NEWS — Design Decisions and Methodological Evolution

What this file is: A reverse-chronological log of design decisions, methodological shifts, and schema changes in the process tracing database. This is not a Git changelog — it records intellectual changes: why the schema changed, what epistemological commitment drove the change, and what was gained or lost.

O que é este arquivo: Log em ordem cronológica reversa de decisões de design, mudanças metodológicas e alterações no esquema do banco de dados. Registra mudanças intelectuais: por que o esquema mudou, qual compromisso epistemológico motivou a mudança, e o que foi ganho ou perdido.

2026-05-26 (evening) — `llm_conversations_database-development/` subfolder structure and HISTORY.md

What changed:

Inside llm_conversations_database-development/:
- Renamed xlxs_templates/ → xlsx_templates/ (typo fix).
- Created subfolders conversations/, references/, drafts/.
- Moved the four conversation exports into conversations/ (filenames preserved; proposed normalized names in conversations/rename-suggestions.md, pending researcher decision).
- Moved the six reference PDFs into references/, normalizing filenames to Authors-Year_Short-Title.pdf.
- Moved the generated draft README into drafts/ and added a provenance comment block at its top flagging it as a pre-v3-pivot framing (advocated decibel weighting, since rejected).
- Wrote llm_conversations_database-development/README.md as a folder-level guide.
At repo root:
- Created HISTORY.md — narrative chronology of database evolution from 2 May 2026 (first Perplexity literature search) through 26 May 2026 (v7 working database). Explains the v3 radical pivot (drop of numerical likelihoods + symmetry of hypotheses + passage-level localization + de-manifestoed READ_ME), the two methodological turning points inside the Claude bayesian-tests conversation on 23 May 2026 (the v3→v4 explaining-outcome redesign and the retraction of the Git-as-pre-registration argument), and the start of the V5 Fairfield & Charman 2022 deep-read on 24 May 2026.
- Archived prior README.md, NEWS.md, METHODOLOGY.md, REPOSITORY_PLAN.md as md_documentation_previous_versions/{NAME}-v{N}.md per the documentation-versioning policy in REPOSITORY_PLAN.md §6.
- Patched README.md and this NEWS.md to reflect the new layout and to cross-link HISTORY.md.

Why: The earlier 2026-05-26 reorganization (see entry below) consolidated everything under llm_conversations_database-development/ but left a flat structure: conversations, PDFs, templates, and drafts at the same level. With 12 items at that level — including a typo (xlxs_templates) and a stray duplicate xlsx — the folder lost legibility. The subfolder split is a hygiene move. HISTORY.md exists because NEWS.md is a decision log, not a narrative; the story of the database — the dead ends, the LLM-induced reversals, the moments where the researcher dismantled a beautiful but misaligned design — required its own document. The dates inside the conversation file headers also corrected an earlier misperception: the work begins on 2 May 2026, not 3 May and not 5 May.

Affected files: HISTORY.md (new); llm_conversations_database-development/README.md (new); llm_conversations_database-development/conversations/rename-suggestions.md (new); llm_conversations_database-development/xlsx_templates/ (renamed); llm_conversations_database-development/conversations/, references/, drafts/ (new subfolders, populated); README.md, NEWS.md, REPOSITORY_PLAN.md (patched); md_documentation_previous_versions/{README,NEWS,METHODOLOGY,REPOSITORY_PLAN}-v{N}.md (archived).

2026-05-26 — Full commitment to Fairfield & Charman (2022): numerical probabilities adopted, Van Evera tests dropped

What changed: METHODOLOGY.md rewritten to reflect full alignment with Fairfield & Charman (2022) — the book, not just the 2017 article. Key changes: (1) adopted the verbal probability scale (quase_certa → 0.95, …, quase_impossivel → 0.05) and the decibel weight-of-evidence system from Ch. 3; (2) dropped the Van Evera tests (Hoop, Smoking Gun, Straw-in-the-Wind, Doubly Decisive) entirely — not even as communicative vocabulary; (3) documented the two-stage evidence workflow (raw → analysis) as implemented in v7.

Why: The v3 template (and the earlier version of METHODOLOGY.md) occupied an intermediate position: qualitative Bayesian reasoning without numbers, Van Evera tests retained as vocabulary. This was based primarily on the 2017 article’s permissive language (“transparency does not require assigning numerical values”). The v7 working database moved past this position by implementing the 2022 book’s more developed operationalization — verbal probability scale, likelihood ratios, and decibel weights. This entry records the decision to align METHODOLOGY.md with v7 and with the full framework of the 2022 book. The Van Evera tests are superseded by the likelihood ratio: evidentiary confirmation is a matter of degree (captured by the LR and dB), not of type (captured by categorical labels).

Affected files: METHODOLOGY.md (major rewrite), NEWS.md.

Key reference: Fairfield & Charman (2022), Ch. 3, especially §3.6.1 (“Inhabit the World of Each Hypothesis”).

2026-05-26 (morning) — Repository documentation creation and methodological-development reframing

What changed: Created README.md (v1), NEWS.md (v0), METHODOLOGY.md (v0), and REPOSITORY_PLAN.md. Preserved original README as md_documentation_previous_versions/README-v0.md. Reframed the repository purpose from “research data repository” to “methodological development repository.” The conversation logs, reference PDFs, and template schemas were initially consolidated under llm_conversations_database-development/ (with a flat structure that was later split into subfolders the same day — see the entry above).

Why: The repository’s core contribution is the database design, not the data it holds. The previous README described the repo as a research data project. The new documentation positions it as a record of methodological reasoning — how to structure evidence for abductive process tracing under imperfect theories. The METHODOLOGY.md synthesizes the epistemological premises (drawn from Fairfield & Charman, 2022) to guide any collaborator or AI assistant.

Affected files: README.md, NEWS.md, METHODOLOGY.md, REPOSITORY_PLAN.md, md_documentation_previous_versions/README-v0.md.

Pre-2026 — v3: Adoption of Fairfield & Charman as epistemological foundation

What changed: Removed numerical probability fields (prob_evidencia_dado_h, prob_evidencia_dado_nao_h, likelihood_ratio, forca_inferencial) from TB_EVIDENCIAS. Replaced with qualitative Bayesian reasoning: discriminacao_H_vs_rivals (analytically primary field) and hipotese_favorecida_id (FK). Removed QualCoder fields. Added url, arquivo_local, pagina_trecho for passage-level auditability.

Why: Fairfield & Charman (2022): “transparency does not require assigning numerical values to probabilities.” The numerical fields suggested false precision. The new discriminacao_H_vs_rivals captures the inferential core — “mentally inhabiting the world of each hypothesis” — without it.

Key reference: Fairfield & Charman (2017, §3.3.1).

Affected files: llm_conversations_database-development/xlsx_templates/banco_pt_process_tracing_v3.xlsx.

Pre-2026 — v3: Symmetry of hypotheses

What changed: Removed tipo_hipotese (principal/alternativa) from TB_HIPOTESES. All hypotheses now have equal structural status. Added justificativa_prior for qualitative prior reasoning. Replaced favorece_H_principal with neutral hipotese_favorecida_id FK.

Why: Fairfield & Charman require comparing rival hypotheses without privileging any as “the main hypothesis.” The justificativa_prior field follows F&C (2017, §3.1): “We can specify priors that aim to reflect our background information as best as possible.”

Affected files: llm_conversations_database-development/xlsx_templates/banco_pt_process_tracing_v3.xlsx.

Pre-2026 — v3: Van Evera tests retained with transparent tension

What changed: teste_bayesiano field retained (Hoop, Smoking Gun, Straw-in-the-Wind, Doubly Decisive), but READ_ME tab rewritten to explain them as communicative vocabulary, not formal procedure — with explicit note on disagreement with Fairfield & Charman.

Why: F&C argue these tests are unnecessary in a Bayesian framework (confirmation is degree, not type). This project retains the labels as communicative shorthand (Bennett 2015, Mahoney 2012) while declaring the tension openly. The discriminacao_H_vs_rivals field is analytically primary.

Affected files: llm_conversations_database-development/xlsx_templates/banco_pt_process_tracing_v3.xlsx (READ_ME tab).

2026-05-23 — v4: Explaining-outcome redesign + numerical likelihoods reinstated

What changed: Inside the Claude bayesian-tests conversation (llm_conversations_database-development/conversations/2026-05-03_claude_bayesian-tests_v1-through-v4.md), the schema was reorganized around three first-class concepts that v3 lacked: (a) TB_EPISODIOS — episodes as distinct causal stories with their own actors and counterfactuals (the project has three: PSDB / private sector, ProUni vs. Quotas Law, FIES expansion); (b) explicit rival accounts at the episode level, with partisan theories as foils to be falsified episode-by-episode; (c) a distinction between diagnostic evidence (showing the partisan mechanism does not operate) and confirmatory evidence (supporting the alternative). Simultaneously, the v3 deletion of numerical likelihoods was reversed: v4 reintroduced p_num_h1, p_num_h2, razao_verossimilhanca, peso_db (weight of evidence in decibels), and peso_categoria with conditional formatting — Fairfield & Charman’s 2017 operationalization in full. The same day, the researcher caught Claude in an unrelated overstatement: the prior framing of Git-commit-before-fieldwork as “epistemological protection” was retracted as a theory-testing leftover, and reframed as audit-trail value only.

Why: Even after v3 removed the explicit theory-testing framing, the schema still felt single-mechanism in use because it lacked an EPISODIOS table. The three-episode reorganization closed that gap. The decibel-weighting reintroduction was not requested — it was the LLM defaulting to its training-data-heavy F&C 2017 operationalization once given a redesign opening. See HISTORY.md §3.4 for the full analysis of this oscillation; METHODOLOGY.md §8 documents this as a recurring AI-misalignment pattern.

Affected files: llm_conversations_database-development/xlsx_templates/v4_banco_process_tracing_baesiano_abdutivo.xlsx; llm_conversations_database-development/drafts/Genereted README_v2 about the database generation an objectives.qmd (intermediate README reflecting the same v4 framing — carries a provenance note flagging it as pre-final).

2026-05-05 — v3: Adoption of Fairfield & Charman as epistemological foundation

(This entry was previously listed as “Pre-2026”. The actual date is 2026-05-05 — file mtime of banco_pt_process_tracing_v3.xlsx. Triggered by Q11 of the v1/v2 Perplexity session at 2026-05-05 12:02.)

2026-05-05 — v2: Schema normalization and fingerprint formalization

What changed: Introduced normalized tables: TB_HIPOTESES, TB_ETAPAS_MECANISMO, TB_FONTES, TB_FINGERPRINTS. Added fingerprint_esperado field. Added evidencia_presente boolean. Added TB_REVISOES for audit trail. Added numerical probability fields (later removed in v3, then reinstated in v4 — see entry above and HISTORY.md §3.4).

Why: The v1 flat structure could not answer relational queries (“which evidence supports M2 against H_alt1?”). Normalization enabled SQL JOINs and separated concerns. The fingerprint_esperado field formalized Beach & Pedersen’s (2019) “core inferential logic”: defining expected observable implications before evaluating whether they were found.

Affected files: llm_conversations_database-development/xlsx_templates/banco_pt_process_tracing_v2.xlsx.

2026-05-03 — v1: Initial flat database schema

What changed: Created first evidence database schema — flat structure with 16 fields: id_evidencia, etapa_mecanismo, teste_bayesiano, direcao_evidencia, hipotese_alvo, confiabilidade_fonte, among others.

Why: Starting point based on Beach & Pedersen’s (2019) theory-testing framework. Assumed a well-specified theory with identifiable fingerprints — an assumption relaxed in later versions as the project moved toward explaining-outcome abductive PT. Generated via Perplexity AI conversation (claude46sonnetthinking in COPILOT mode) at Q6 of the v1/v2 Perplexity session on 2026-05-03 at 09:04, after preceding literature-search questions Q1–Q5 starting 2026-05-02 at 23:09.

Affected files: llm_conversations_database-development/xlsx_templates/banco_pt_process_tracing.xlsx; llm_conversations_database-development/conversations/2026-05-02_perplexity_pt-literature-and-v1-v2-schema.md.