🧪 Feature-Realness Instrument-Grounding Loop

Before a promoted feature can be called real or useful, the instrument that judges it must itself be proven on data where the right answer is known. This bounded loop grounds the 20 candidate instruments (#0–#19) — one per iteration, in dependency order — then STOPS. Every iteration emits one HTML page, appended to the ledger at the bottom of this index.

3 / 20

instruments grounded

discovered (candidates)

0 / 30

columns with grounded evidence

P2 ▸

#2 grounded · #3 next

What this loop does

Each candidate instrument sits an entrance exam: the §4 five-control battery (positive / negative-null / substitution / power-calibration / invariance-leakage), plus its instrument-specific Harden item and a published detection envelope. It ends with a terminal verdict:

DISCOVERED → UNDER EVALUATION → GROUNDED (or REJECTED / UNVERIFIABLE-PARK)

Only a GROUNDED instrument may later judge a real feature. The graduation tally is the loop's STOP condition: when all 20 carry a terminal verdict, the loop publishes a completion page and ends.

Substrate gate. Instruments #1 (future-perturbation invariance) and #0 (CPCV purge/embargo) ground first. If either fails to ground, the campaign HALTS before the realness/utility instruments — every downstream out-of-sample number rides on the substrate.

Hard operating rules

Real-data-only (operator ruling 2026-07-17, re-confirmed 2026-07-22 — directive 03c governs): no synthetic data. Controls are deterministic transforms of real series — look-ahead-shift (known-positive), block-permute (known-null), exact-clone (known-duplicate), subsample (power-vs-N). An archetype with no clean real exemplar → UNVERIFIABLE-PARK.
Hard-capped, read-only ClickHouse: every connection readonly=2, SELECT only, both substrates.
5 CPU cores / 5 GB RAM / no swap: every compute step runs inside a systemd-run --user --scope cap on bigblack. Cap-kills are evidence; the cap is never raised.
Laptop drives, bigblack computes: the /loop orchestrates on the laptop and dispatches every read + compute step to bigblack over SSH. Published to /home/nasimubd/sites.

Dependency order (frozen)

Phase	Instruments (in order)
SEAL	real-data-only control substrate self-validation (first firing)
Substrate	#1 future-perturbation invariance → #0 CPCV purge/embargo (both GROUNDED or HALT)
Realness	#2 block-perm null → #3 effective-n+FDR → #6 SFI → #8 DSR → #9 PBO → #10 Harvey-Liu-Zhu
Usefulness	#4 Rank IC → #5 IC-decay → #7 MDA → #19 quantile monotonicity
Conditional	#11 CMI → #12 knockoffs → #13 DML
Increment / mechanism	#14 spanning → #15 mechanism-intensity
Robustness / detector	#16 structural-break → #17 plateau → #18 detector lead-lag

Full per-instrument Admit / Graduate / Harden rows: METRIC-EVALUATION-FRAMEWORK.md §7. Metric tables twin: The Realness Question.

Next iteration

#3 — effective-n deflation + BH-FDR (realness / multiple-testing). #2 grounded (iter 5) — the first realness instrument is cleared. Next: ground #3 — the HAC long-run-variance effective-n estimator + Benjamini-Yekutieli FDR (autocorrelated data has fewer independent points than it looks; count them honestly, then correct for many tries), with the mandatory GARCH heavy-tail null (the archetype the scalar effective-n deflator fails) and iid/AR1/white/block null calibration. Then #6 → #8 → #9 → #10.

Iteration ledger (append-only)

One row per firing. The loop appends below the marker; rows are never edited or deleted (corrections are new rows with supersedes:). Each row links the iteration's own HTML page.

#	Date	Instrument / step	Verdict	Page
1	2026-07-22	SEAL · real-data-only control-substrate self-validation	SEAL-PASS lab open	iter 1 →
2	2026-07-22	#1 · future-perturbation invariance (causality floor)	GROUNDED	iter 2 →
3	2026-07-22	#0 · CPCV purge/embargo (leakage-free partition)	CHECKPOINT 6/7 · PBO open	iter 3 →
4	2026-07-22	#0 · CPCV purge/embargo (supersedes iter 3) — ★ substrate gate PASS	GROUNDED	iter 4 →
5	2026-07-22	#2 · block-permutation shuffled-label null (realness) · P2	GROUNDED	iter 5 →