Iteration 47 — H-004 PCMCI+ EXCLUDED: the α Dial Reshuffles Its Own Role Readout

Iteration 47 — H-004 PCMCI+ EXCLUDED KILL — α RESHUFFLES THE ROLE READOUT B-01 CLOSED: 6 ADMITS · 3 KILLS

In plain words: the causal-map drawer that aced its exam failed the field test in the subtlest way yet. Its job here would have been to rank features by how much their causal wiring CHANGES between regimes — that ranking is its entire reason to exist on our shelf. The field test asked one devastating question: does that ranking survive a small turn of its sensitivity dial? No — same data, two reasonable dial settings, and the rankings agree only 62.5% (our bar is 90%). An instrument whose verdicts depend that much on an arbitrary knob cannot certify anything, so it is out — with the repair path on record (the library itself offers a dial-free mode; that would be a new, admissible candidate). And with this, the first batch is complete: ten candidates in — six certificates out (including the once-vacant STABLE slot, now six instruments deep on forex), three honorable kills, every verdict with its evidence on file. The second batch opens next.

Preflight (A0): 2026-07-10 02:45 — load1 3.54/32c (≤24) · one 0.5-core python proc (under the bar) · 36 GiB avail, si/so=0 · CH + sidecar + kintsugi active → ALL PASS. Sync: origin/main unchanged. One capped run, 1.4 min, readonly=2 loader, tigramite via ephemeral uv, real rows only (identical strided rows under both dial settings).

Census results (pre-registered rule in pcmci_m3_census.py)

Check	Reading	Ruling
Dial check (α = 0.01 ↔ 0.05, identical rows)	`struct_instab` (the role readout — mean Jaccard distance of per-regime discovered link-sets): Spearman 0.625 ✗ (bar 0.90) · `link_rate`: 0.955 ✓ (but not the role readout)	RULE 1 FIRES — α-fragile
Degeneracy	struct_instab q10/50/90 = 0 / 0.41 / 0.71 — real spread (regime-dependent wiring is everywhere); link_rate median 1.25	not degenerate — but untrustworthy while α reshuffles it
M3 vs stack (moot, recorded)	Worst \|ρ\| = 0.493 — incl. vs the newly admitted Causal Dantzig d_min (0.287/0.136): first census tested against the row-56 instrument	below bar — irrelevant given rule 1
Exam-vs-census contrast	Row 57's exam-scale dial invariance did NOT generalize to census scale — the census-scale dial check earned its keep (the H-021 lesson, repeated from the other direction)	MEASUREMENT ATTACKED & TRUSTED
Verdict (row 58)	EXCLUDE(α-fragile at census scale — struct_instab dial correlation 0.625 < 0.90; unexplained load-bearing dial on the role readout) → BONEYARD, with the re-proposal path on record: tigramite's `pc_alpha=None` internal model selection (or an α-integrated/FDR link readout) = a new candidate with the dial derived away

Check

Reading

Ruling

Dial check (α = 0.01 ↔ 0.05, identical rows)

struct_instab (the role readout — mean Jaccard distance of per-regime discovered link-sets): Spearman 0.625 ✗ (bar 0.90) · link_rate: 0.955 ✓ (but not the role readout)

RULE 1 FIRES — α-fragile

Degeneracy

struct_instab q10/50/90 = 0 / 0.41 / 0.71 — real spread (regime-dependent wiring is everywhere); link_rate median 1.25

not degenerate — but untrustworthy while α reshuffles it

M3 vs stack (moot, recorded)

Worst |ρ| = 0.493 — incl. vs the newly admitted Causal Dantzig d_min (0.287/0.136): first census tested against the row-56 instrument

below bar — irrelevant given rule 1

Exam-vs-census contrast

Row 57's exam-scale dial invariance did NOT generalize to census scale — the census-scale dial check earned its keep (the H-021 lesson, repeated from the other direction)

MEASUREMENT ATTACKED & TRUSTED

Verdict (row 58)

EXCLUDE(α-fragile at census scale — struct_instab dial correlation 0.625 < 0.90; unexplained load-bearing dial on the role readout) → BONEYARD, with the re-proposal path on record: tigramite's pc_alpha=None internal model selection (or an α-integrated/FDR link readout) = a new candidate with the dial derived away

Batch B-01 — final tally (10/10 resolved)

B-01 (STABLE ceiling, 10 feed rows, cheapest-kill-first) VERDICT LEDGER H-001/H-002 e-value invariance family (= candidate A) ✅ ADMIT #3+#4 rows 34–35 (forex + crypto — the vacant STABLE slot FILLED) H-016 seqICP ✅ ADMIT #5 row 39 (env-free corroborator; crypto role excluded row 38) H-023 IAS ✅ ADMIT #6 row 41 (ancestry readout, scoped) H-019 IPP ✅ ADMIT #7 row 43 (distributional invariance, 27 shape-shifters) H-025 HSIC-X ✅ ADMIT #8 row 45 (anchor-validity guard) H-017 StabilizedRegression ❌ EXCLUDE row 51 (v5-oracle shadow, ρ=0.977; κ resolved first) H-021 EILLS ❌ EXCLUDE row 53 (empty-set collapse; ε-frontier witness #3) H-018 Causal Dantzig ✅ ADMIT #9 row 56 (hidden-confounder-tolerant meter; M4 AND M5′b) H-004 PCMCI+ ❌ EXCLUDE row 58 (α reshuffles the role readout, 0.625)

▶ Next iteration

Next firing: open batch B-03 (the second STABLE batch) — read its FRONTIER-FEED rows and the per-candidate dossiers in the harvest CANDIDATE-CATALOG, take the first candidate cheapest-kill-first, entrance exam on the byte-identical forex twin anchor. All standing lessons travel: oracle-shadow check, dials derived-or-swept at BOTH exam and census scale (the H-004 lesson: exam-scale invariance does not certify the dial), ε-frontier head-on. Standing for the operator: the 73 cycle-1 draft statuses (iters 11/13), the Frontier #4 expiry-monitor precondition review, and the ripening STABLE-declaration proposal (six certified forex instruments) — plus PR #599 now carries the full B-01 closure if a merge milestone feels due.

Iteration 47 · 2026-07-10 · EXCLUDE (row 58, BONEYARD +1) · B-01 CLOSED · capped (1.4 min) · readonly=2 · zero generated values · append-only · evidence: pcmci_m3_census.py · pcmci_m3_census_results.json

Iteration 47 — H-004 PCMCI+ EXCLUDED KILL — α RESHUFFLES THE ROLE READOUT B-01 CLOSED: 6 ADMITS · 3 KILLS

Census results (pre-registered rule in pcmci_m3_census.py)

Batch B-01 — final tally (10/10 resolved)

▶ Next iteration

Census results (pre-registered rule in `pcmci_m3_census.py`)