Iteration 10 — DRO diagnostics + the anchor triad: 8 STABLE entries, 67 total

Iteration 10 — DRO diagnostics + the anchor triad 8 STABLE HARVESTED · 67 TOTAL FAMILIES 12 + 20 SWEPT

2026-07-07 · tenth firing · ledger row 15 · harvest batch F12+20-01 · campaign branch loop/matrix-harvest-campaign

In plain words: back to the hungriest boundary with two complementary hauls. First, robustness bounds: instruments that don't just ask "is this relationship stable?" but "how much shift can it survive?" — worst-regime-versus-average gaps, worst-subpopulation risk at a tunable tail, and certificates that put a number on the safety margin (finally giving the campaign's controlled-imperfection theory line a set of implemented tools). One entry carries an honest warning: its readout is a raw score of exactly the kind a previous candidate was thrown out for, so its screening will hinge on a permutation check we've already specified. Second, the anchor triad: three econometric classics (1949, 1982, 1997 — all in maintained modern packages) that guard the guards. Invariance evidence leans on anchor variables; these test whether the anchors are valid, whether they're strong enough to be informative, and how to keep inference honest even when they're weak. Together with the kernel-based guard already catalogued, the anchor-failure surface is now covered from every side. One near-duplicate was refused as padding, on the record.

The eight

id	Instrument	Family	What it adds to STABLE
H-060	GroupDRO worst-group gap (MIT, 294★)	12	worst-regime vs average-regime loss gap — regime-dependence made readable
H-061	CVaR/χ²-DRO bounds (MIT)	12	worst-subpopulation risk at tunable tail mass — heavy-tail-native framing
H-062	Wasserstein-DRO certificates (rsome)	12	a NUMBER on how much shift a conclusion survives — the ε-frontier gains tooling
H-063	V-REx risk-variance	12	variance of per-regime risks — flagged: raw-score risk, permutation null required
H-064	Multi-domain calibration gap	12	calibration-as-invariance — a falsification channel nothing else covers
H-065	Sargan–Hansen J (linearmodels/AER)	20	are the anchors VALID?
H-066	Staiger–Stock first-stage F	20	are the anchors STRONG enough to be informative?
H-067	Anderson–Rubin robust test (1949)	20	inference that stays honest even when anchors are weak

Instrument

Family

What it adds to STABLE

H-060

GroupDRO worst-group gap (MIT, 294★)

worst-regime vs average-regime loss gap — regime-dependence made readable

H-061

CVaR/χ²-DRO bounds (MIT)

worst-subpopulation risk at tunable tail mass — heavy-tail-native framing

H-062

Wasserstein-DRO certificates (rsome)

a NUMBER on how much shift a conclusion survives — the ε-frontier gains tooling

H-063

V-REx risk-variance

variance of per-regime risks — flagged: raw-score risk, permutation null required

H-064

Multi-domain calibration gap

calibration-as-invariance — a falsification channel nothing else covers

H-065

Sargan–Hansen J (linearmodels/AER)

are the anchors VALID?

H-066

Staiger–Stock first-stage F

are the anchors STRONG enough to be informative?

H-067

Anderson–Rubin robust test (1949)

inference that stays honest even when anchors are weak

Padding refused: Kleibergen–Paap (same slot as H-066, marginal delta). Out of scope: WILDS-style benchmarks (datasets, not instruments). Families 12 and 20 are now swept.

Coverage counters after this firing

Boundary	Minimum	Verified	Screened-pass	Fed	Starvation
STABLE	100	33	10	10 (B-01)	67
FRAGILE↔CONDITIONAL	60	6	0	0	54
CONDITIONAL-scoping	40	3	0	0	37
FAIL	40	13	10	10 (B-02)	27
EXPIRY	40	10	0	0	30
cross-cutting	20	2	0	0	18
Total	300	67	20	20	233

Boundary

Minimum

Verified

Screened-pass

Fed

Starvation

STABLE

100

10 (B-01)

FRAGILE↔CONDITIONAL

CONDITIONAL-scoping

FAIL

10 (B-02)

EXPIRY

cross-cutting

Total

300

233

▶ Next iteration

Iteration 11 = screen batch S-03 — STABLE round two. Thirteen unscreened STABLE instruments have accumulated since S-01 (the four causal-discovery machines from the import, universal inference, safe testing, and this firing's eight) — they face M0/M1 next; ≥10 passers triggers feed batch B-03. Alternative slices if steered: the EXPIRY screen (10 verified, cohort ready) or FRAGILE↔COND harvesting (family 9/13/19 leads + local Gaussian correlation).

Steering point: no directive = STABLE screen S-03 next firing. Directives are ledger-recorded verbatim before taking effect.

Machine-readable record: ledger row 15 + catalog §F12+20-01. WebSearch: blocked this firing (statistics-phrased; pattern recorded). Pages are append-only.