Hypotheses

Testable predictions linking the project's theory frameworks (../theory.md) to pre-existing evidence (literature, news anecdotes, institutional facts) and concrete empirical tests on the project's data (../data.md, ../methods.md). Clusters map to the planned paper sections.

Status (2026-06-02). The headline result (β ≈ +7 pp) is estimated (H1, H2 confirmed). Clusters B–E are the next round of work; H3 (Channel A vs B shrinkage) is the highest-leverage open lever.

---

How to read

Each hypothesis is a standalone page under docs/hypotheses/<slug>.md. Cross-reference by slug: [hyp:headline-sponsor-bias] resolves to headline-sponsor-bias.md. The - id: hyp:<slug> anchor on each page is what the cite lint checks against.

---

Cluster A: Headline sponsor effect

The descriptive claim the paper rests on: do sponsor-paid polls inflate the sponsor's candidate? Planned Section: Main result.

• H1: Self-sponsored polls overstate the sponsoring candidate
• H2: Opponent-sponsored polls understate the opposing candidate

Cluster B: Mechanism decomposition (Channel A vs Channel B)

How is the slant produced? Planned Section: Mechanism.

• H3: β shrinks substantially when methodology features are controlled for
• H4: Channel A contribution is larger where methodology flexibility is greater
• H5: Channel B contribution decays as the election approaches

Cluster C: Discriminating voter-side theories

How polls move voters (heterogeneity). Planned Section: Why polls matter (heterogeneity).

• H6: β peaks at the M+1 viability cutoff (coordination)
• H7: β peaks just below the front-runner (bandwagon)
• H8: β is larger for first-time / lesser-known candidates (information cue)

Cluster D: Sponsor-side heterogeneity

Who slants, and through which sponsor structure? Planned Section: Who slants.

• H9: Slant is largest in polls financed by partisan funds
• H10: Reputable national pollsters show smaller β than small regional firms
• H11: Pollster-self-sponsored polls operate as a hidden-sponsor channel

Cluster E: External validation (litigation channel)

Cross-checking the design with poll-related electoral litigation. Planned Section: External validation.

• H12: High-β polls are over-represented in poll-related Representações
• H13: Shell-contratante polls show larger residual β

---

Summary

H#	Slug	Cluster	Theory	Evidence	Data	Status
H1	headline-sponsor-bias	A	Bayesian persuasion	Strong	in hand	confirmed (β≈+7)
H2	opponent-sponsored-mirror	A	Bayesian persuasion	Moderate	in hand	confirmed (β_opp≈−2)
H3	channel-a-shrinkage	B	Bayesian persuasion	Strong	LLM extractor pending	queued (top lever)
H4	methodology-flexibility-a	B	Bayesian persuasion	Weak	IBGE demo pending	queued behind H3
H5	verifiability-decay-b	B	Verifiable disclosure	Weak	in hand	queued behind H3
H6	coordination-peak	C	Coordination devices	Strong	in hand	queued
H7	bandwagon-peak	C	Bandwagon	Strong	in hand	queued (joint w/ H6)
H8	information-cue-experience	C	Quality cues	Moderate	in hand	queued
H9	funding-source-gradient	D	Principal-agent	Strong	in hand	queued
H10	pollster-reputation	D	Career concerns	Moderate	in hand	queued
H11	pollster-self-cover	D	Concealed-sender persuasion	Strong	in hand	queued
H12	litigation-validation	E	Perceived-bias channel	Strong	EJ cross-link pending	queued
H13	shell-contratante	E	Concealed-sender persuasion	Moderate	LLM pass pending	queued

All theory frameworks referenced are in ../theory.md.

Theory.md coverage

All four current frameworks are used by ≥1 hypothesis:

• Coordination devices → H6
• Bandwagon triggers → H7
• Quality cues → H8
• Bayesian persuasion (supply-side) → H1, H2, H3, H4, H11, H13

Brainstormed theories in ../thinking.md (mobilization / turnout effect; fundraising-cascade donor signaling; media-coverage allocation; strategic candidacy / entry-exit; clientelism; coligações) are not yet operationalized here — most lack a project- side outcome dataset. Worth a separate pass if/when corresponding outcomes arrive.