1. The inferential task

Process tracing asks how an outcome came about inside a particular case. The object of inference is not a correlation between variables across many cases, but a causal explanation connected to observable events, decisions, constraints, and transmissions within one bounded historical record.

The project sits between two Bayesian traditions. Narrative or likelihoodist process tracing preserves historically specific explanations and asks how expected each observation would be under each rival. Formal model-based work declares a causal graph and derives implications from its structure and parameters. Both can use Bayes' rule, but they differ in what counts as a hypothesis, how mechanisms are represented, and where likelihood judgments come from. For a single source-rich case, this system is narrative-primary: it retains evidentiary texture while externalizing the reasoning that would otherwise remain implicit. Formal cross-case estimation is kept separate for settings in which cases share variables and the design can identify population or counterfactual quantities.

The system therefore does not ask a model for a free-form historical answer. It moves through five reviewable operations:

2. Research design, estimand, and rival explanations

Every run declares how explanations encounter evidence. In a theory-first design, rivals and observable predictions are fixed before evaluation evidence is examined. A discovery/evaluation split develops explanations from one source set and evaluates them against a disjoint held-out set. An exploratory design may use the full corpus to discover mechanisms and alternatives, but evidence used to formulate an explanation is not presented as independent numerical confirmation.

Rival explanation

A distinct account of the outcome that specifies a causal mechanism and at least one observation expected to differ from another account.

Observable prediction

A trace that should be more or less likely to appear if an explanation is correct, stated before that trace is used diagnostically.

Admitted evidence

Source content within the declared corpus whose text, identity, provenance, and role in the analysis are retained.

The confirmatory gate examines every pair of rivals for overlap, complementarity, or absorption. Numerical comparative support is restricted to a defensible local mutually exclusive partition. Compatible, conjunctural, sequential, nested, and equifinal components remain visible but are not forced into winner-take-all percentages.

The target numerical estimand is deliberately narrow: posterior odds over a finite set of mutually exclusive, mechanism-specific explanations, conditional on the admitted corpus, declared priors, and calibrated diagnostic likelihoods. The current implementation has not earned that probability interpretation. It reports normalized comparative evidentiary support derived from experimental model scores. These scores rank how diagnostically favorable an item appears under each rival; algebraic coherence does not establish that they are proportional to P(E | H).

The current pipeline also includes a flat residual label for “none of the specified stories.” That is an open-world warning, not yet a genuine Bayesian hypothesis: without a predictive distribution for the heterogeneous alternatives it contains, it cannot compete on the same numerical footing. A future version will either model an imprecise mixture over named residual subtypes or exclude the residual from normalization and state that support is strictly conditional on the declared rivals.

Selection and prior provenance

Exposure changes the claim. If an observation helped generate a hypothesis, treating the same observation as independent confirmation would count it twice. The system currently records evidence visible during hypothesis generation and sources informing the prior. Theory-first and genuinely held-out evidence can support confirmatory updating; full-corpus exploratory evidence remains available for explanation-building and mechanism appraisal but is not relabeled as independent support. This is a conservative workbench design policy, not the uniquely correct interpretation of iterative Bayesian inquiry. A future version will extend lineage to acquisition, retrieval queries, inclusion and exclusion, extraction, outcome definition, observable-prediction construction, and prior formation.

Partition sensitivity

The present rival audit detects overlap and incompatible relationships but does not yet make support invariant to how explanations are split or merged. Dividing one parent explanation into two mechanism variants can redistribute normalized mass even when the substantive evidence has not changed. Plan 043 therefore adds a hierarchical parent → mechanism-variant representation, parent-level reporting, and mandatory split/merge sensitivity.

3. Two models behind every observation

A likelihood judgment about evidence rests on more than a story about what caused the historical outcome. It also depends on how the alleged event became an observable trace. The method therefore distinguishes two models that are often collapsed:

Substantive causal model

How actors, conditions, decisions, and events produced later events and the outcome of interest.

Trace-production model

How an event or claim was solicited, recorded, copied, preserved, translated, retrieved, and extracted into the evidence available to the analysis.

This distinction matters because a trace can be present for at least two reasons: the underlying event occurred and was recorded, or the claim was produced even though the event did not occur as stated. Propaganda, bureaucratic formulae, retrospective rationalization, genre conventions, selective survival, and extraction error create false-positive channels. Conversely, censorship, secrecy, destruction, and archival absence can suppress traces of events that did occur.

Retaining H in both trace channels is essential: even when the underlying fact is held fixed, different explanations can imply different incentives to fabricate, suppress, classify, standardize, or preserve its record. The current workbench captures source provenance, coverage, dependence lineage, and source-silence opportunity, but it does not yet represent this complete typed trace-production model. That is an explicit future-version target, not a capability implied by the present interface.

4. Evidence and experimental diagnostic scoring

Every evidentiary proposition must resolve to accepted source text and stable provenance. The exact quotation is a citation anchor, not the whole evidentiary object. Meaning can depend on the surrounding paragraph, genre, authorship, audience, purpose, competence, authenticity, temporal proximity, translation, editorial history, and incentives to misrepresent. The current exact-anchor requirement provides strong provenance but only partial contextual source criticism; Plan 043 makes contextual windows and those source fields mandatory for update-bearing evidence.

For each evidence item, the testing stage proposes one positive vector of relative evidentiary scores across all rivals. Pairwise ratios are derived from that common vector rather than elicited separately. This matters when there are three or more hypotheses: independent pairwise judgments can violate reciprocity or transitivity. Equal vector components mean the model judges the evidence nondiscriminating between those rivals.

This construction establishes algebraic coherence. It does not establish calibration: q may be a stable model score without being proportional to P(E | H). Paci's 2026 benchmark found that GPT judgments could be internally coherent and run-stable while remaining substantially different from expert estimates, predominantly overweighting evidence and sometimes changing the leading hypothesis. Automated magnitude estimation is therefore the principal validity bottleneck.

Priors remain conceptually distinct from evidence scores. Typed prior specifications retain their method, weights, rationale, and source provenance so the same source is not silently used twice. The current deterministic layer aggregates centered scores and exposes sensitivity:

The resulting weights are normalized comparative evidentiary support within the declared rival set. They are not posterior probabilities or posterior odds. Legacy internal fields still use likelihood/posterior names; those implementation names do not license the stronger public interpretation.

Diagnostic tests as a legible interpretation layer

Hoop, smoking-gun, doubly-decisive, and straw-in-the-wind tests summarize the shape of a likelihood comparison. They do not replace the likelihood model. A hoop test describes evidence strongly expected if a hypothesis is correct; failure is damaging, while passage may be weak support. A smoking-gun trace is unlikely under rivals; its presence is strongly supportive, while its absence may say little. The labels make diagnostic logic readable, while the multi-rival vector carries the coherent comparison.

Raw judgment and policy transformations

The system retains the originally elicited vector and records each deterministic transformation separately: relevance adjustment, dependence pooling, eligibility decisions, caps, and sensitivity perturbations. No workload cutoff or display convention is treated as a law of process tracing. A conclusion that changes under plausible priors, likelihood magnitudes, or guardrail settings is reported as fragile.

Dependence and repetition

A valid joint likelihood generally factors conditionally:

The current system instead pools evidence through a claim → event → source → author hierarchy. This reduces obvious duplication, but a scalar discount is not equivalent to a joint likelihood model. Dependence may also be hypothesis-specific—for example, two documents may share one propaganda origin under one rival but represent separate observation channels under another. Plan 043 replaces the hierarchy as the sole representation with an evidentiary-ancestry graph, joint cluster judgments where needed, hypothesis-specific dependence, and leave-one-cluster-out sensitivity.

Absence of evidence

Source silence currently enters numerical aggregation only when an admitted source was positioned to record the predicted trace, the search is complete, no partial observation exists, and a separate audit accepts that opportunity-to-observe judgment. Missing source classes limit the conclusion; they do not prove absence in the world. A fuller trace-production model will replace the binary opportunity gate with rival-relative probabilities of nonappearance while preserving qualitative fallback when observability cannot be defended.

Selection and corpus conditionality

Exact quotations do not prevent selection on diagnosticity or latent model knowledge. A future version will retain the complete candidate-evidence universe—including rejected and nondiagnostic passages—freeze the retrieval and stopping rule, and audit inclusion independently of score extremity. It will also compare famous-name packets with anonymized entities and dates, fictional substitutions, permuted rival labels, and counterfactual packet variants. Material score changes caused by “Napoleon” rather than the admitted evidence violate the intended corpus-conditional claim.

5. Mechanism reconstruction

The method constructs a forward temporal graph from source-grounded stages. Each transition is classified as observed sequence, supported causation, contested causation, or unresolved. Temporal order alone never certifies causation. An unresolved link must identify the evidence that would clarify the causal break.

This distinction prevents a fluent narrative from hiding inferential gaps. A source may establish that event A preceded event B while leaving open whether A caused B, whether a third process produced both, or whether the relevant transmission is missing from the admitted record.

The graph is a qualitative causal audit, not a machine for reading likelihood magnitudes from topology. Without defensible parameters, a graph can reveal a missing pathway, confound, temporal contradiction, or translation loss; it cannot determine whether a likelihood ratio should be two or five. The narrative estimator supplies the magnitude, while the graph forces the causal and evidentiary structure behind that judgment into inspectable form.

The current category supported causation is still a corpus-conditional model judgment. Plan 043 will require each promoted edge to face a miniature rival-relative test: direct transmission versus a plausible alternative transmission, a common cause, and no causal connection. Until that control exists, the graph should be read as a typed and separately reviewed causal argument—not as independently identified causal structure.

6. Separate audit, traceability, and reproducibility

The component that produces a judgment does not certify it. Separate review stages inspect the rival partition, every materially discriminating likelihood comparison, the mechanism stages and edges, and the central claims in the final synthesis. Reviews are bound to the exact candidate artifacts and accepted source spans.

1. Audit whether the rival set supports the proposed comparison.
2. Review discriminating evidence against the exact quotation and opposed prediction.
3. Review every mechanism stage and transition independently.
4. Check terminal claims against their accepted evidence anchors before publication.

Rejected judgments, corrections, attempts, artifact hashes, model routes, costs, and final outputs remain inspectable. This creates traceability of one analytical execution; it does not by itself establish rerun reproducibility, because proprietary model versions, stochastic inference, retrieval indices, and tools can change. Future artifacts will bind exact model and version, prompts, reasoning and decoding settings, tool/runtime versions, source hashes, retrieval-index version, and seed where supported, then measure repeated-run variation.

Three meanings of auditability

Process auditability

Can a reviewer determine what stages ran, with which inputs, models, prompts, costs, and artifact versions?

Interpretive auditability

Can a reviewer trace a specific likelihood, mechanism link, or conclusion to the exact source text and contest its reasoning?

Interpretive validity

Are those judgments substantively defensible? Architecture can expose this question, but only representative empirical validation can answer it.

Independence is graded, not assumed

Separate prompts and blinded roles provide procedural separation by reducing anchoring and self-certification. Different model families can add model independence. Epistemic independence—different information or differently trained priors—is much harder: several models may reproduce the same historiographic stereotype. The default label is therefore separate audit. “Independent audit” is reserved for explicitly blinded cross-family or human review with recorded independence provenance. Agreement is supporting evidence, not proof, and same-model repetition is never described as independent corroboration.

7. Uncertainty, robustness, and falsifiability

Uncertainty enters through the source packet, rival partition, priors, diagnostic likelihoods, dependence assumptions, mechanism translation, and missingness judgments. The report keeps these sources separate rather than compressing them into one generic confidence score. It emphasizes whole-number comparative support, sensitivity ranges, rank stability, contested mechanism edges, and explicit source gaps.

A high comparative weight can coexist with weak evidentiary resolution when every rival is poorly observed or when one explanation merely loses less badly. Conversely, an indeterminate conclusion can be methodologically successful when the admitted record does not discriminate. Contract completeness, historical argument quality, and evidentiary resolution are separate evaluations.

Against ritualized Bayesianism

Equations, graphs, and audit logs can lend opaque judgments a false aura of objectivity. The architecture is answerable only if it can be tested against a simpler narrative-only baseline. Representative validation should measure calibration, discrimination, dependence detection, duplicate-overcounting control, rank accuracy, and expert contestability on held-out or adversarial cases. Thresholds and scoring rules must be fixed before evaluation; otherwise procedural sophistication becomes decoration.

The validation design has two strata. Synthetic and semi-synthetic cases expose a known data-generating process, allowing objective tests of calibration, rank recovery, duplication, dependence, trace production, and missingness. Real historical cases use blinded expert adjudication to evaluate defensibility, context use, error discovery, agreement, and useful disagreement—not an imaginary gold-standard historical posterior.

The decision gate

The benchmark must be able to reject numerical use. The post-MVP Plan 043 design predeclares tolerances for magnitude bias, rank error, proper-scoring performance, anonymization and label-permutation sensitivity, split/merge instability, dependent-cluster influence, contextual reversals, and rerun variation. Its corrected full benchmark remains unexecuted and is not an MVP release criterion. If future calibrated automation does not beat a simpler qualitative baseline on held-out evidence—or if these stress tests materially change the leading conclusion—the default product will suppress normalized numbers and retain scoring only as a structured deliberation aid.

What the unfamiliar-corpus pilot actually showed

On 17 August 2026, the hosted workbench received four previously uninstalled documents about Portugal's 25 April 1974 regime collapse. Eight accepted Gemini provider calls completed extraction through the experimental update for $0.110634; a ninth call stopped at mechanism reconstruction when the provider returned HTTP 503. The attempt was retained without a model switch or substantive retry, so it produced no final causal narrative.

The partial result is methodologically informative. Nine propositions from all four sources resolved to exact admitted quotations. Because the full corpus also generated the four candidate explanations, every proposition was excluded from diagnostic updating and all five displayed weights—including the residual—remained at their equal starting value. That equality is not a finding. The rival audit also classified all six pairs as potentially complementary and two pairs as absorptive, even while allowing the exploratory workflow to continue because each pair had an opposed prediction. This is direct implementation evidence that the present exploratory path can generate a plausible research agenda but does not yet implement the target typed relationships among compatible, conjunctural, nested, and sequential mechanisms.

8. What the result can—and cannot—establish

A completed run establishes what the admitted corpus supports, weakens, or leaves unresolved; how the declared rivals compare under the accepted diagnostic judgments; which causal links remain contested; and what evidence would most improve the inquiry.

9. Extension to comparative studies

Process tracing remains a within-case method. Cross-case inference therefore begins only after each case has a separately inspectable full causal study and then applies a separate selection and comparison design. A compact trace is a useful intermediate projection, but it does not replace the rival specification, evidence testing, mechanism qualification, conclusion, and uncertainty analysis required of a completed case. A common discovery corpus may orient case selection, but it is not reused as independent case evidence.

The ten-case revolution design

The current study asks: across bounded revolutionary and comparison episodes, what combinations and sequences distinguish episodes that establish a materially different governing order from episodes that do not? The unit is one connected challenge to an identified incumbent sovereign order with an observable mobilization, response, and endpoint. “Transformation” means that challengers displaced that order and established materially different governing authority within the frozen episode boundary. Mobilization, violence, leadership turnover, territorial control, or the label revolution is not sufficient.

The eligible universe was first assembled from three declared English-language discovery surfaces and reconciled into 1,963 provisional episodes. Twenty finalists then faced eligibility and pre-outcome comparability review before the separately retained endpoint ledger was joined. The final five–five balance is purposive contrast, not genuine outcome blinding, random sampling, prevalence, or a population estimate: names and boundary descriptions could reveal outcomes, and source availability was a feasibility constraint rather than a reason to select a famous case.

Selection follows structured, focused and theory-relative logic. Cases fill declared roles—anchor pathway, similar-context opposite outcome, same-polity temporal contrast, alternative path, external intervention, near miss, deviant non-transformation, and construct boundary—rather than pretending that one generic “most similar” rule solves every inferential problem. Replacements must fill the same role and preserve the frozen outcome and block design; inconvenient cases cannot be swapped merely for richer or easier sources.

What crosses the case boundary

Within each completed case, rivals are tested against source evidence and the temporal mechanism is qualified. Across cases, the system compares only declared constructs, configurations, and sequences carried by those full studies. It does not pool case-level support weights, create one ten-case rival simplex, count overlapping blocks as independent observations, or infer necessity and sufficiency from recurrence. Every cross-case observation must step back to named cases, then to their exact evidence anchors; dependence and construct failures can exclude a case from a particular claim without removing its value as a boundary test.

The current implementation has ten completed source-limited full case studies and a separate full 18 Brumaire depth reference. Its five-block synthesis was rerun from those ten terminal case artifacts and separately audited. The result remains source-limited: it is a bounded comparative interpretation of the installed corpus, not a population estimate or a general theory of revolution.

When many cases are genuinely comparable, the project has a separate formal path built around explicit causal models and CausalQueries. That path can target identified counterfactual or population estimands only when common variables, case comparability, measurement, and model assumptions justify them. A single case is never sent through both engines and presented as if narrative support and a population causal effect were the same quantity.

10. Present implementation and methodological frontier

The public method distinguishes demonstrated controls from the quality-optimal architecture that guides future work. “Implemented” means the software path exists and has focused technical evidence; it does not mean that substantive historical validity has been established.