Evaluate a batch of decisions in ONE parallel pass: choice (<=255 options), score (fractional scale) and noul (calibrated probability). Extra questions do not increase latency or cost.

Route one decision through the 'BELKİ' gatekeeper: confidence above the tuned threshold executes directly ($0 LLM), mid range triggers speculative sub-decisions, low confidence escalates to System 2.

Produce a steeled plan: System-2 hypothesis + hostile anti-thesis, scored by the Jev loop on feasibility, failure risk, cost and maintainability. Works offline with deterministic fallbacks.

Generate concrete anti-theses against a thesis, score their severity/evidence with the Jev loop and arbitrate the safest route with residual risks and mitigations.

Scan a project directory and score five dimensions (architecture, security, marketing, legal, budget) with findings, evidence and ranked fix actions.

Sweep web (Wikipedia/Brave), academic (arXiv), code (GitHub/npm) and social (HN/X) channels in parallel, re-rank with Jev relevance and synthesize a cited brief. Free/keyless sources by default.

Delete irrelevant log/grep/diff lines at $0 cost WITHOUT summarizing: kept lines are byte-identical. File paths, commands, error codes, URLs and diff headers are always preserved.

Audit a repository license (permissive vs copyleft/unknown) and, for restricted licenses, extract an architecture-only clean-room spec with a 5-gram similarity guard proving originality.

Label raw rows with Jev primitives in one fan-out pass and write a JSONL dataset into artifacts/datasets for distillation.

Score every candidate answer with the Jev Score primitive and emit chosen/rejected JSONL pairs for local DPO training.

Emit a concrete LoRA distillation recipe (axolotl/unsloth + vLLM) for Qwen2.5-0.5B or a ModernBERT-421M decision head; optionally write the axolotl YAML into artifacts/training.

Run verification commands (default: npx tsc --noEmit, npm test) in a working directory, record +1/-1 reward into the skill memory and auto-tune the gatekeeper thresholds.

Inspect and steer the arena-style skill memory: calibration buckets, verified win rate, threshold auto-tuning and manual outcome recording.

Read shared memory + task intent and pick the next agent role (researcher / planner / implementer / reviewer / writer / auditor) with a ready handoff payload.

Classify a tool call before execution: dangerous pattern blacklist + Jev Noul danger and Score severity produce an allow / ask / block verdict with reasons.

Mask PII and secrets (emails, cards, IBAN, Turkish ID, API keys, bearer tokens, private keys, env assignments) locally before anything is sent to an external model.

Re-rank retrieved passages with Jev Noul relevance and keep the top-K, dropping the noise.

Generate a deterministic edge-case test matrix (input, concurrency, dependency, auth, state, time, billing) for a feature spec.

Check a unified diff for breaking export removals, hardcoded secrets, console leftovers, new TODOs and dependency manifest changes; returns allow/ask/block.

List the 20 enterprise features with honest implemented/scaffolded status and their hosting tool.

Politely crawl the 11 curated sources (Farnam Street, LessWrong, Derek Sivers, Julian Shapiro, Internet Archive, Open Library, Project Gutenberg, Wikibooks, PhilArchive, PsyArXiv, CORE) into the local SQLite+FTS5 research hub. Respects robots.txt and each source's off-hours crawl window unless force=true.

FTS5 search over the crawled research hub, re-ranked by Jev Noul relevance into VERIFIED/PROBABLE/REJECTED tiers, then synthesized into a cited, step-by-step answer that grounds claims in mental models and cognitive-science findings over popular advice.

Report document counts, word totals and last-crawl timestamps per source in the research hub.

Report the resolved backend chain, pricing, gatekeeper policy, presets, memory summary and workspace paths so agents can self-diagnose the setup.