Menu
Sketched Isotropic Gaussian Regularization primitive. Scalar loss matching the embedding distribution to a standard-normal target via Cramér-Wold slicing and the Epps-Pulley empirical characteristic function test. Port of rbalestr-lab/lejepa (MIT). Default-off in v1.49.571.
Generate intelligence briefings for the planning dashboard. Use this skill whenever a request file appears in `.planning/console/inbox/pending/` whose `type` field starts with `intelligence.` (refresh_briefing, triage_finding, snapshot_diff, investigate_section, dismiss_finding). The skill reads the per-project KB at `.gsd/intelligence/intelligence.db`, synthesizes a briefing with a causal hypothesis + acknowledged uncertainty + confidence label and ranked moves, then writes the result back to the KB. Always trigger this skill for these request types — do not generate briefings manually.
Adversarial spec-compliance PR review — cross-references diffs against approved specs, verifies runtime claims against source, detects competing PRs, audits scope/convention compliance. Use before merging.
Pick among candidate outputs (code, configs, plans) by running them on diverse inputs and clustering by behavioural fingerprint, rather than by textual aggregation or log-probability. Activates when an executor returns multiple plausible candidates that need disambiguation, when output-majority voting would be the default choice, or when reviewing generated code that has not yet been validated. The 2026 evidence (Semantic Voting, arxiv 2605.08680v1) is that any execution-based selector dominates output-majority voting by 19-52pp; sketch-generated inputs beat random fuzz by 11.3pp. Triggers: "pick the best candidate", "majority vote on code", "select from N samples", "validate the generated output", "behavioural verification".
Extract creative intent from images into executable build specs. Activates on images + build intent, "image to mission", "i2m", or capturing visual energy in code/design.
Classify the information-need of a query and dispatch it to the appropriate retrieval or reasoning strategy. Use before read-side memory access, before multi-strategy retrieval, or any time you'd otherwise default to "one retriever for everything". Returns a strategy label, a token budget, and a retrieval depth so downstream handlers can be specialised. Backed by Pre-Route (arxiv 2605.10235v2) and MemFlow (arxiv 2605.03312v1), which together show LLMs possess latent routing ability elicitable via a structured prompt — and that externalising the routing decision improves small-model performance by ~2x. Triggers: "route this", "what strategy", "before retrieving", "intent classification", or any query whose ideal handling depends on what KIND of question it is.
Audit a skill by running a paired probe — the same task once with the skill loaded and once without — segment both traces into goal-directed phases, align phases, and emit a SIP report (surface anchoring, template copy, excess planning, task recovery, off-task artifact). Use whenever a skill is created, modified, or proposed for retirement. Pass-rate is BLIND to most skill effects: CTA (arxiv 2605.11946v1) shows a single skill can produce 522 measurable behavioural changes across 49 tasks while pass-rate moves only +0.3%. Triggers: "audit this skill", "is this skill helping", "retire skill", "before shipping skill", "behavioural impact of skill X", or any skill review event.
| name | sigreg |
| description | Sketched Isotropic Gaussian Regularization primitive. Scalar loss matching the embedding distribution to a standard-normal target via Cramér-Wold slicing and the Epps-Pulley empirical characteristic function test. Port of rbalestr-lab/lejepa (MIT). Default-off in v1.49.571. |
| version | 1.49.571 |
| tier | T1 |
| phase | 729 |
| source_paper | arXiv:2511.08544v3 |
| license_attribution | rbalestr-lab/lejepa (MIT) |
Port of SIGReg from Balestriero & LeCun (2025, LeJEPA) arXiv:2511.08544v3. The primitive computes a scalar loss measuring how far an embedding distribution is from the standard-normal target, using Cramér-Wold slicing plus the Epps-Pulley empirical characteristic function test. Linear O(N·M·K) time, naturally differentiable, multi-GPU friendly (all_reduce over ECF averages).
import { sigreg } from './src/sigreg/index.js';
const loss = sigreg(embeddings); // embeddings: number[num_samples][num_dims]
// scalar loss; use as L_total = L_pred + λ · loss
For telemetry:
import { sigregWithBreakdown } from './src/sigreg/index.js';
const { loss, perSliceStatistic, maxSliceStatistic, runTag } = sigregWithBreakdown(embeddings);
Default matches the LeJEPA reference implementation:
const LEJEPA_DEFAULT_CONFIG = {
numSlices: 1024,
univariateTest: { numPoints: 17, sigma: 1.0 },
};
Default-off. Opt-in via .claude/gsd-skill-creator.json:
{
"heuristics-free-skill-space": {
"sigreg": { "enabled": true }
}
}
Ported from https://github.com/rbalestr-lab/lejepa under the MIT license.
© Randall Balestriero and LeJEPA Contributors. See ../../license_notices.md.
src/skill-isotropy/ — Skill Space Isotropy Audit (Phase 728) — read-only audit use case