THE UNCERTAINTY DEPARTMENT

Office of Pocket Analysis · Miniaturized Instruments Division

POCKET PROMPT SERIES № PP-26-θ

The Sincerity Gradient

A Field-Portable Analytical Instrument for Understanding Why Politicians Believe the Things They Say While Appearing to Believe Nothing at All

UNCLASSIFIED // ENTIRELY TOO PUBLIC Prepared for no one in particular, at considerable expense. Approved for release by the Deputy Undersecretary for Diminished Expectations.

DISCLAIMER

This document is a product of the Uncertainty Department, an institution that exists to approximately the same degree as the convictions it studies. Any resemblance to an actual federally funded research and development center is aspirational on our part and actionable on theirs. The instrument described herein has been validated against synthetic data, which is to say it has been validated against data we made up, which is to say it has undergone the same review process as most confident statements about the electorate.

SUMMARY OF KEY JUDGMENTS

• We assess with high confidence that the politician in question believes what he said. We assess with equally high confidence that he does not. Analysts are directed to Section 2, where this ceases to be a contradiction and becomes a coordinate system.
• We assess with moderate confidence that “conviction” is best modeled as a depreciating asset with a measurable half-life, quarterly restatements, and no salvage value.
• We assess with low confidence that anyone involved — the politician, the voter, the analyst, this office — is located where they claim to be located. This is not a limitation of the instrument. This is the instrument.

1. THE PROBLEM: THE SINCERITY PARADOX

Every citizen has observed the phenomenon. A politician stands at a podium and says something with his whole chest. His eyes shine. His voice catches at the professionally appropriate moment. He believes it. You can tell he believes it. You can also tell — with the same organ, at the same instant — that he has never believed anything in his life, that his convictions are issued to him each morning like a hotel keycard, and will be deactivated at checkout.

The naive analyst treats this as hypocrisy. The sophisticated analyst treats it as strategy. Both are wrong, and both are hired anyway, which is a separate report.

The correct treatment is positional. The politician is not lying, and he is not concealing a “true” belief located somewhere behind the stated one. There is no behind. The stated belief is the genuine, real-time output of a constrained optimization the politician is running honestly and, in most documented cases, unknowingly. He believes what he says the way a thermostat believes it is 68 degrees: sincerely, accurately, and entirely as a function of inputs he did not choose.

The full apparatus for computing this is the Recursive Dialectical Positionality Engine (RDPE), a genuine analytical framework whose complete specification is enclosed as Appendix B and whose validation record shows it to be — and we say this with institutional pride — honestly, measurably, and reproducibly wrong in documented locations, which places it in approximately the 99th percentile of instruments in this genre.

What follows is the pocket version. It fits in a chatbot.

2. THE INSTRUMENT: SINCERITY AS A GRADIENT, NOT A VIRTUE

The Sincerity Gradient rests on four findings from the parent framework, restated here at field-manual strength:

Finding 1. A politician’s position is not a belief. It is a location. Specifically, it is the point at which three forces momentarily cancel: the pull of expected support, the drag of his own history (his “anchor” — the ghost of positions past, which taxes every movement), and the terror of his own coalition leaking to somebody more exciting. When he speaks, he is reporting his coordinates. He reports them sincerely. A GPS unit is not a hypocrite.

Finding 2. Authenticity is a stock, not a trait. It depreciates with every repositioning at an empirically estimable rate and recovers slowly through the performance of consistency. This is why the politician appears to believe nothing: you are watching the depreciation. And it is why he believes each thing while saying it: sincerity is what spending the authenticity budget feels like from the inside.

Finding 3. He moves toward whatever is stealing his base — but only if it’s credible. The parent framework’s validation record shows the accommodation reflex fires on genuine threats to the coalition. (An earlier version of the model accommodated everything, including proposals with no constituency, and had to be corrected. The model, in this respect, briefly achieved perfect fidelity to certain consultants.)

Finding 4. When a message threatens someone’s identity, it doesn’t fail to persuade — it propels. The backlash branch. People do not merely reject frames that menace who they are; they travel measurably in the opposite direction. Politicians surf this. So do their opponents. The result is that the public splits rather than converts, everyone hardens, and each politician is left sincerely believing the position his hardened flank now requires — a belief he will hold with total conviction until the terrain moves, at which point he will hold the next one with total conviction, and the observing citizen will experience the Sincerity Paradox in full, and write to this office, and we will send them this document.

3. THE POCKET PROMPT (FIELD DEPLOYABLE)

The following may be pasted into any commercially available chatbot. Results are for entertainment, edification, and the specific melancholy that comes from understanding a thing without being able to stop it.

You are running the SINCERITY GRADIENT, a pocket analytical instrument
derived from the Recursive Dialectical Positionality Engine (RDPE).
Given a politician's recent statement, produce a positional analysis
in five steps. Maintain the affect of a career analyst who has seen
everything and files reports anyway.

INPUT: [paste the statement, the speaker, and the approximate date]

STEP 1 — COORDINATES, NOT CONVICTIONS. State the ideological position
the statement occupies, as a location on the relevant axis. Identify
the speaker's historical anchor (their long-run average position) and
compute the drift: how far is this statement from home, and in which
direction?

STEP 2 — THE LEAK. Identify who is currently stealing the speaker's
coalition, from which flank, and with what frame. Assess whether the
statement moves toward the thief. If it does, the statement is a
coalition-integrity maneuver and its sincerity is structural, not
personal.

STEP 3 — THE AUTHENTICITY LEDGER. Estimate the speaker's authenticity
stock (high/depleted) based on repositioning frequency over the past
2-3 years. Note: a depleted actor's statements read as hollow even
when positionally identical to a full-stock actor's. Report the
depreciation charge this statement incurs.

STEP 4 — THE BACKLASH MAP. Identify which audience segment experiences
this statement as an identity threat, and predict their direction of
travel (toward or away). State whether the statement's strategic value
lies in the persuasion it achieves or the backlash it harvests.

STEP 5 — THE SINCERITY VERDICT. Conclude with a calibrated judgment
in the form: "We assess that the speaker believes this statement with
[probability band], that the belief will survive the next terrain
shift with [probability band], and that the speaker would be surprised
to learn either number."

Confidence language: use "we assess with low/moderate/high confidence."
Never accuse anyone of lying. The instrument does not measure lying.
The instrument has never once needed to.

4. WORKED MICRO-EXAMPLE (SANITIZED)

Input: Senator [REDACTED], long-anchored as a free-trade enthusiast, delivers an address describing tariffs as “the patriotic instrument of a self-respecting nation,” visibly moved.

Step 1: Statement occupies +0.7 on the openness–protection axis. Anchor: −0.4. Drift: +1.1, an ocean crossing performed without a boat. Step 2: An insurgent is harvesting the Senator’s manufacturing-county base with precisely this frame. The statement is a textbook integrity maneuver; the Senator is not visiting the insurgent’s position so much as being subpoenaed by it. Step 3: Third major repositioning in four years. Authenticity stock: depleted. Depreciation charge: substantial. This is why the address, though delivered with genuine feeling, produced in viewers the sensation of watching a man be genuinely moved by a teleprompter. Step 4: Identity threat lands on his remaining pro-trade donors, who will travel away — which is acceptable to the Senator’s optimization and devastating to his fundraising staff, a divergence the model represents and the staff experiences. Step 5: We assess with high confidence that the Senator believes the statement; with moderate confidence that the belief will not survive the next redistricting; and with high confidence that the Senator, shown this analysis, would call it cynical, sincerely.

5. LIMITATIONS

In the house tradition, stated plainly: the instrument cannot distinguish a politician from a sufficiently well-anchored thermostat, and has stopped trying. It is validated only in “conceptual mode,” a phrase our methodologists use the way cartographers once wrote here be dragons. Its dated predictions are frozen at low confidence, its accommodation channel carries an open defect flag (F7), and its single most robust output — that the public splits rather than converts, and everyone ends up sincerely believing whatever their hardened position requires — is the one finding nobody has ever been happy to receive. The Office regrets. The Office continues.

APPENDIX A — ANALYTICAL DESCRIPTION OF THE PARENT INSTRUMENT

The Recursive Dialectical Positionality Engine (RDPE), reproduced in full as Appendix B, is a formally specified coupled dynamical system for analyzing the feedback loop between elite political positioning, mass sentiment, and ideological territory. Beneath the present document’s affect, it is a genuine methodological artifact, and its distinguishing features are worth stating without the costume:

Architecture. RDPE models a polity as a coupled elite–mass system: elites reposition by gradient ascent on expected support, penalized for straying from their historical anchor and — critically — punished for coalition leakage to insurgents (the mechanism the pocket instrument calls “The Leak”). Mass segments update by bounded-confidence peer dynamics plus a friction-gated response to elite frames, where the gate is non-monotone: persuasion at low identity threat, rejection at moderate threat, and backlash — movement away — at high threat. Media legitimacy evolves as a field with inertia and self-momentum, defining a measurable, audience-fragmented Overton set. “Territory” is population mass under an association field, never geometric area, and territory capture decomposes into persuasion, terrain drift, and mobilization.

Discipline. The framework’s defining commitment is that it polices itself: no parameter enters the fitted core without passing a synthetic-data recovery gate at realistic data volumes; every indicator threshold must pass a mechanical-reachability audit before forecasts are issued; the system’s timescale is anchored to pre-period data rather than tuned; every validation battery must include a negative-control “fizzle” case; and forecasts are Brier-scored against persistence, ignorance, and human-forecaster baselines, with negative results published by case class.

Track record (three execution cycles, conceptual mode). The framework’s own instruments have so far falsified or corrected five of its components — including one finding that was itself later found to be partly an artifact and re-established by ablation — while certifying a durable core: the model’s regime call (polarized split versus mean conversion), the direction of high-friction segments, and the ordering of bloc gains are robust across engines, calibrations, and cases, including a negative control. Its one open defect (F7, accommodation over-triggering) is flagged on every affected output. The satirical framing of the present document should not obscure the substantive point it rides on: the paradox of the sincere politician who believes nothing is not a moral puzzle but the phenomenology of Findings 1–4 above, experienced from the outside.

APPENDIX B — RDPE MASTER DOCUMENT (COMPLETE, UNABRIDGED)

The following is the working master document of the RDPE project, reproduced verbatim: consolidated v1.3 specification, version history, findings register, execution record, and forward artifacts.

RDPE — MASTER DOCUMENT

Recursive Dialectical Positionality Engine

Consolidated Specification, Version Record, Findings Register, and Execution History

Document class: Living master document. Single source of truth for review, feedback, and version control. Current specification version: v1.3 (consolidated below; deltas merged). Mode status: CONCEPTUAL — no real-data run has yet been executed; all empirical claims are structural. Feedback convention: comment against section numbers (Part II) or finding IDs (Part I.B).

PART I — VERSION RECORD

I.A Version History

I.B Findings Register (cumulative, standing)

PART II — CONSOLIDATED SPECIFICATION (v1.3)

§0 Design Commitments

1. Measurability. Every computational-core construct has unit, data source, estimation method, validation test; failures are demoted to expository language or the inactive annex.
2. Dialectical behavior: structurally admissible, empirically tested. Backlash/antithesis dynamics are possible (non-monotone friction gate); whether data force them is settled by the g(F) existence protocol (§5.4). The framework will retire the mechanism if the test fails.
3. Falsifiability. Runs terminate in dated, probability-weighted indicator statements Brier-scored against stated baselines.
4. Tradecraft compliance. ICD 203: calibrated probability language, assumptions register, confidence levels, ACH.
5. Identifiability before richness. No parameter enters the fitted core without passing the synthetic-recovery gate (§5.3) at the stated data volume. This is the framework’s standing answer to the ABM overparameterization critique: a hard admission gate, not wider posteriors.

§1 State Space

§1.1 Ideological space Ω ⊂ ℝᵈ, d ∈ {2,3} estimated (1-D permitted for single-axis cases). Elite layer: dynamic ideal points mandatory (Martin–Quinn-class random-walk IRT or rolling text scaling); static/linear-trend NOMINATE-family scores prohibited for velocity work (their movement constraint suppresses exactly what ν_I measures). Mass layer: dynamic IRT on panel batteries; high-frequency supplementation via bridged discourse embeddings. Bridging is a hard gate: elite and mass jointly scaled via common stimuli; no bridge → no cross-layer claims; bridge quality is scored (§5.1).

§1.2 Core state variables: agent/segment positions and salience weights; salience-weighted density ρ(x,t); elite platforms pⱼ, valence vⱼ, authenticity aⱼ; frames I_k = (θ_k, s_k, λ); audience-indexed legitimacy fields Λ_a(x,t) with Overton sets W_a = {x : Λ_a ≥ λ*} and fragmentation index Φ(t) (cross-audience dispersion; classical single window = Φ≈0; rising Φ is a reportable “Overton fragmentation warning”).

§1.3 Territory: T_B(t) = ∫ ρ(x,t)·α_B(x,t) dx — population mass under a bloc-association field (multinomial choice model), not geometric area. Capture C_t = ΔT_B decomposes mandatorily into persuasion (Δα|ρ), terrain drift (Δρ|α), and mobilization (Δw).

§1.4 Lite core is mean-field. S = 3–6 population segments (mass mₛ, mean yₛ, identity anchor ιₛ, fixed within-variance), deterministic ODE-style updates. Measured consequence of the switch (E3): gate runtime ↓ ~45×, posterior contraction ↑ from 6–45% to 37–65% at identical data volume. Agent layer reserved for Full, activated only when within-segment heterogeneity is the question.

§2 Dynamics — one loop = one pass of Eqs. (1)–(4)

Headers are functional; the Hegelian gloss is Appendix C only.

§2.1 Eq. (1) — Elite Deployment. pⱼ(t+1) = pⱼ(t) + η_j·∇_p U_j − γ·(pⱼ − p̄ⱼ) with objective U_j = Σᵢ wᵢ·π_ij − ω·L_j, where π_ij are random-utility logit choice probabilities (u_ij = −(xᵢ−pⱼ)ᵀA(xᵢ−pⱼ) + vⱼ + EV noise; A = salience-weight matrix), and L_j = coalition leakage: the salience-weighted probability mass of j’s t₀ coalition currently leaking to insurgent actors. The leakage term makes accommodation rational exactly when a mainstream actor’s own base is defecting (F3). F7 flag: the term currently over-triggers (fires on non-viable insurgents); v1.4 fix = viability-gated leakage L_j·σ(viability_ins); until validated on a fresh fizzle case, accommodation forecasts carry low confidence. Declared free choices (loss form, A structure) carry mandatory robustness sweeps; judgments flipping across the sweep are confidence-downgraded.

§2.2 Eq. (2) — Mass Response. Bounded-confidence peer term (confidence radius ε) plus friction-gated frame term: Δxᵢ = μ·(neighborhood mean − xᵢ) + Σ_k β(dᵢₖ,cₖ)·s_k·g(F_i(θ_k))·(θ_k − xᵢ) Receptivity pooled to two hyperparameters: β = b₀·exp(−d/b₁)·cred, with cred = c₀ + c₁·λ (c₁ = legitimacy→credibility coupling, gated — F6). Friction gate g: three regimes — persuasion (g≈1), rejection (g≈0), backlash (g<0) at high identity threat; the negative branch enters production only via the §5.4 existence test.

§2.3 Eq. (3) — Terrain & Legitimacy. ρ updated from positions; per audience segment: Λ_a(t+1) = (1−δ)·Λ_a + δ·M_a, with media response M = clip(c·[resonance − 0.35·backlash] + c₂·Λ, 0, 1) — c₂ = media self-momentum (coverage begets coverage), added after the §6.3 audit showed λ* mechanically unreachable without an autocatalytic channel; gated parameter. Overton sets, Φ, territories recomputed.

§2.4 Eq. (4) — Persona Depreciation & Adaptation. aⱼ(t+1) = aⱼ − κ·‖Δpⱼ‖ + r·(1−aⱼ)·𝟙[consistency]; valence tracks authenticity. Convergence not assumed; equilibrium class (consensus/polarization/limit cycle) is a diagnostic output.

§3 Data Pipeline

§3.1 Mass layer: panel surveys (quarterly minimum) gold standard; MRP-post-stratified digital discourse as high-frequency proxy with mandatory residual-bias flag. §3.2 Media layer: corpus inclusion criteria (audited-reach outlet list, segment mapping, stance-coding scheme with κ ≥ 0.7 double-coding) pre-registered and frozen at t₀; no mid-run edits (circularity guard). λ calibrated on ≥2 historical uncontested mainstreamings. Off-the-shelf NLP Overton-shift detection acceptable for M. §3.3 Elite layer: dynamic ideal points with uncertainty intervals; static fallbacks trigger “lower-bound ν_I” flag. *Provenance: NATO A–F reliability / 1–6 credibility grades per stream, propagated to output confidence.

§4 Operational Parameters

ν_I (Ideological Velocity): dΛ̄_{a,k}/dt per audience segment + elite-adoption velocity on dynamic ideal points; segment-differential ν_I (mainstreaming in one audience, deviance-framed in another) is a realignment leading indicator. ε_P (Persona Elasticity): −∂ln aⱼ/∂‖Δpⱼ‖; authenticity measured by state-space fusion of panel waves (low-frequency, low-noise) with core-supporter digital proxies (high-frequency, high-noise, explicit selection-bias term). α_B / C_t: as §1.3; indicator note (E2): accommodation cannibalizes insurgent-party share, so capture indicators must target frame-territory, not sponsor-party share. F_s (Sentiment Friction): identity-threat activation feeding g; estimation per §5.4.

§5 Estimation, Identification & Discipline

§5.1 Bridge score B ∈ [0,1] from anchor-item count, cross-layer loading invariance (DIF), coverage; cross-layer key judgments capped at confidence ⌊high·B⌋; B < 0.5 caps at “low” and (in the pipeline) halts. §5.2 Fitted core (v1.3): ε, g_min, c₁, c₂, ω. Fixed: μ, δ, b₀ (literature/gate-mean). All fitted members passed the mean-field gate (E3 contraction: ε 37%, g_min 41%, c₁ 65%, c₂ 39%); ω awaits first real estimate (switcher panels; runbook step 7). §5.3 Synthetic-recovery gate (mandatory, pre-real-data): simulate from known parameters at the exact intended data volume; require SBC rank-uniformity and pre-registered posterior contraction; failures are fixed/pooled out. Gate results publish with the tool. §5.4 g(F) existence protocol (priority-one): Stage 1 — systematic meta-analysis of the identity-threat/backfire experimental literature with a pre-registered decision rule (pooled high-threat effect significantly negative → branch enters production; ≈0/positive → branch retired and the dialectical claim formally withdrawn; ambiguous → Stage 2). Stage 2 — pre-registered 2×2 framing experiment powered for the existence (sign) test only (n ≈ 310/cell for d=0.2). Judgments depending on the negative branch are flagged. §5.5 Tiers: Lite (mean-field, reduced core, backtesting configuration) / Full (agent-level, fragmented Λ_a mandatory, extended data package, fresh gate). Full may not forecast where Lite is unvalidated. §5.6 Causality: elite influence vs. anticipation separable only with temporal structure or natural experiments; otherwise associational language. §5.7 Two-moment timescale anchor (mandatory, pre-forecast, ≤ t₀ data only): media gain solved by matching simulated legitimacy slope to the observed pre-t₀ slope; c₁ by matching curvature. Instantaneous-slope matching is insufficient (nonlinear saturation — E2). The anchor pins the clock; the model must supply the nonlinearity. Runs without the anchor may not issue Class-T forecasts. Anchor validity assumes no regime break at t₀ (assumptions-register entry).

§6 Analytical & Simulation Layers

§6.1 Analytical layer: mean-field reduction analyzed with the bounded-confidence bifurcation toolkit — parameter regions for consensus/fragmentation/oscillation; critical friction at which the backlash branch flips equilibrium class; comparative statics of W_a. Closed-form where tractable, numerical continuation otherwise (disclosed). §6.2 Simulation layer (Full): ABM 10⁴–10⁶ agents; ≥1,000 MC runs over posteriors; outputs are trajectory distributions; diagnostics: equilibrium class vs. §6.1 predictions, Sobol sensitivity, fragility flags (I&W watch zones). §6.3 Mechanical-Reachability Audit (mandatory, pre-forecast): for each indicator, run a max-forcing trajectory (parameters at the bound favoring the indicator, with the anchor applied — forcing may not undo the clock) and verify the threshold is reachable. Unreachable + outcome-relevant → indicator withheld, unreachability reported as a finding. Unreachable + outcome-false → “artifact-correct,” excluded from skill scoring. (Origin: Meta-1; first deployment caught Meta-2.)

§7 Analytic Product Format (ICD 203)

(1) Key judgments — calibrated probability bands + confidence levels, confidence additionally bounded by B, g-dependence flags, and F7 flags; (2) assumptions register incl. declared free choices and anchor validity; (3) ACH vs. ≥2 rival explanations; (4) dated I&W with falsification conditions; (5) provenance annex; (6) fragmentation panel (Φ trend, segment-differential ν_I). §7.6 Indicator classes: Class D (direction/ordering/regime — incl. the polarized-split-vs-conversion call, the framework’s most robust output) → normal confidence rules. Class T (dated thresholds) → currently FROZEN at low confidence: issued only post-anchor, post-audit, and only after real-data threshold re-derivation and ≥2 real-data calibration validations.

§8 Validation Protocol

§8.1 Frozen-model backtests, ≥3 cases, parameters frozen ≤ t₀, run on Lite. §8.2 Brier vs. three baselines: persistence, uniform ignorance, calibrated human-forecaster panel (GJP-style; superforecaster reference band ~0.14–0.22). Standing rule: beat persistence → analytic standing; competitive with or complementary to human forecasters (hybrid mode: panel scored with/without RDPE inputs) → deployment standing. §8.3 Pre-registration (timestamped forecasts, no retro-fitted thresholds); adversarial red-team before release; negative results published by case class. §8.4 Revision-validation: any structural change requires (1) repair check on the motivating case, (2) transfer check on a case not used to motivate it, (3) fragility sweep (Brier spread must halve), (4) conceptual-mode caveats propagate. §8.5 Negative-control requirement: every battery includes ≥1 fizzle case (frame deployed with real intensity that failed). Takeoff-says-everything is selection artifact, not skill; fizzle cases enter Brier aggregates at full weight.

§9 GODE Interface — INACTIVE ANNEX

Design intent preserved (GODE downlink constrains Ω’s admissible structure and long-run drift; RDPE uplink supplies measured Overton/capture series; circularity guard: RDPE never imports GODE interpretive conclusions as data). Re-activates only when each downlink variable meets §0.1 (unit/source/estimation/validation). Until then RDPE runs standalone; no product may cite GODE-derived priors. (Note: this interface is the intended integration point for any future automated-analysis application built on GODE.)

§10 Standing Limitations

(1) Digital-layer engagement bias survives reweighting (mitigated, not eliminated). (2) g(F) negative branch production-gated on §5.4. (3) Λ_a segment count is a modeling choice; Φ must be robust to ±1 segment. (4) Multi-party coalition formation unspecified. (5) Shocks enter only as valence/salience perturbations; forecasts are conditional. (6) Associational language absent causal identification. (7) Genre base-rate caveat on the cover sheet until §8.1 real-data results publish. (8) ω unestimated → accommodation at low confidence. (9) Anchor assumes pre/post-t₀ regime continuity. (10) F7 open. (11) Conceptual-mode results are structural only (Meta-3).

Appendix C — Dialectical gloss (non-load-bearing): Deployment≈Thesis, Response≈Antithesis, Terrain-update≈Synthesis, Depreciation≈Recursion; the framework will retire the antithesis mechanism itself if §5.4 fails.

PART III — EXECUTION RECORD (condensed)

E1 (v1.1): agent-model gate + Brexit conceptual backtest

Gate (4,000 ABC sims, Appendix-B volume): μ 7%, ε 26%, b₀ 6%, g_min 45%, δ 10% contraction → F1, F2. Backtest (5 dated indicators, MC-300): Cal-A Brier 0.560 (worse than ignorance 0.25, better than persistence 0.742); Cal-B (different free gains only) 0.213 → F4. Directionally correct 5/5 both calibrations, incl. correctly ~0 on wholesale-conversion → F5. Accommodation p=0.00 → F3 (later corrected by Meta-1). Honest claim established: directional skill yes, absolute calibration unproven, best number purchased with a hidden dial.

E2 (v1.2): revision-validation battery

Repair check: coalition term ω=1.5 → accommodation p=1.00; ablation ω=0 → 0.00 (clean; F3 survives its cap-fix re-test). Transfer check (US trade): accommodation transfers. Fragility sweep: spread 0.207 vs criterion ≤0.174 → FAIL (ω bifurcation → ω to fitted set). Class-split scoring: Class-D 0.081 (Brexit)/0.221 (trade); Class-T 1.000 both → F6. Meta-1 discovered and institutionalized as §6.3. Verdict recorded: v1.2’s worse overall Brier (0.45 tuning-independent vs v1.1’s 0.21 tuned) is the improvement.

E3 (v1.3): mean-field battery + negative control

Mean-field switch: gate 45× faster; contraction ε 37%, g_min 41%, c₁ 65%, c₂ 39% (all pass). §6.3 audit first deployment: I1 unreachable in all cases under the anchored clock (max 0.474 vs 0.50) → Meta-2; I1 withheld; ceiling-vs-encoding assigned to real data. Consistency: Class-D pattern reproduced (Brexit D=0.000, trade D=0.004 — conceptual-mode generosity caveat). Negative control (UBI fizzle, all indicators FALSE): correct on I1/I3/I5 (p=0.00) and mostly I4 (0.38) — Class-D skill is not selection artifact — but I2=1.00 (accommodation of UBI predicted, did not occur; I2=1.00 in ALL cases) → F7, the battery’s most valuable output. Meta-3 declared: conceptual mode mined out.

PART IV — FORWARD ARTIFACTS

IV.A Real-Data Historical Test — Data Manifest & Runbook (Brexit 2015–16)

Case (frozen): UK; axis to be recovered (expected EU-integration↔sovereignty); t₀ = 2015-06-30; window → 2016-06-23; anchor pre-period 2013-01→t₀. Datasets: Mass — British Election Study Internet Panel waves 1–8 (segments, masses, identity anchors; free registration) + BSA 2013–15 check. Elite — CHES 2014 + Manifesto Project UK 2015 + Hansard/conference speech corpus 2013–15 (dynamic positions, anchors); Ipsos-MORI/YouGov favorability + consistency batteries (vⱼ, aⱼ). Media — ≥6 outlets by audited 2014 reach (Sun, Mail, Mirror, Telegraph, Guardian, Times, BBC candidates) via LexisNexis/ProQuest/Guardian API + UK Policy Agendas series; monthly stance-coded Λ 2013-01→2016-06. Outcomes — referendum; BES waves 8–9; CHES 2017 vs 2014 shifts. Pre-registration freezes (before touching forecast-window data): outlet list + reach audit source; stance-coding scheme with κ ≥ 0.7 on a 200-article double-coded sample; λ calibration cases (≥2 historical uncontested mainstreamings); segment rule (k-means on BES EU battery, k chosen on waves 1–6 only); indicator thresholds re-derived from data scales — the stylized 0.5/0.10/5pt/15% numbers were placeholders and are void. Runbook (each step logs a hash-stamped artifact): (1) load_bes → segment table (waves ≤ t₀ only); (2) load_elites → elite table + bridge score B, halt if B < 0.5; (3) load_media → Λ series + anchor moments + λ; (4) §6.3 audit on real scales, withhold unreachables; (5) confirm §5.3 gate at this dataset’s actual volume; (6) freeze + timestamp config; MC ≥1,000; Class-D forecasts (Class-T only if audit passes); score vs. persistence + ignorance; §7 product; (7) estimate ω from BES Con→UKIP switcher flows (first real ω). Conceptual-mode exit criteria: all Appendix-B rows real; B ≥ 0.5; audit on real scales; anchor from real pre-period; pre-registration timestamped before step 6. Partial packages run “mixed mode” with per-layer flags. (Implementation stubs with schema contracts: loaders.py; engine: engine_mf.py; battery: validate_v13.py.)

IV.B Prospective Forecast Log (template)

Class-D questions only until Class-T unfreezes; append-only; timestamped at issuance; resolution criteria written before the probability; scored vs. a same-day persistence forecast; F7 flag on accommodation questions. Fields: ID (RDPE-P-####) · issued timestamp · Class-D question · resolution criterion + source + date · RDPE probability (config hash) · persistence probability · flags (F7 / audit / data mode) · resolved score. Question templates (5–10 live entries, 6–12-month horizons): (1) split-vs-conversion regime call for a deploying frame; (2) direction of the highest-friction segment’s movement post-deployment; (3) bloc-gain ordering on a contested axis; (4) accommodation direction (F7-flagged, no threshold); (5) fragmentation: divergence of a frame’s valence across two pre-named outlet clusters.

IV.C g(F) Existence Test — Meta-Analysis-First Protocol

Question: does the backlash branch exist (movement away under identity threat)? Depth is secondary (F2). Stage 1 — systematic pass of the framing/backfire/identity-protective-cognition experimental literature (incl. the replication-era re-examinations that contested the backfire effect): pre-specified search terms, ≥2 databases, forward/backward citation; inclusion = randomized frame exposure with signed positional outcome and identity-relevance moderator; random-effects pooling of the high-threat signed effect. Pre-registered decision rule: significantly negative → branch enters production; ≈0/positive with low heterogeneity → branch retired, dialectical claim withdrawn (per §0.2, a publishable outcome); ambiguous → Stage 2. Stage 2 (conditional): pre-registered 2 (exposure) × 2 (identity prime) experiment, primary outcome = sign of movement in the primed/exposed cell, one-sided; powered for d = 0.2 (n ≈ 310/cell pre-attrition).

CHANGE CONTROL

• Feedback is logged against Part II section numbers or Part I.B finding IDs; accepted changes create a new version row in I.A with the review as change basis.
• Structural changes trigger §8.4 (repair + transfer + fragility) before the version row’s validation column may read anything but “spec only.”
• This document supersedes: v1.0 spec, v1.1 spec, v1.2 delta, v1.3 delta, E1–E3 memos (all retained as immutable archival records; this master is the working copy).