases cooperation CI-reproduced small-scale — refreshed on every push + weekly r = -0.818 (n=5 seeds) Verified Sep 21, 2026 against commit 55c4ada of the real simulation code, not a copy of this page's numbers. [CI run](https://github.com/jstiltner/gcl/actions/runs/35615288861) · [source](https://github.com/jstiltner/gcl) · [reproduce full scale in Colab](https://colab.research.google.com/github/jstiltner/gcl/blob/main/notebooks/reproduce_headline_stats.ipynb) ## Hart-Moore Hold-up Reduction [View on project page →](https://jasonstiltner.com/projects/grounded-commitment-learning/#hart-moore) Published, full-scale: 40.4% hold-up reduction — GCL's failure-first specification vs. incomplete contracts CI-reproduced small-scale — refreshed on every push + weekly 42.9% hold-up reduction (n=5 seeds) Verified Sep 21, 2026 against commit 55c4ada of the real simulation code, not a copy of this page's numbers. [CI run](https://github.com/jstiltner/gcl/actions/runs/35615288861) · [source](https://github.com/jstiltner/gcl) · [reproduce full scale in Colab](https://colab.research.google.com/github/jstiltner/gcl/blob/main/notebooks/reproduce_headline_stats.ipynb) ## CNL Bridge Ablation [View on project page →](https://jasonstiltner.com/projects/collaborative-nested-learning/#results) Published, full-scale: +89% accuracy at high regularization — bidirectional knowledge bridges vs. CMS alone CI-reproduced full-scale — refreshed weekly +88.5% at reg=5 (bridges helped 5/5 settings) Verified Sep 21, 2026 against commit a2d08b7 of the real simulation code, not a copy of this page's numbers. [CI run](https://github.com/jstiltner/collaborative-nested-learning/actions/runs/35612744655) · [source](https://github.com/jstiltner/collaborative-nested-learning) · [reproduce full scale in Colab](https://colab.research.google.com/github/jstiltner/collaborative-nested-learning/blob/main/notebooks/reproduce_bridge_ablation.ipynb) ## How this works - Each source repo ([jstiltner/gcl](https://github.com/jstiltner/gcl), [jstiltner/collaborative-nested-learning](https://github.com/jstiltner/collaborative-nested-learning)) runs its own GitHub Actions workflow that re-derives the number from the real simulation code and commits the result back to `ci_results/latest.json`. - This site fetches that file at build time (`scripts/fetch-ci-results.ts`) — a build never fails because a fetch didn't succeed; it falls back to the last-known-good copy. - A CI-derived number is deliberately not the same run as the published claim — see each repo's CHANGELOG for why (in short: GCL's reduced-scale run preserves the right direction with wider intervals; CNL's bridge effect needs closer to full scale, so its badge only refreshes weekly, at full scale, not on every push). - Every “Open in Colab” link runs the exact same code at the original published scale, interactively, for anyone who wants to check it themselves.