# Frozen Analysis Specification v3

**Issued 2026-08-30. Binds `PREDICTIONS_v5_2026-08-24.md` and every prediction in it.
Supersedes v2, issued earlier the same day. v1 and v2 are preserved verbatim and are not
edited; both remain in the timestamped commitment. Thresholds below are frozen literals —
they are the claim. Basis figures are measurements, computed by `scripts/figures.py` at
issue and stated "as at issue".**

**v3 changes no scoring map, no threshold, no population and no confidence.** It corrects
two false statements v2 froze, restores one publication commitment v2 silently dropped,
codifies one already-adopted evidentiary safeguard that currently binds zero rows, and —
the reason it exists at all — carries the changelogs this document series was always
supposed to carry.

---

## Why there is a v3

Three failures, all v2's, all found by independent review on the day v2 was issued.

**1. v2 froze a figure its own register has since corrected.** §1 A5 said "A1a′ holding
implies A5 holding in 20,000 of 20,000 bootstrap resamples." 20,000 is the number of
resamples *drawn*. A1a′ holds in 14,048 of them; A5 holds in every one of those 14,048.
The implication is unchanged at 100% and the figure was still wrong — it was typed, not
computed, in the document that exists to bind computation. Found by an independent
reviewer; the corrected figure is computed by `figures.py::dependence()`.

**2. v2's §5.5 case law misdates the event it cites as precedent.** It says the
no-withdrawal rule "was tested on 2026-08-24." A5 was registered on 2026-08-23 and its
characterisation was found wrong *the same day*; the ruling recording the retraction
issued with register v5 on 2026-08-24. A case-law paragraph that cannot date its own
case is not case law. Found by an independent fact-checker.

**3. v2 changed rules from v1 without a changelog — the exact practice this series
exists to prevent.** Five differences between v1 and v2 were never enumerated anywhere;
they were found by independent review, declared on the published pages on 2026-08-30
under the operator's ruling ("declare it, score as-is"), and are now recorded here in
Changelog A. Under that ruling **the v2 rules stand and score**; v3 does not flip them
back, because a second silent-then-corrected change is not better than a first. What v3
restores is the one item that was a publication *commitment* rather than a scoring rule
— the Murphy decomposition — which the build had already reinstated in code.

---

## Changelog A — what v2 changed from v1 without saying so

Recorded retrospectively. None of these is re-litigated; all were declared on the pages
on 2026-08-30 and score as v2 states them.

1. **G1's population** moved from "incumbents and challengers across the extension set"
   (v1) to "the roster" (v2). This is the change declared at length on G1's card: it
   turned a bet about data *beyond* the roster into a bet about data in hand, and a hit
   on G1 is therefore a drafting artifact, not a forecast. Declared 2026-08-30; scored
   as v2 says, and the card says a hit here is worth nothing.
2. **D1's event definition narrowed and its unit changed.** v1 counted *districts*
   materially changed by "withdrawal, replacement nominee, party switch, campaign
   suspension, death, or disqualification" dated 2026-08-22 to 2026-11-03. v2 counts
   *races with a material change* — "withdrawal, replacement nominee, party switch, or
   suspension — after 2026-08-21," silently dropping death and disqualification from the
   list. v2's list governs. The narrowing runs *against* the prediction (fewer
   qualifying events), which is the direction that makes scoring it as-written honest.
3. **E1's establishability rule loosened in wording.** v1 pinned "establishable" to rule
   G4 of `EVIDENTIARY_STANDARD_v1.1.md` and stated the denominator ("of the seven"); v2
   says "meeting the frozen evidentiary standard, not appearing on a data broker" with
   no denominator. For scoring, v2's map is applied with v1's G4 definition read in —
   the two are consistent, v2 is merely vaguer, and vagueness resolves to the stricter
   published rule.
4. **F2's Murphy decomposition was dropped.** v1 committed to publishing the Brier
   score's Murphy decomposition (reliability − resolution + uncertainty); v2 committed
   only to the Brier score and â. Nothing in the record explains the deletion; the build
   has carried a Murphy self-test throughout (`build_all.py` step 3a4). **Restored in §4
   below.** A publication commitment is additive — restoring it changes no score.
5. **§5.5's text changed** from "No prediction is withdrawn after 2026-08-22" (v1) to
   "after registration" (v2) — necessary, since six predictions were registered
   2026-08-23 — and gained the case-law paragraph whose date defect item 2 above
   corrects.

---

## Changelog B — v2 → v3

1. **§1 A5:** the "20,000 of 20,000" sentence replaced with the computed figures
   (14,048 of 20,000; A5 in all 14,048). No scoring change.
2. **§5.5:** the case-law date corrected — found wrong 2026-08-23, the day of
   registration; recorded in register v5 on 2026-08-24. No scoring change.
3. **§4 F2:** the Murphy decomposition commitment restored (Changelog A, item 4).
   Additive publication commitment; no scoring change.
4. **§4 G2:** basis restated on the coding as corrected 2026-08-30 (a settled
   personal-capacity civil suit had been coded "None found" against the index's own
   comparably-coded rows; the correction is declared on G2's card and in the Source
   Log): **6/56 R vs 7/57 D, p = 1.0000, OR = 0.857**. v2's printed basis (5/56,
   p = 0.7616) predated that same-day coding correction. The scoring map is untouched
   and the verdict region is unchanged — the correction moved the p-value *further into*
   the region G2 bets on, which is declared on the card in exactly those terms.
5. **§4 G2, evidentiary safeguard codified:** a row whose adverse coded value the index
   has downgraded to "Not established" (sourcing the index cannot stand behind, per the
   adverse-downgrade rule adopted 2026-08-30 on independent legal review) is excluded
   from G2's numerator and denominator, and both bases are published whenever the
   exclusion binds anything. **As at issue it binds zero rows** — the two rows it
   briefly covered were verified to named contemporary outlets and court records the
   same day and returned to the basis. This is the one rule v3 adds; it is a data-quality
   floor, not a threshold, and its provenance and its currently-vacuous status are both
   stated so nobody discovers it in October.
6. **§4 E2:** the flagged case is now named in the spec itself — it always was named in
   register v1, the naming was lost in the v2 rebuild of the card, and a spec a third
   party can score must not depend on cross-reading. No scoring change.
7. **§1 A5, declared context:** the fitted two-coefficient model (step and slope) that
   answers the question A5 asked badly is referenced as published-unregistered-unscored.
   Informational; no scoring change.

---

## §0 Definitions binding all predictions

Everything in v1 §0 carries forward unchanged except where restated here. Restated in
full so this document stands alone.

**Universe.** The 59 US House districts fixed as a literal table in `scripts/build.py`.
Selection rule as printed there: 2024 House margin ≤ 10 **OR** 2024 presidential margin
≤ 10, from the Cook Political Report competitive list (ratings 2026-08-13, retrieved via
270toWin 2026-08-20). **Three districts do not satisfy that rule and are named
exceptions: OH-09 (presidential margin 10.5), TX-28 (10.3), TX-34 (10.1)** — added as
marquee races, House margins not comparable after redistricting. They are flagged in
`districts_59.csv`. The list is frozen and is not revisable.

**Dual-universe rule (register v5, Ruling 1), binding on every A-block prediction.**
Every A-block verdict is scored on the **as-published 59-district universe** — the
universe the bets were priced against. The **rule-compliant universe** (the 56 districts
that satisfy the printed rule) is computed by the same script and **published beside
every A-block verdict, unscored and labelled**. As at issue the two differ materially —
pooled r 0.5756 against 0.7116, incumbents-only
0.3108 against 0.4993 — which is precisely why the
choice is fixed here, in advance, rather than in October.

**Roster.** 115 candidates in `scripts/nominees.py`. Party composition
**57 D / 56 R / 2 IND**
(register v5, Ruling 3: Kevin Kiley left the Republican Party on 2026-03-09 and filed
CA-06 as an independent). This is the baseline C2 is scored against.

**Incumbent.** Tenure > 0, i.e. any prior elected House or Senate service. As at issue
n = 48 of the 112 candidates with a finance row.
**This group is not "sitting members"** — it contains one former member seeking return
and one sitting member whose FEC record still codes her as a challenger. Publication
says "candidates who have served".

**Tenure.** Cumulative elected House + Senate service in years through 2026-11-03, in
`AGE_TENURE`. Challengers and open-seat non-members are 0.0.

**Corporate/trade PAC receipts.** FEC bulk `itpas2.txt`, summing `transaction_amt` where
transaction type ∈ {24K, 24Z, 24R, 24C, 24F} and the filing committee's bucket is
Corporate PAC or Trade assoc PAC. Memo entries (`memo_cd == "X"`) excluded. Independent
expenditures are not in this sum.

**Committee bucket**, from `cm.txt`, applied in this order, frozen:
`dsgn == "D"` → Leadership PAC (never corporate) · `cmte_tp ∈ {Y,X,Z}` → Party ·
`org_tp ∈ {C,W}` → **Corporate** · `org_tp == "T"` → **Trade** · `L` → Labour ·
`M` → Membership · `V` → Cooperative · otherwise Other.

**Corporate/trade share of receipts** = `100 × (Corporate + Trade) / total_receipts`,
`total_receipts` being `weball26.txt` column 5, cycle-to-date. Not itemised-only.

**Viability filter.** A candidate committee is excluded if receipts < $25,000 **and**
disbursements < $25,000. **Declared gap:** the frozen CSV does not carry disbursements,
so only the receipts leg is enforced. This is stated rather than hidden, and is applied
identically in October.

**Missing finance rows.** As at issue 3 of the 115 have no row in the
frozen finance file: Jennifer Balkcom, Kevin Kiley, Matt Little. One has no federal committee; two have
committees the August bulk files misfiled, documented in the master coding. **They are
excluded from every A-block computation and named in the published excluded-list.** If
exclusions ever exceed 10% of incumbents, the affected prediction scores **MISS**, not
unscoreable.

**Data vintage.** FEC bulk data frozen 2026-08-21. **77 of
112** committees report coverage through 30 June 2026;
the full range runs 16 June 2026 to 18 August 2026. Publication says
this, not "through 30 June". October uses the Q3 bulk files downloaded once, on a date
recorded in the Source Log.

**Reading grade.** Flesch-Kincaid as implemented in `scripts/style_analysis.py`, **with
the terminal-punctuation correction** (register v5, Ruling 2): each sample is terminated
with a full stop before cell pooling. The shipped pipeline joined samples with a bare
newline, and 72 of 468 samples
(66 of them social) end without terminal punctuation,
so sentences merged across sample boundaries and social-cell grade inflated. Both values
are published with every B-block verdict; **the corrected pipeline is what scores.**

**Register cells.** `release` · `quote` (written attributed) · `spoken` (excluded from
written comparisons) · `post`.

**Veteran coding, restated unambiguously.** Each candidate is coded `Yes`, `No`, or
`Unclear` per §F of the evidentiary standard. **In every party comparison, candidates
coded `Unclear` are removed from BOTH the numerator and the denominator — they are not
counted as non-veterans — and their number and names are published with the result.**
As at issue: 27 Yes, 85 No,
3 Unclear.

**Adverse-downgrade rule (adopted 2026-08-30 on independent legal review; codified
Changelog B, item 5).** A row may carry an adverse coded legal value only where the tier
gate has verified its sourcing to a court record or a named independent outlet. Where it
has not, the coded value is published as "Not established — see detail", the full detail
is retained, and the row is excluded from every computed legal statistic, with both
bases published whenever the exclusion binds anything. As at issue it binds zero rows.

**Calibration fitting set.** The log-odds calibration shift **â** is estimated on the
**23 scored predictions excluding F2′ and F2″**. Those two are then scored against it.
This is what makes them predictions rather than a fixed-point equation, and it is not
revisable.

**Scoring date: 2026-10-31.**

**Rounding.** All test statistics rounded to 2 decimals **before** comparison to a
threshold, **except** where a scoring map states a Fisher-z criterion, which is applied
at full precision because rounding an r-band edge can place it outside its own rule.
A value landing exactly on a boundary counts as **satisfying** the claim.

---

## §1 The money predictions

Common to all: population as defined in §0, unit = candidate, Y = corporate/trade share
of receipts, X = tenure in years, no covariates unless stated, **no outlier removal ever,
under any justification**. The highest-leverage point (Marcy Kaptur,
44 years, 8.41%) is named in publication
and retained.

### A1a — frozen, known bad, scored anyway
- **Scoring map:** incumbents-only Pearson `r ≥ 0.60` → HIT, else MISS. No other outcome.
- **Basis at registration:** 0.70 (wrong — see the defect report). **As at issue:
  0.3108.** Registered at 92%; expected MISS.

### A1b — frozen, known bad, scored anyway
- **Scoring map:** HIT iff `0.70 ≤ r_pooled ≤ 0.90` **AND** r_DEM > 0 **AND** r_REP > 0.
- **As at issue:** pooled 0.5756; DEM 0.5183, REP
  0.6830. Registered at 88%; expected MISS.

### A1a′ — PRIMARY
- **Population:** incumbents (tenure > 0) passing the viability filter.
- **Estimator:** Pearson r. Also reported, not scored: Spearman ρ (the outcome is
  bounded, skewed, zero-inflated) and a 10,000-resample bootstrap 95% CI, seed 20260823.
- **Scoring map:** `|z_Oct − 0.3214| ≤ 0.20` → HIT, else MISS, where z is the
  Fisher transform. **The z form governs**; its r image is
  [0.1208, 0.4788], printed to four decimals because both
  two-decimal roundings fall outside the rule they state.
- **Basis as at issue:** r = 0.3108, n = 48,
  p = 0.032, bootstrap CI
  [0.058, 0.683].
- **Declared:** October is the same 48 people with one further
  quarter of receipts, not an independent draw. This tests continuity, not the finding.

### A1b′ — SECONDARY
- **Population:** all candidates passing the filter. **Labelled in publication as
  containing the challenger/incumbent contrast; not a gradient.**
- **Scoring map:** HIT iff `|z_Oct − 0.6559| ≤ 0.20` **AND** r_DEM > 0
  **AND** r_REP > 0. All three conjuncts scored.

### A5 — RETRACTED IN SUBSTANCE, STILL SCORED
- **Scoring map:** `z_pooled − z_incumbents ≥ 0.20` → HIT, else MISS.
- **Basis as at issue:** 0.6559 − 0.3214 = 0.3344.
- **Declared, and this is the point:** the claim it was written to express — that the
  association is "mostly a step, not a slope" — **is retracted**. It is a claim about
  magnitude tested with a correlation, and the magnitudes disagree: the step is
  11.89 points, the fitted line climbs 14.62 points
  across the observed range, residual scatter 8.64 points.
  Across 20,000 bootstrap resamples of the frozen file, A1a′ holds in **14,048**; A5
  holds in **every one of those 14,048** *(v2 printed "20,000 of 20,000", which was the
  number of resamples drawn, not the number in which the condition held — Changelog B,
  item 1; computed by `figures.py::dependence()`)*. **A hit here is evidence about the
  weakness of the test, not about the claim**, and publication says so. It is scored
  because §5.5 forbids withdrawing a registered prediction.
- **Declared context, not scored (Changelog B, item 7):** the question A5 mis-asked is
  answered by a two-coefficient fit published with the register — share on
  served-at-all and years-served: step 9.41 points (95% CI 6.83–11.98), slope 0.348
  points/year (0.145–0.551, p = 0.0011), crossover 27.0 years, computed by
  `figures.py::step_and_slope()`. It was fitted 2026-08-30, after the freeze, so it is
  published **unregistered and unscored**, and no causal claim attaches to it.

### A3 — frozen, known bad · A3′ — corrected
- **Covariate, both:** binary — committee chair **or** ranking member of any standing
  committee as of the October roster, from official committee membership pages (Tier 1).
  Subcommittee chairs do not count. **The coded list is published before scoring.**
- **Estimator:** partial Pearson correlation of tenure with share, controlling for it.
- **A3 scoring map:** `partial r ≥ 0.50` → HIT, else MISS. Registered 65%; expected MISS.
- **A3′ scoring map:** `0.10 ≤ partial r ≤ 0.45` → HIT, else MISS.
- **Declared limitation, both:** one binary control is not an identification strategy.
  Majority status, committee assignment, seat safety, district industry mix and total
  receipts are uncontrolled. **No causal claim is made or will be made.**

### A2
- **Population:** incumbents in all 435 districts, same filter, same definitions.
- **Estimator:** Fisher z of the all-435 incumbents-only r, minus Fisher z of **A1a′'s
  October value on the same vintage** — referent fixed here, not the June figure and not
  the pooled figure.
- **Scoring map:** `|Δz| ≤ 0.20` → HIT, else MISS.
- **Declared:** the samples are nested; this is not a two-independent-sample comparison
  and no p-value is claimed.

### A4 — frozen, known bad · A4′ — corrected
- **Population:** incumbents in the 2024 competitive set, rebuilt by re-running
  `build.py` against `weball24 / cm24 / pas224 / cn24 / ccl24`. **The 2024 district list
  is fixed from Cook's final 2024 ratings and published before the rebuild is run.**
- **A4 scoring map:** `0.50 ≤ r ≤ 0.85` → HIT, else MISS. Registered 70%; expected MISS.
- **A4′ scoring map:** `0.10 ≤ r ≤ 0.55` → HIT, else MISS.
- **Declared:** largely the same members appear in both cycles. **This is not an
  independent replication and will not be described as one.**

---

## §2 The writing predictions

### Scoreability floor — binds B1, B2, B3, B4
Scoreable only if **≥ 90 of 115 candidates yield ≥ 5 collectable social posts each**
under the frozen protocol. **Failing the floor scores every B prediction MISS, not
unscoreable.** Attrition is what C1 and C2 predict; letting it delete predictions would
let the denominator be selected on a variable correlated with the outcome.

**All four are scored on the terminal-punctuation-corrected pipeline (§0). Both values
are published with every verdict.**

### B1
- **Population:** candidates whose written-official and social cells both meet the frozen
  minimums.
- **Groups:** "professional media background" = pre-candidacy paid employment in
  broadcast, journalism or on-air media, per §H of the evidentiary standard. **The group
  assignment is published before October.**
- **Estimator:** difference in mean social-register FK grade. Two-sided Welch t; **and**
  TOST equivalence, margin 1.0 grades, α = 0.05.
- **Scoring map:** HIT iff `p > 0.05` two-sided **AND** equivalence established at 1.0.
  MISS if `p ≤ 0.05` in either direction, **or** equivalence not established.

### B2
- **Population:** the **first 100** written attributed quotations collected, ordered by
  district code then candidate surname — **the ordering is fixed here so the sample
  cannot be chosen later.**
- **Scoring map:** **any single instance, in any quotation, of any length** → MISS. Zero
  across all 100 → HIT.
- **Declared:** a Beta(1,1) posterior on 23-for-23 gives 19% for 100 more. The stated 70%
  departs from that on the mechanistic ground that releases quote their principal in the
  first person by construction. **A miss falsifies the mechanism, not merely the number.**

### B3
- **Population:** candidates with a computable written-quote → social gap and an
  established age.
- **Estimator:** Mann-Whitney U on the under-55 vs 55-and-over split; **plus** leave-one-
  out removing the oldest subject; **plus** TOST, margin 1.5 grades.
- **Scoring map (v4 correction):** HIT iff the split fails at `p ≥ 0.05` two-sided **AND**
  equivalence is established at margin 1.5. **Every other outcome, inconclusive included,
  is a MISS.** v1's map omitted the equivalence conjunct its own claim carried, so an
  underpowered test could make the claim false while scoring a hit.

### B4
- **Estimator:** median written-quote → social FK gap, national office-holders (benchmark
  file) vs candidates.
- **Scoring map:** HIT iff office-holder median ≥ 3.5 **AND** candidate median ≤ 2.5.
  Else MISS.
- **Basis as at issue, corrected pipeline:** candidate median
  **3.54** — above the 2.5 conjunct, so **expected MISS**.
  On the shipped pipeline it read 2.38 and the conjunct
  held. **The instrument was wrong; the bet was the bet.** Confidence stays 55%.
- **Declared:** n = 4 office-holders at registration. Underpowered and labelled as such.
  A second instrument defect, declared 2026-08-30, is published with every B4 figure:
  the protocol's per-register corpus minimums (600 words official / 300 social) admit
  3 of the 19 pooled candidates; the compliant-only median is published beside the
  scored one. The scored pipeline is unchanged — the bet was the bet, twice.

---

## §3 Collection

### C1
- **Estimator:** count of the 115 with no social cell meeting the frozen minimum ÷ 115.
- **Scoring map:** `≥ 15%` → HIT, `< 15%` → MISS.
- **Published with the result:** the per-candidate reason for every failure.

### C2
- **Baseline (register v5, Ruling 3):** **57 D /
  56 R**, baseline odds 1.018. The two independents are excluded
  from this test. *(v1 fixed the baseline at 57:57, "exactly 1.00"; Kiley's March party
  change makes that false, and the roster column had fossilised the Phase 1 source.)*
- **Estimator:** odds ratio of meeting the social minimum, Democrat vs Republican, on the
  2×2. Fisher exact p reported but **not** part of the scoring rule.
- **Scoring map:** `OR ≥ 1.30` → HIT. `OR < 1.30`, or OR < 1 → MISS.

---

## §4 Roster, ages, calibration, service and legal record

### D1
- **Estimator:** count of the 59 races with a material change — withdrawal, replacement
  nominee, party switch, or suspension — after 2026-08-21, from Tier 1/2 sources.
  *(Changelog A, item 2: v1's definition also listed death and disqualification and
  counted districts; v2 narrowed it silently. v2's definition governs and is restated
  here unchanged — the narrowing runs against the prediction, and a third rewrite would
  be worse than the second.)*
- **Scoring map:** `≥ 2` → HIT, `< 2` → MISS.

### E1
- **Scoring map:** `2 ≤ newly establishable ≤ 4` of the seven → HIT, else MISS.
  Establishable means meeting rule **G4** of `EVIDENTIARY_STANDARD_v1.1.md` (stated,
  identified, editorial, not partisan-origin) — not appearing on a data broker.
  *(Changelog A, item 3: v2 dropped the G4 pin and the denominator; both are restored
  as the stricter published reading, which is a tightening of wording, not of rule.)*

### E2 — the only prediction with an unscoreable branch
- **Subject (Changelog B, item 6): Carey Coleman, OH-13** — named in register v1's own
  heading ("Coleman's asterisk resolves to 67, not 66") and on the card since
  2026-08-30; restated here so this document scores without cross-reading.
- **Scoring map:** resolves to the older value → HIT. Resolves to the younger → MISS.
  **Does not resolve → UNSCOREABLE**, counting as neither.
- **Declared:** the 182:139 window assumes uniform birth dates; US births are seasonal by
  roughly ±5–8%, which places the true probability in the 53–61% band. Publishing 57% to
  the point without that caveat was overprecision.

### E3
- **Scoring map:** `≤ 1` of the four markers resolves → HIT. `≥ 2` → MISS.

### F1
- **Scoring map:** the October publication contains ≥ 1 item explicitly labelled
  RETRACTED or DOWNGRADED, referencing a claim published before 2026-09-01 → HIT. None
  → MISS.
- **Declared conflict:** this outcome is partly under the author's control. It is scored
  on the published artifact rather than on judgment, which is the most externally
  checkable form available — **but it is not a clean forecast and is not read as one.**

### F2 — COMMITMENT, NOT SCORED
Publish, for the scored set: the **Brier score**; the **log-odds calibration shift â**
with its 95% CI; and the **Murphy decomposition** (reliability − resolution +
uncertainty). No threshold attached — attaching one is what created v1's paradox.
*(Changelog A, item 4 / Changelog B, item 3: v2 silently dropped the Murphy
decomposition; it is restored here, and the build has carried a decomposition-identity
self-test throughout.)*

### F2′ and F2″ — same proposition, two prices, both scored
- **Estimator:** â, the MLE of the shift in `y ~ logistic(logit(p) + a)`, fitted on the
  **23 predictions excluding these two** (§0).
- **Scoring map, both:** `â < 0` → HIT. `â ≥ 0` → MISS.
- **As at issue** the register's own confidences give
  P(â < 0) = **0.9356**.
- **Declared:** F2′ is priced at 60%, F2″ at 93% — the same claim, repriced once four
  predictions were known dead. Both are scored. **The 33-point gap is the size of the
  advantage declined**, and F2′ will hit cheaply.

### G1
- **Population:** the roster, **Unclear removed from both sides** (§0). *(Changelog A,
  item 1: this population is itself a declared post-freeze change from v1's "extension
  set"; it governs, and G1's card says what that makes a hit worth.)*
- **Estimator:** Fisher exact, two-sided, on the 2×2 of veteran status by party.
- **Scoring map:** `p > 0.05` → HIT, `p ≤ 0.05` → MISS.
- **Basis as at issue:** 17/54 R vs
  10/56 D, **p = 0.1223,
  OR = 2.114**. *(The figure first published, p = 0.186 on
  denominators of 57 and 57, counted the Unclear candidates as non-veterans — the
  opposite of the rule. The verdict is unchanged; the basis figure was wrong.)*

### G2
- **Any recorded legal matter** = the legal column carries one of the enumerated adverse
  values — **Conviction, Charge filed, Settlement, or Civil judgment**. The positive
  enumeration governs; "None found" and "Not established — see detail" are not matters.
  **The exculpatory qualifier travels with the charge in every published form** —
  `roster_115.csv` carries the disposition column and the full detail column, not the
  one-word summary.
- **Population:** rows excluded under the adverse-downgrade rule (§0) are out of both
  the numerator and the denominator; both bases are published whenever that binds
  anything. As at issue it binds zero rows.
- **Estimator:** Fisher exact, two-sided, by party.
- **Scoring map:** `p > 0.05` → HIT, `p ≤ 0.05` → MISS.
- **Basis as at issue (Changelog B, item 4):** 6/56 R vs
  7/57 D, **p = 1.0000,
  OR = 0.857**. *(v2 printed 5/56 vs 7/57, p = 0.7616 — the basis before the same-day
  coding correction declared on G2's card. The correction moved the p-value further into
  the region this prediction bets on, and the card says so in those terms.)*

---

## §5 Adjudication

1. **Every prediction above has a total scoring map. No outcome requires judgment.** If
   one is found that does, that is **a defect in this document**, logged as such in the
   Source Log — not resolved by the author's discretion.
2. **Scoring runs 2026-10-31**, on data downloaded once, on a date recorded in the log.
3. **Every input is published with the result** — the excluded-candidate list, the
   media-background assignment, the committee-chair coding, the per-candidate collection
   failures, the Unclear list — so a third party can recompute every number.
4. **Disputes:** any reader may submit a challenge with the recomputation attached.
   Sustained challenges are published as errata beside the original, never as edits.
5. **No prediction is withdrawn after registration.** The only permitted post-registration
   status is HIT, MISS, or E2's single UNSCOREABLE branch.
   **Case law, corrected 2026-08-30 (Changelog B, item 2):** this rule was tested when
   A5's characterisation was found wrong on **2026-08-23 — the day it was registered**;
   the ruling recording the retraction issued with register v5 on 2026-08-24. The
   prediction was **not** deleted. It stands, is scored, and is marked
   retracted-in-substance with the reason published on its card. **A rule that binds
   only until it is inconvenient is not a rule** — and the cost of honouring it is one
   near-certain undeserved hit, disclosed in advance.
6. **Every published figure comes from a script in the repository.** Any figure appearing
   in publication without a corresponding run is a defect and is logged as one. *(v2
   itself violated this rule in its A5 section — Changelog B, item 1 — which is the most
   instructive defect in this series: the binding document typed the number.)*

---

*v3 issued 2026-08-30, superseding v2 (also 2026-08-30) and v1 (2026-08-22). Corrections
go to v4 with a changelog; nothing here is edited in place. Basis figures are as at issue
and are not live. v1 and v2 stand verbatim in the record and in the timestamped
commitment; reading them beside this document shows every change this series has made,
which from v3 onward is the only way changes happen.*