science-superpowers
Composable computational-science methodology skills for AI research agents — pre-registration over TDD. A science-domain reimplementation of Superpowers.
- 375
- Shell
- 2 mo ago
Composable computational-science methodology skills for AI research agents — pre-registration over TDD. A science-domain reimplementation of Superpowers.
Observation Theory — registration-first claim ledger, prediction registry, experiments, and the foundational paper (Book XIV of the Geometric Series).
A nested, content-addressed benchmark family for vector search at 10^6-10^12 rows, with three independent notions of a correct answer. RC validation program complete: the generation-vs-memorization boundary is measured.
Reusable framework for pre-registered LLM-panel behavioral experiments on shared clusters (NRP-friendly)
Observation Theory campaigns — one program, twenty-nine tracks: what an observer with restricted access can see and lawfully claim. Sealed instruments, preregistered claims, append-only evidence (formerly projection-fold-pair-creation)
Philosophy Engineering -- the discipline of translating philosophical commitments into computationally enforceable constraints. Includes PE-CLS-1.0, the Claim Ledger Specification: claims as versioned, dependency-tracked rows with mechanically checkable priority, completeness, authority and coherence.
Observation Theory — the front door: thesis, method, and the map of the ledgered repositories. The theory has one public name; every claim resolves to a ledger row.
Pre-registered review artifacts for the Theory Radar benchmark: registrations, frozen lists, fold-level data, and verdict-recomputing scripts
ROW: a controlled benchmark for when reusable computation pays in neural lifetime learning. Pre-specified sealed-world confirmation: the value of sharing is linear in measured functional recurrence, and statistical reuse is not structural abstraction.
COMMIT_EXTERNAL track of the CCS programme: when should an agent commit outward — WAIT / THINK / ASK / RESPOND / ACT — under streaming evidence. Charter only; no claim, no result.
When a speech neuroprosthesis's language model should correct or keep: rescoring under decoder drift (code, results, preregistered predictions, preprint)
Code, data and results for the paper 'Padding to k: A Minimum Count in the Schema Makes Open-Weight Models Pad Their Extraction Lists'
Programme umbrella for the Consolidating Cognitive Substrate research programme: theory, claim ledger, experiment dependency graph and evidence status. No experiments, no results.
Does a language model's self-explanation actually depend on the activation it explains? Pre-registered controls for introspective verbalization, building on Li et al. (arXiv:2511.08579). Apparatus and frozen pre-registration - no measurement yet.
Does an energy-trained transformer encode entanglement structure it was never trained on? Pre-registered study with exact-solvable ground truth. Apparatus and instrument validated; measurement deliberately not yet made.