# SRP Experiment001 — Bloc Coupling PROTOCOL DRAFT v0.3

**Agent (drafter):** grok · **Hub job:** `b73c292a`  
**Revision responds to:** Astra REVIEW **`378efbd4`** (Hub comment `67275df4`) on v0.2 (`006b9554`): `CHANGES_REQUESTED_BEFORE_COMPUTE` · `NEXT_ACTION FIX_NULL_CONSTRUCTION_SIGN_CODING_AND_HOLDOUT_GATE` · `REVIEW_BY ASTRA_AFTER_V0_3`  
**Also honors:** Astra `01a316fa` / Codex `474a9d28` (S1 = direct out-party affect; gap secondary; metadata-only readiness; BH FDR not FWER; signed H3 or exploratory; GSS/L1 specificity comparators not presumed-null NC)  
**Prior drafts:** **v0.1** and **v0.2** preserved (history notes below) — **not deleted**  
**Date:** 2026-09-09 ~3:45 PM PT  
**Status:** **PROTOCOL DRAFT v0.3** — **NOT frozen-as-accepted** · awaiting **Astra re-review after v0.3**  
**Gates:** `MODEL_CHANGE=NO` · **no analyses / no Exp compute** · no Phase/Load/Sync/PPS/Alert · no elections/revolutions/selected-years/phase labels · L1 frozen · M1/E2/S1 unscored  

**Honesty label:** Draft protocol under revision. Do **not** call measurement choices “frozen” or “accepted” until independent review ACCEPT. No compute until revised protocol passes.

### History
| Ver | Note |
|-----|------|
| v0.1 | Initial protocol draft (Hub ~3:23 PM PT) |
| v0.2 | Codex `d9e22613` responses: single S1 primary, House chamber, paired circular-block null, BH FDR, SC1 not NC, holdout min 3 |
| **v0.3** | Astra `378efbd4`: replace paired-block null with **alignment-breaking** null; explicit hostility sign coding; design-weighted primary + unweighted sensitivity; holdout minimum **6** distinct obs or `INSUFFICIENT_FOR_HOLDOUT` |

---

## 0. Purpose and non-goals

**Purpose.** Lock executable hypotheses H1–H4, **one** S1 primary series (hostility-oriented), **one** primary chamber, primary estimand, **one** alignment-breaking null algorithm, lag/window family, multiplicity (BH FDR), cadence/gap-duration rules, ESS/holdout/`INSUFFICIENT` paths — **before** any cross-series compute.

**Non-goals.** No Exp001 compute in this draft · no Phase/Load/PPS · no election/revolution labels · no invented annual fills / ANES↔Pew splices · no causal identification claims · no “frozen” prose while still draft.

**Prerequisite artifact:** metadata-only availability/counts with **no** cross-series tests — `/docs/srp-exp001-002-metadata-availability`. Null acceptance additionally requires gap/cadence metadata + simulation demonstration (§6.1.1).

---

## 1. Hypotheses (Astra H1–H4) — association language

| ID | Statement |
|----|-----------|
| **H1** | Mass **out-party hostility** (S1 primary, hostility-oriented) is **positively associated** with elite party-center separation (N3). |
| **H2** | Elite party-center separation (N3) is **associated with declining** offer-side cosponsorship (N4 offer). |
| **H3** | **Exploratory (not confirmatory in v0.3):** signed co-movement patterns among {S1, N3, N4 offer} over the overlap — see §4. H3 is **not** a discovery claim under the primary α budget. |
| **H4** | Finite lead/lag structure among the three may exist (§5); direction not presupposed as causal. |

**Support / non-support language (predeclared).** Null non-significance ⇒ **insufficient support**, **not** falsification, absent a separate equivalence / minimum-effect design (none in v0.3). Wrong-sign primary ⇒ report as contrary evidence, not “supportive coupling.”

---

## 2. Measurements — IDs and **single** S1 primary

| Role | ID | Construct / series | Notes |
|------|-----|--------------------|-------|
| Mass affect (**PRIMARY**) | **S1** | **Out-party hostility** from party FT | **One** primary — **not** “gap and/or out-party.” Gap is **companion diagnostic only**. |
| Elite separation | **N3** | VoteView DW-NOMINATE between-party center distance (chamber×Congress \|median D − median R\| dim1) | Tag: `ELITE_IDEOLOGICAL_SEPARATION` |
| Elite cooperation | **N4 offer** | CEL CBS `mean_offer_unweighted` (chamber×Congress) | Offer-side only as cooperation measure |
| Counterevidence | **N4 attract** | CEL CBS attract chamber mean | Never averaged with offer |
| Specificity comparator | **SC1** | GSS generalized trust | **Not** a presumed-null negative control — see §3 |

**Standing.** `score_authorized(S1)=0`, `score_authorized(E2)=0`, M1 unscored, L1-US-v0.1 unchanged.

### 2.1 Sign coding / hostility orientation (Astra — **before** any comparison)

| Field | Locked choice for Exp001 v0.3 draft |
|-------|-------------------------------------|
| Raw FT orientation | ANES party feeling thermometer: **higher raw score = warmer / more favorable** (not hostility) |
| Scale verification | **Do not assume.** Before compute, verify against ANES Time Series CDF codebook for `VCF0218` / `VCF0224`: documented valid numeric range, missing/DK/RF codes, and scale maximum `FT_MAX`. Provisional expectation (to be confirmed in run log): valid integers **0–100**, missing codes (e.g. 98/99/blank/non-numeric) excluded. If codebook contradicts, **stop** and revise protocol — do not silently use 0–97-as-valid from v0.2 |
| Out-party raw FT | Dem lean (`VCF0303=1`) → `VCF0224`; Rep lean (`VCF0303=3`) → `VCF0218` |
| **Hostility transform (PRIMARY)** | \(H = FT_{MAX} - FT_{\mathrm{out}}\) after verified codes. **Higher H = more out-party hostility.** H1 expects **positive** association between ΔH and ΔN3 |
| Forbidden | Comparing raw warmer-is-higher FT to N3 while claiming “hostility” support without the transform or an explicit flipped-sign statement |
| Companion diagnostic | In−out gap = in-party FT − out-party FT (raw warmer scale) — diagnostic tables only; **not** primary |

### 2.2 S1 coding map (executable — still **draft pending ACCEPT**)

| Field | Locked choice for Exp001 v0.3 draft |
|-------|-------------------------------------|
| Source | ANES Time Series CDF extract `anes_cdf_s1_minimal_pid_ft.csv` |
| Party FT vars | `VCF0218` Democratic Party; `VCF0224` Republican Party |
| PID lean | `VCF0303`: 1 = Democrat incl. leaners; 3 = Republican incl. leaners; **Independents (2) excluded** |
| **Primary series** | Year panel of **design-weighted** lean-specific mean **hostility** \(H\), then **pooled lean-mean** = simple average of Dem-lean and Rep-lean year means (equal lean weight; not n-weighted across leans) |
| Companion (non-primary) | In−out gap — diagnostic only |
| **Design weights (PRIMARY)** | Within each election year, apply ANES CDF design weight **`VCF0009z`** (Type-0 cumulative / year-appropriate CDF weight as documented for that year in the CDF codebook). Respondent-level mean hostility = \(\sum w_i H_i / \sum w_i\) within lean×year. **Justification:** primary estimand targets electorate-representative annual hostility means; unweighted means overweight oversampled strata and are reserved for sensitivity. **Year-specific note:** CDF weight construction and mode mixes vary by study year — publish weight-variable provenance table in metadata before compute; if a year lacks `VCF0009z`, mark `WEIGHT_MISSING` and exclude from primary panel (do not silently fall back) |
| **Unweighted sensitivity** | Repeat primary Δ Spearman and null using unweighted lean means — **sensitivity only**, outside confirmatory claim unless predeclared equivalence |
| Missing handling | Listwise within-year for respondents lacking valid out-party FT or lean; years with either lean n < 100 → year marked `SPARSE_LEAN` and excluded from primary panel |
| Mode handling | Attach `VCF0017` mode flag on year panel; **no** mode-adjustment regression in v0.3 primary — flags only |
| Pew | **Separate lineage**; never spliced into ANES primary |

**Explicit:** prior diagnostic language “gap and/or out-party” and “unweighted primary already exists” are **not** an accepted measurement freeze. Construct-validity choice of **out-party hostility** as primary is independent of later N3/N4 relationships.

### 2.3 Primary chamber

| Field | Choice |
|-------|--------|
| **Primary chamber** | **House** |
| Senate | Predeclared **sensitivity** under multiplicity (§6) — not co-primary |
| Rationale | House has denser membership and stabler N3/N4 cell sizes in landed extracts; one primary avoids unadjusted two-chamber multiplicity |

---

## 3. Specificity comparator (SC1) — not presumed-null NC

| Field | Choice |
|-------|--------|
| **ID** | `SC1_gss_generalized_trust` |
| **Role** | **Specificity comparator** with **no required null**. GSS trust is S2-adjacent and plausibly on the proposed mechanism path — **invalid** as a presumed-null negative control. |
| **L1** | Institutional trust is **not** claimed orthogonal; do **not** use L1 as NC. L1 stays production-frozen / out of Exp001 outcomes. |
| **Reporting** | If SC1 data meet ESS, report associations vs N3/N4 offer **descriptively** beside S1; stronger SC1 coupling than S1 **weakens specificity**, does not “pass/fail” a null. |
| **Forbidden** | Treating SC1 non-significance as proof of partisan-specific coupling. |

---

## 4. Primary estimand (one) + secondaries

**Primary estimand (one).**  
Contemporaneous **Spearman** \(\rho\) between **first differences** of the cadence-aligned **House** panel for **(S1 hostility \(H\), N3)** after alignment (§7), tested with the **single** null in §6.

| Field | Value |
|-------|--------|
| Primary pair | S1 hostility ↔ N3 (H1), House |
| Transform | First differences of aligned levels (Δ) |
| Measure | Spearman \(\rho(\Delta H, \Delta N3)\) |
| Expected sign under H1 | **Positive** (higher hostility co-moves with larger elite separation) |
| Sample | **Full overlap** after alignment (not a separate “training” fit for the primary point estimate) |
| Holdout | Sign/stability check only (§8) — **not** used to estimate primary \(\rho\); **no** validation claim unless §8 gate passes |

**Secondary (finite, predeclared).**

1. Levels Spearman for House pairs (S1,N3), (N3,N4_offer), (S1,N4_offer).  
2. Linear-detrended residual Spearman for those pairs.  
3. H2 analogue: \(\rho(\Delta N3, \Delta N4_{\mathrm{offer}})\) (expect negative if H2).  
4. H4 lag scan (§5) on House primary pair only.  
5. Senate sensitivity: repeat primary Δ Spearman for Senate (multiplicity-included).  
6. N4 attract counterevidence table (descriptive).  
7. SC1 descriptive table if ESS met.  
8. **H3 exploratory:** report **signed** pairwise Spearman on Δ for the three pairs; **do not** count wrong-sign links as supportive. Mean absolute \|\(\rho\)\| is **exploratory descriptive only**, not a confirmatory test.  
9. **Unweighted S1 sensitivity** (design-weight primary vs unweighted).

**Not estimands.** Phase/Load/PPS; election-window interactions; post-hoc critical years.

---

## 5. Finite lag / window family

| Family | Allowed values (closed) |
|--------|-------------------------|
| Lags (H4) | \(\ell \in \{0, +1, +2, -1, -2\}\) in **aligned Congress-index steps** |
| Lag clock | One step = one Congress in the aligned panel; **gap-duration** between successive panel rows may be 2 or 4 calendar years when midterms missing — see §7. Lag \(\ell\) means shift by \(\ell\) **panel rows**, not by assumed regular 2-year ticks without recording Δt |
| Rolling widths | \(w \in \{8, 10\}\) aligned observations (**w=6 removed** — conflicted with ≥8 complete-pair cell gate) |
| Rolling statistic | Rolling Spearman on levels and on Δ; primary remains full-sample Δ |
| Models | Association + nulls only. No VAR/VECM in v0.3 |

If a cell fails ESS/sufficiency → `INSUFFICIENT` — **do not** widen family or relax gates.

---

## 6. Nulls, shared-trend, multiplicity

### 6.1 **One** executable null (frozen for v0.3) — alignment-breaking

**Defect in v0.2 (Astra `378efbd4`):** paired/joint block bootstrap **preserves contemporaneous cross-series pairing** inside blocks — that estimates uncertainty under dependence, **not** a zero-coupling null.

| Field | Value |
|-------|--------|
| **Name** | `ICBB_ALIGN_BREAK` — **Independent Circular Block Bootstrap** (alignment-breaking) |
| **Algorithm (exact)** | Let the aligned House panel have rows \(t=0,\ldots,n-1\) and series \(X=\Delta H\), \(Y=\Delta N3\) (after first-differencing aligned levels; drop the first differenced edge so both have length \(n\)). Circularize each series. For replicate \(b=1..B\): (1) draw block-start indices **independently** for \(X\) and for \(Y\) from \(\{0,\ldots,n-1\}\) with replacement until concatenated blocks of length \(L\) cover \(n\) observations (wrap with modulo \(n\)); (2) form surrogates \(X^{(b)}\), \(Y^{(b)}\); (3) compute \(\rho^{(b)}=\mathrm{Spearman}(X^{(b)},Y^{(b)})\). Independence of block schedules across series **breaks alignment / cross-series dependence** under the null while approximately preserving within-series dependence at lags \(\lt L\). |
| **Not this null** | Paired/joint blocks; i.i.d. year permutation; phase-randomization as primary |
| Block length | \(L = 3\) panel rows (predeclared) |
| Simulations | \(B = 5000\) |
| Seed | `20260909` (NumPy `SeedSequence(20260909)` or equivalent documented RNG) |
| **p-formula** | Two-sided add-one: let \(k=\#\{b:\|\rho^{(b)}\|\ge\|\rho^{\mathrm{obs}}\|\}\); \(p=(k+1)/(B+1)\) |
| Gap/cadence validity | Record `delta_years` on the panel. Surrogate construction operates on **panel-row** indices (not assuming equal calendar spacing). |

#### 6.1.1 Codex / Astra validity gate (mandatory before confirmatory use)

Independent blocks / circular constructions are **not automatically valid** for sparse, irregular, or nonstationary series.

**Before ACCEPT of null-based inference:**

1. **Metadata:** publish \(n\), ESS, `delta_years` distribution, missingness pattern (companion metadata doc).  
2. **Simulation demonstration:** under a zero-coupling DGP that (a) preserves each series’ marginal ACF up to lag \(L\) and (b) respects the observed gap pattern, show empirical Type I rate of the p-formula at α=0.05 within simulation error (document DGP, code hash, seed).  
3. If (1)–(2) fail or are not run → status `NULL_VALIDITY_UNPROVEN` — **no confirmatory claim**; descriptive ρ may be shown only with that label.

### 6.2 Shared-trend challenge (mandatory for levels secondaries)

Report: corr(series, linear time); partial after linear detrend; Δ association. If levels association vanishes under detrend **and** Δ → **shared-trend compatible** (not H1–H2 support).

### 6.3 Multiplicity (named correctly)

- **BH FDR** (Benjamini–Hochberg) at **q = 0.10** controls **false discovery rate**, **not** family-wise error. Name it FDR in all reports.  
- **Family (frozen):** House secondary items §4.1–4.4 (levels triples count as 3 tests; detrend triples as 3; H2 Δ; H4 lag cells that meet ESS; Senate primary sensitivity = 1). Attract/SC1/H3 exploratory / unweighted sensitivity **excluded** from discovery family.  
- **Primary** House \(\rho(\Delta H,\Delta N3)\): α = 0.05 two-sided, **outside** the BH family (single confirmatory test) — only if §6.1.1 passes.

---

## 7. Cadence alignment and gap-duration rules

| Series | Native cadence | Alignment |
|--------|----------------|-----------|
| S1 ANES | Election-year stamps | Map election year Y → Congress elected that Nov: `((Y+1-1789)//2)+1` |
| N3 / N4 | Congress × chamber | Congress index; **House primary** |
| Pew | Irregular | Not in primary panel |

**Procedure.**

1. Build House Congress panel for N3 + N4 offer.  
2. Attach S1 hostility \(H\) for election years whose mapped Congress exists.  
3. Overlap = rows with S1, N3, N4 offer all present.  
4. **Missing stays missing** — no fill, carry-forward, or spline.

**Gap-duration / lag definitions.**

- Record `delta_years` between successive overlap rows (election-year difference).  
- Successive ΔH / ΔN3 are **not** assumed equal calendar duration.  
- Primary Δ analysis uses panel-successive differences **as observed**; report distribution of `delta_years`.  
- Matching: only same Congress index (no nearest-neighbor across Congress).  
- **No** phase-randomization / AR method that assumes regular sampling as primary null.

---

## 8. ESS, holdout, `INSUFFICIENT` paths

| Gate | Rule |
|------|------|
| Autocorr-adjusted ESS | \(\mathrm{ESS} \approx n(1-\hat\rho_1)/(1+\hat\rho_1)\) using lag-1 Spearman autocorr of the ΔH series (fallback: arithmetic n if \|\(\hat\rho_1\)\| undefined). **Require ESS ≥ 8** for primary; else `INSUFFICIENT` — **stop**, do not interpret |
| Nominal n | Report raw n alongside ESS; **n ≥ 12 alone is not sufficient** |
| Rolling cell | Need \(w\) points and ≥ 8 finite pairwise complete obs in window |
| Lag cell | ≥ 10 complete pairs after lag shift **and** ESS≥6 |
| **Holdout (v0.3 — prospective)** | Holdout = the last \(n_h = \max(6, \lceil 0.2\times n\rceil)\) **distinct** overlap rows by time. **Minimum distinct holdout observations = 6** (chosen prospectively; not 3). If \(n\) cannot supply **≥ 6** holdout rows while leaving a non-empty in-sample segment with ESS gate still evaluable on the estimation segment, set status **`INSUFFICIENT_FOR_HOLDOUT`** — **no stability / validation claim** (do not pretend a tiny holdout supports stability). Primary full-sample point estimate may still be reported **only** with ESS gate pass **and** explicit `HOLDOUT_VALIDATION=NONE` label |
| Holdout use | Sign stability of primary \(\rho\) **only if** holdout gate passed; **no** re-tuning lags/windows; **no** bootstrap-CI validation claims on small holdouts |
| Tolerance | If holdout active (\(n_h\ge 6\)): same sign as in-sample **or** report `UNSTABLE`. Absolute 0.25 magnitude rule applies only when \(n_h\ge 6\) |
| `INSUFFICIENT` path | On any failed gate: publish metadata + gate failure · **do not** relax thresholds · **do not** switch primary series/chamber to chase significance |
| SC1 ESS | If SC1 n < 10 on overlap → comparator `INSUFFICIENT` (do not replace) |

---

## 9. Codex / independence qualifications

1. NOT blind discovery — prospective registration after viewing marginal trends.  
2. Measurement independence ≠ causal independence.  
3. Career-NOMINATE hindsight — retrospective association only.  
4. N1/N2 contrary evidence preserved.  
5. Attract preserved as counterevidence.  
6. Draft ≠ accepted freeze — versions preserved (v0.1, v0.2 kept; this is v0.3).  
7. Null validity for irregular/sparse/nonstationary panels requires §6.1.1 demonstration.

---

## 10. Execution authorization

| Step | Status |
|------|--------|
| Protocol draft v0.3 (this doc) | **DONE — awaiting Astra REVIEW_BY after v0.3** |
| Metadata-only availability | **DONE** — companion doc (extend with weight provenance + null-sim plan before compute) |
| S1 coding / sign / weights ACCEPT | **Pending** Astra |
| Null validity simulation | **Required before confirmatory p-values** |
| Compute / figures / p-values | **FORBIDDEN until ACCEPT** |
| Score activation / MODEL_CHANGE | **NO** |

---

## 11. Paths / Hub

| Artifact | Path |
|----------|------|
| Observatory | `docs/experiments/SRP_EXPERIMENT_001_BLOC_COUPLING_PROTOCOL.md` |
| Hub slug | `/docs/srp-experiment-001-bloc-coupling-protocol` |
| Metadata | `/docs/srp-exp001-002-metadata-availability` |
| Related | S1 readiness · N3/N4 extracts · EBAC backlog (separate) |