Iteration 1 — Family-1 invariance harvest: 6 verified entries

Iteration 1 — Family-1 invariance harvest HARVESTED 6 · CHECKPOINT

2026-07-04 · first firing · ledger rows 1–3 · campaign branch loop/matrix-harvest-campaign

In plain words: the question this whole research program cares about is "if a trading feature is genuinely 'its own thing' today, will it still be tomorrow, given that markets remember yesterday?" The branch of statistics built for exactly that question is called invariance testing — checking whether a relationship holds unchanged across different environments (for us: market regimes). This firing swept that family's literature and brought back six instruments, each verified end-to-end: the paper resolved to a real journal DOI via Crossref, and the open-source code checked on GitHub or CRAN (or an explicit "none exists" recorded, with the reason it still qualifies). Three more good candidates are waiting only because their venue isn't in Crossref's index and the arXiv channel was rate-limited all firing — they are recorded as leads, not padded into the catalog. That restraint is the point: every number on this dashboard is backed by a verification you can replay.

Bootstrap notes (first firing)

Event	Detail
Cadence	Operator launched at 10m (SSoT LOOP-PROMPT says 45m) — recorded as operator directive, ledger row 1
Branch bootstrap	Scaffold PR #581 still open → campaign branch cut from the scaffold tip (main lacks the campaign files); will rebase onto main when #581 lands and its squash SHA becomes the frozen rulebook hash — ledger row 2
Dashboard twin	Created this firing (this folder), registered in `build_probes.py`

Event

Detail

Cadence

Operator launched at 10m (SSoT LOOP-PROMPT says 45m) — recorded as operator directive, ledger row 1

Branch bootstrap

Scaffold PR #581 still open → campaign branch cut from the scaffold tip (main lacks the campaign files); will rebase onto main when #581 lands and its squash SHA becomes the frozen rulebook hash — ledger row 2

Dashboard twin

Created this firing (this folder), registered in build_probes.py

What entered the catalog (all STABLE boundary — the most starved: 9/100 at pick time)

id	Instrument	One-line story	Primary source (Crossref-verified)	FOSS (checked)
H-016	seqICP — sequential invariant causal prediction	Invariance testing built FOR time series — environments come from the time structure itself, no labels needed	JASA 2018 · DOI 10.1080/01621459.2018.1491403	CRAN `seqICP`
H-017	Stabilized Regression	Selects feature sets that are stable AND predictive across environments	Ann. Appl. Stat. 2021 · DOI 10.1214/21-aoas1487	CRAN GPL-3 + GitHub AGPL-3
H-018	Causal Dantzig	Invariance estimation that tolerates hidden confounders — an angle nothing in our stack covers	Ann. Stat. 2019 · DOI 10.1214/18-aos1732	NONE-FOUND; trivially implementable (closed-form) — stated per H0 §3
H-019	Invariant Probabilistic Prediction	Tests whether the WHOLE forecast distribution (not just the mean) is invariant — heavy-tail relevant	Biometrika 2024 · DOI 10.1093/biomet/asae063	GitHub `AlexanderHenzi/IPP` (no license file — noted)
H-020	Distribution-generalization framework	The possibility-side theory: exactly WHEN invariance transfers to unseen regimes (complements the Rosenfeld impossibility anchor H-010)	IEEE TPAMI 2022 · DOI 10.1109/tpami.2021.3094760	n/a (theory instrument; NILE code NONE-FOUND)
H-021	EILLS — environment-invariant least squares	Invariance as a penalty inside the regression objective, with a consistency proof	Ann. Stat. 2024 · DOI 10.1214/24-aos2435	GitHub `wmyw96/EILLS` (MIT)

Instrument

One-line story

Primary source (Crossref-verified)

FOSS (checked)

H-016

seqICP — sequential invariant causal prediction

Invariance testing built FOR time series — environments come from the time structure itself, no labels needed

JASA 2018 · DOI 10.1080/01621459.2018.1491403

CRAN seqICP

H-017

Stabilized Regression

Selects feature sets that are stable AND predictive across environments

Ann. Appl. Stat. 2021 · DOI 10.1214/21-aoas1487

CRAN GPL-3 + GitHub AGPL-3

H-018

Causal Dantzig

Invariance estimation that tolerates hidden confounders — an angle nothing in our stack covers

Ann. Stat. 2019 · DOI 10.1214/18-aos1732

NONE-FOUND; trivially implementable (closed-form) — stated per H0 §3

H-019

Invariant Probabilistic Prediction

Tests whether the WHOLE forecast distribution (not just the mean) is invariant — heavy-tail relevant

Biometrika 2024 · DOI 10.1093/biomet/asae063

GitHub AlexanderHenzi/IPP (no license file — noted)

H-020

Distribution-generalization framework

The possibility-side theory: exactly WHEN invariance transfers to unseen regimes (complements the Rosenfeld impossibility anchor H-010)

IEEE TPAMI 2022 · DOI 10.1109/tpami.2021.3094760

n/a (theory instrument; NILE code NONE-FOUND)

H-021

EILLS — environment-invariant least squares

Invariance as a penalty inside the regression objective, with a consistency proof

Ann. Stat. 2024 · DOI 10.1214/24-aos2435

GitHub wmyw96/EILLS (MIT)

Coverage counters after this firing (recomputed from the catalog)

Boundary	Minimum	Verified	Movement
STABLE	100	15	+6
FRAGILE↔CONDITIONAL	60	0	—
CONDITIONAL-scoping	40	1	—
FAIL	40	0	—
EXPIRY	40	4	—
cross-cutting	20	1	—
Total	300	21	+6

Boundary

Minimum

Verified

Movement

STABLE

100

FRAGILE↔CONDITIONAL

—

CONDITIONAL-scoping

—

FAIL

—

EXPIRY

—

cross-cutting

—

Total

300

Channel report (the honest part)

arXiv API ✗ HTTP 429 the ENTIRE firing (backoffs 10s / 45s — still throttled) Crossref REST ✓ 6/6 primary sources resolved to journal DOIs gh search/api ✓ FOSS checks (repos, stars, licenses, activity dates) CRAN pages ✓ seqICP · StabilizedRegression · (causalDantzig: 404 → NONE-FOUND) WebSearch ✗ filter-blocked for these topics (standing constraint, recorded 2026-07-03/04)

Leads held back (checked FOSS, unresolvable paper this firing — NOT catalogued): Active ICP (juangamella/aicp 21★) · Invariant Ancestry Search (PhillipMogensen/InvariantAncestrySearch) · EIIL (ecreager/eiil 52★). All are PMLR conference papers — Crossref does not index PMLR; they queue for the arXiv retry.

▶ Next iteration

Continue family 1 (checkpoint honored): (1) retry the arXiv API for the three PMLR leads + the H-015 Wasserstein-relaxed-invariance source; (2) if arXiv is still throttled, try the firecrawl fallback on littleblack; (3) continue the family-1 sweep — GROW e-statistics pipeline variants, e-value invariance extensions. After family 1 is swept, the starvation rule (H0 §2) picks the next slice: family 2 / 12 / 20 for STABLE, or the predecessor-import batch (~95 instruments, mostly FAIL/cross-cutting).

Steering point: type a directive into the loop session to redirect (e.g. "import batch first", "skip to family 7", "pause") — it is ledger-recorded verbatim before taking effect.

Machine-readable record: ledger rows 1–3 + catalog rows H-016–H-021 in findings/evolution/audits/2026-07-04-matrix-harvest-status-grounding-loop/. Pages are append-only; corrections arrive as new pages with supersedes pointers.