Strategy reference

Rule Lab — Canonical Reference

Status: LIVE TOOL (not a strategy). The Rule lab is an evidence INSTRUMENT: it takes a user-composed rule and returns an honest verdict from the same gauntlet every house study passes. It never originates a strategy, a rank of its own, or a recommendation. Surface: /dash/rule-lab (owner-gated composer; anonymous sees a read-only demo verdict). · Governing record: the strategy ledger (verdicts append there only after human approval in the Review Inbox) · .

Origin: 🏠 HOUSE ( §4-H "bring your idea, we'll kill it honestly" — the evidence factory productized; design by LANE-H , build ). See the origins notes.

Charter: the single canonical definition + current-state reference. Result numbers live ONLY in the strategy ledger; code + exact constants live in the rule lab code (grammar · BLOCKING wall · verdict law) · the rule lab executor research code (the gauntlet orchestrator) · the rule lab inbox code (queue + Review-Inbox producer). This page states the CONTRACT and links the rest.

One-line definition: a closed-vocabulary rule composer ("SELECT liquid500 WHERE not_extended RANK BY mom12 TAKE 25 HOLD quarterly") whose output is judged by the existing evidence factory — pre-registration, walk-forward halves, a random-selection placebo, realistic costs, a capacity read, the Nifty-500 net benchmark — and returned as a verdict in the strategy ledger's own vocabulary.

---

1. What it is

Testing an idea used to mean hand-writing a study module. The Rule lab makes falsification a product surface: compose a rule from a fixed vocabulary and the factory runs it through the same gauntlet as house research. Its value is not that it finds winners — it is that it makes falsification cheap and it refuses to let a known-dead idea be re-walked silently: any rule matching a recorded failure shape carries the matching ❌ BLOCKING row(s) from the strategy ledger, verbatim, stapled to the result before anything runs.

2. The grammar (closed vocabulary — the Pat pattern)

RULE := SELECT <universe> [WHERE <filter> [AND <filter>]*]
        RANK BY <signal> TAKE <n:5..50> HOLD <horizon> [VETO <veto> [AND <veto>]*]

Every token binds 1:1 to a function that already exists and is already tested (the rule lab code declares the bindings; the suite proves each callable exists). Universes are PIT liquidity-percentile bands (never rupee constants). Deliberately inexpressible: arithmetic composers (weight-tuning is the overfitting engine the ledger exists to prevent), rupee thresholds, entry/exit timing micro-language (the measured exit law says looser-is-better), and single-stock verdicts. Promoting a veto-layer term (Credibility, Accumulation) to a ranker is a compile-time refusal carrying its ledger citation .

3. The gauntlet (reuse, not new math)

Pre-register (sha256 of the frozen rule text, first registration wins, rule_lab_prereg) → PIT tables with the -F1 one-day execution lag → factory.run_strat walk-forward halves 2012-18 / 2019-26 → random-selection placebo (observed must beat the null p95) → participation-real costs (net is reported first) → capacity breakpoint (no stated capacity, no result) → Nifty-500 buy-and-hold net benchmark → verdict. A missing stage is a refusal to rule (NO-VERDICT), never a partial verdict.

4. Verdict vocabulary (the ledger's, verbatim)

REJECTED · WEAKER-THAN-BENCHMARK · CONDITIONAL(<condition>) · NEW-BENCHMARK · NO-VERDICT(<why>), each carrying a permanent qualifier (descriptive-only / paper-only / flat-cost-only / fundable). NEW-BENCHMARK is machine-impossible without net-beats-benchmark in BOTH halves AND a stated capacity AND observed > placebo p95. Decided verdicts land in the Review Inbox (kind='rule_verdict') — the ledger is canon and canon carries a human signature, so nothing auto-appends.

5. Boundary (SEBI posture — plan §3)

User-directed analysis is analytics, not advice: the user composes the rule; the machine reports cohort arithmetic (SEBI RA Regulations 2014 reg. 2(1)(w) — statistical summaries / user-criteria tools). No single-stock verdicts, no action verbs, no target prices; verdict labels are ledger vocabulary only. v1 is personal-first: the composer is owner-gated, anonymous visitors see a clearly-labeled synthetic demo.

6. Operations

POST /dash/rule-lab/run only queues (rule_lab_queue); the owner drains it with the research-venv CLI — python -m explosive_moves.rule_lab_executor --work — and the verdict lands on the page and in the inbox. Deliberately no timer (AUD-95); the runner is a human-invoked CLI. Glossary: *Placebo p95 · Capacity breakpoint · Both halves · Prereg gate hash · Flat cost only* (the metrics glossary § Rule lab).