A story. A causal graph.
A philosophical object.
Tragedy Engine makes its assumptions inspectable. A repeatable result is a property of a specified model, not proof that a literary interpretation is uniquely correct.
The textual witness
We use Sophocles’s Oedipus the King in Francis Storr’s 1912 translation, archived from Project Gutenberg #31. Source bytes and cited passages have SHA-256 hashes. Claims are checked against dialogue and stage directions. Read the pinned passages →
What stays separate
- Canonical records: the source-supported events and claims. A character saying something does not establish its truth.
- Interpretation: the reviewed causal rules, information annotations and initial-state reconstruction.
- Deterministic trace: the results of replaying guards, event prerequisites, information requirements and decision policies through the original horizon, t10.
- Projected continuation: an authored, checked extension supported by the resulting state and the pinned source. Its events and witnesses remain separate from the canonical records and original-window metrics.
Change a set of conditions
The five interventions target real units of the current model. Timing uses the original delayed-transfer policy. The other operations suppress a complete rule episode: all of its effects, assertions and transfers are disabled together. For the opening investigation, its associated decision option is disabled too. These operations use an isolated copy of the IR; the original engine and source remain unchanged.
Only the explicitly selected conditions may differ initially. Every other difference must be causally downstream. Held constraints remain in force in the continuation. Removing an encounter is not the same as selectively removing just one of its statements.
The five choices have 31 nonempty subsets. Seventeen are compatible and all are precomputed. Fourteen are rejected: the final answer cannot both arrive later and remain withheld, and a delayed final answer requires the parentage inquiry to reach its fork. An upstream cut and a later held suppression can coexist; the latter may then be redundant.
Continuity after the original window
Continuations preserve established goals, character information, remaining evidence and the original world. They introduce no new rescuer, secret clue or arbitrary punishment. The delayed-testimony continuation reuses the original testimony, recognition and reporting rules in that order; each guard and information gate must pass again. Other continuations preserve an active search or a blocked civic duty without inventing new disclosures or harms.
The reporting rule represents a canonical sequence, not a universal law that public recognition causes suicide. Jocasta has already inferred the identity before the final testimony. Her offstage death cannot be dated from Oedipus’s public recognition. A missing report is therefore never converted into proof of rescue.
The mathematical object
G = (V, E) is the realized event-support incidence graph through t10. V contains emitted events. An edge joins a producer event to a later event that uses its state, assertion, prerequisite or decision input. Parallel reasons between the same pair are stored on one directed edge. Initial facts and assertions remain in the IR; they are outside this graph’s distance calculation.
An incidence path shows a dependency route. It is not an independent sufficient cause: a recognition rule may need several incoming inputs together. All / any / not logic remains in the full rule system. Target reachability is checked by replay, never by an OR-only graph traversal.
Definitions you can reproduce
- Reachability: a target is reached exactly when its event is emitted by the gated replay.
- Path preservation: enumerate directed paths from the original graph’s indegree-zero roots to each target. A path is retained only if all its nodes and edges remain in G′. The six recognition routes and eight report routes share necessary inputs; these counts do not measure causal redundancy.
- Structural distance: let U(G) contain typed node IDs and edge IDs. d(G,G′) = |U(G) △ U(G′)| / |U(G) ∪ U(G′)|. Every unit has equal weight. This Jaccard distance is conditional on this event granularity and horizon.
- Rule cut: κ(G′) is the minimum number of additional active non-target event rules whose suppression prevents at least one target from occurring. We exclude direct deletion of recognition or report. All candidate subsets are checked until the smallest cut size is found. If the targets are already cut, κ = 0.
- Local tolerance: R(G′) is the number of single further rule removals preserving both targets, divided by all eligible active non-target rules. The original model preserves both targets under one of seven probes: removing the seer episode.
- Minimal selected set: enumerate all compatible curated sets S. Report inclusion-minimal sets that cut the original target sequence and min |S|. In this catalogue the minimum is one.
- Multiple-change robustness: Rₖ = compatible sets of size k retaining both targets / all compatible sets of size k. The table separately counts original-window results and results including the reviewed continuation. These are catalogue fractions, not probabilities.
Interactions, not an addition of endings
Every set is replayed jointly. Let Eᵢ = U(G) △ U(Gᵢ), and Eₛ be the joint edit set. We compare Eₛ against the union of single-change edit sets. We also report η(S) = (|Eₛ| − Σ|Eᵢ|) / |U(G)|, using a fixed baseline denominator. Distances with varying denominators are never added.
- Independent: disjoint single-change edit sets exactly predict the joint edits.
- Redundant: the exact union predicts the joint edits, but some edits are shared.
- Masking: a combination suppresses added branches predicted by an isolated change, without restoring any original unit. Here an upstream cut can prevent the later waiting branch.
- Synergistic: joint replay creates edits outside that union.
- Antagonistic: joint replay cancels predicted edits to original units. Mixed results have both extra and cancelled edits.
The current catalogue has independent, redundant and masking responses. It does not demonstrate a synergistic cut requiring several individually insufficient interventions. That absence is a result of this model, not evidence that all tragedies have the same structure.
Research and review
Research View preserves Structure, Knowledge, Transformation, checkpoint navigation, source locators, full raw traces and legacy metrics. World facts, reported claims, beliefs, memory and knowledge remain distinct. Raw records retain the original formal state vocabulary for reproducibility.
This release has Codex source-and-consistency review and automated validation, not independent literary peer review. No runtime language model, provider key, free-text story generator or cross-tragedy similarity is used. The application loads the precomputed joint result selected by its stable intervention-set key.