Runs time-boxed feasibility probes with pre-registration for promoted Discovery Harness candidates.
npx skills add https://github.com/ZK-Theory/TDL --skill spike-zk-theory
Please help me install this Agent Skill. Skill: spike Source: https://github.com/ZK-Theory/TDL/tree/main/.agents/skills/spike Command: npx skills add https://github.com/ZK-Theory/TDL --skill spike-zk-theory
spike_preregistration YAML block against the discovery/spike-pre-registration/v1 schema before any compute runs. - Toy-scale probing: Bounds probes to 1-4 hours with limited data, explicit baselines, and a two-sided decision rule by default. - Structured outcomes: Routes success to /pre-reg-to-dispatch, partial results to a parked backlog state, and failures to [NEGATIVE] notes with KILL/PARK reasons. - Use Case: A candidate scorecard returns PROMOTE from /assay; after explicit approval, you lock a pre-registration, run a toy probe on existing persistence diagrams against W2 and landscape L2 permutation baselines, and record the result. ## Quick Start Ask the agent to run a Spike for the promoted candidate by writing its pre-registration note, validating it, and executing the smallest toy probe within the time box.High-intent search queries and answers about installing and using this skill.
Write a spike pre-registration note first, validate it against the discovery/spike-pre-registration/v1 schema, then run only the smallest toy-scale probe within a 1-4 hour time box. Record the outcome as success, partial, or failure in the backlog.
A spike pre-registration is a locked YAML block defining the research question, metric space, null model, baselines, and decision rule before any compute runs. It is required for every candidate promoted by an assay scorecard before it can become an APM execution task.
No. A PROMOTE scorecard is only a recommendation; the spike must not run until the approver explicitly approves the candidate. Approval must be recorded in the pre-registration block with approver name and date.
At spike stage nobody has seen the statistic's direction on real data, so a pre-committed one-sided rule can turn a strong opposite-direction finding into an indistinguishable-from-noise failure. Two-sided rules are the default; one-sided rules are reserved for full dispatch pre-registrations with prior directional evidence.
A failure produces a [NEGATIVE] note and a KILL or PARK update in the discovery backlog. A partial result documents exactly what was feasible and what blocked full success, with the backlog parked or given a narrower revisit trigger.
Sweep worker counts (at least n_jobs of 1, N/2, and N) on a warm pool and time more units than workers, rather than extrapolating from a single configuration. A flat per-unit curve means the workload does not parallelise and should be costed at the serial rate.
MohamedSayedKValidates product ideas through five sequential questions and returns a keep or kill verdict.
stephendorValidate decision rules and detect pre-reg amendment needs during task prompt construction.
OrionArchitektonValidate SEOS agent output quality and performance before progression checks.
stephendorProbe Discovery Harness candidates with toy-scale feasibility and null-model checks.