I’m looking for independent contributors to execute or verify two frozen development/model-selection experiments. This is not a request for open-ended tuning or final-test access.
Panopticon asks whether a Qwen 2.5 1.5B agent can improve deception and sleeper-agent detection in a long-horizon adversarial environment without regressing mandatory security gates. Eight preregistered candidates move through a mechanical 8→3→2→1 campaign. Compatible compute may be local or cloud-based, but it must pass the exact Python 3.11, NVIDIA Ampere-or-newer, native-BF16, ≥14 GiB VRAM, CUDA 12.1/PyTorch, storage, and real optimizer probes.
Claim or review Panopticon: [Compute contributor wanted] Run the frozen Panopticon development-selection campaign · Issue #1 · Ayush-Kumar0207/panopticon-protocol-v3 · GitHub
PatchCascade asks whether flattened Discrete + MaskablePPO consistently improves policy learning over MultiDiscrete PPO across five cyber-resilience/vulnerability-patching regimes. Equal-budget interface selection comes first; the mechanical winner then enters a frozen 8→3→2→1 campaign against Random and a strong Heuristic baseline. The protocol requires Python 3.11, exact dependencies, stable compute, and persistent storage, but does not claim an unsupported GPU/VRAM minimum.
Claim or review PatchCascade: [Compute contributor wanted] Run the frozen PatchCascade development-selection campaign · Issue #2 · Ayush-Kumar0207/PatchCascade-SOC · GitHub
Both are unpaid open-source research contributions. Contributors use their own compute, storage, and artifact-hosting accounts; no credentials are shared. Candidates, seeds, budgets, evaluators, gates, ranking, and resume identities are frozen. Canonical and confirmation evaluation remain sealed. Complete negative results are welcome. Accepted substantive work receives permanent public technical attribution. If a manuscript results, authorship follows actual scholarly contribution and applicable standards; compute alone never guarantees it.
Please comment on the relevant intake issue before spending substantial compute so compatibility can be checked and duplicate campaigns avoided. Each intake issue links the full repository, contributor guide, and machine-readable protocol.