Josué Guevara

J-Editorial — State of the research

Toward a semantic model of editorial state, history, evidence, and intervention.

Abstract

Writing systems are good at storing text, metadata, and revisions. They usually represent much less about why an artifact exists, which obligations apply to it, what is currently wrong or unresolved, what evidence supports those judgments, what editorial intervention occurred, and whether that intervention actually improved the artifact.

J-Editorial investigates whether that missing information can be represented as a reusable semantic layer between raw document history and writing-product behavior. The project is not yet a finished framework, runtime, benchmark, or writing product. Its current work is pre-implementation research: defining candidate editorial semantics, testing them against unlike forms of writing, preserving evidence and disagreement, and trying to falsify the model before implementation choices make the assumptions expensive to change.

Evaluation systems, specialist agents, benchmarks, and possible training datasets are downstream uses of this work. The editorial core is intended to remain useful without any of them.

From revisions to editorial state

A version history can show that a sentence changed. It does not necessarily show that a source changed first, that the old sentence consequently became stale, that a reviewer identified the problem, that evidence was attached, that several remedies were considered, that one remedy was authorized, that an edit attempted to resolve the finding, or that verification later succeeded or failed.

Those distinctions matter to editorial systems. They separate an observation from a fact, an obligation from a finding, a proposed operation from its outcome, and release acceptance from permanent correctness. J-Editorial asks whether these relations can be made explicit without turning Git commits, files, line numbers, Markdown, a particular database, or a particular editor into the canonical domain model.

Revision history and editorial state as linked but separate layersThree material document checkpoints form a revision history. A separate causal chain records a changed source or norm, a finding, supporting evidence, an authorized intervention, and the verified outcome. Dashed links anchor the causal record to exact document checkpoints.MATERIAL REVISION HISTORYCheckpoint 01Exact artifact statebefore interventionCheckpoint 02Exact artifact stateafter authorized editCheckpoint 03Exact artifact stateafter verificationEDITORIAL STATE AND CAUSAL RECORDSource or normA relevant externalcondition changesFindingThe artifact is judgedagainst that conditionEvidenceThe basis for thejudgment is preservedInterventionAn authorized operationattempts resolutionOutcomeVerification recordssuccess or failureCheckpoints preserve what the artifact was.Editorial state preserves why a transition occurred, on what evidence, under whose authority, and with what result.
A raw revision chain records material checkpoints. The candidate semantic layer records the conditions, findings, evidence, authorized operations, and outcomes that explain those transitions. The layers are linked by exact artifact states but are not collapsed into one representation.

A working model under falsification

The current research shape is deliberately more concrete than a vague design principle and less final than an architecture. It combines goals, applicable norms, exact artifact checkpoints, typed editorial data, evidence, causal transactions, time-scoped conditions, and versioned projections.

GoalNormsExact checkpointsTyped editorial dataEvidenceCausal transactionsTime-scoped conditionsVersioned projections

Goals state what an artifact is trying to accomplish. Norms describe applicable standards, policies, conventions, and requirements. Checkpoints preserve exact material states. Editorial data represents distinctions such as observations, obligations, findings, gaps, authority, and outcomes. Evidence records the basis for a judgment; transactions preserve editorial actions and causal context; conditions allow something to remain true between events; and projections derive views such as findings, gaps, and release readiness under a particular goal and interpretation.

The inquiry around that model has five linked questions. It asks what the smallest useful vocabulary is; whether editorial state can be reconstructed causally and temporally; whether the same semantics can cover technical documentation and ordinary prose without erasing their differences; how observations, claims, reviewer disagreement, norms, approvals, and outcomes should remain distinguishable; and whether richer histories can support evaluation without treating every human edit as ground truth or reducing document quality to a single score.

This model is authorized for research and falsification only. Physical event-first, fact-first, event–fluent, checkpoint-plus-journal, and other representations remain open.

How the inquiry is being tested

The first bounded proof deliberately pairs two unlike artifacts: an Amnesia API reference and the portfolio About page. Success on one cannot substitute for success on the other. The technical slice emphasizes exact source parity and structured correctness, while the prose slice requires different forms of adjudication and permits multiple defensible resolutions.

Candidate models are subjected to counterexamples, fixtures, mutation tests, behavioral probes, representation comparisons, and adversarial review before they can become architecture. The current protocol also separates preregistered claims and failure conditions from fixture authoring, validator execution, and red-team work. Earlier packets that predate those independence rules remain in the record at their bounded strength; they cannot by themselves clear the empirical-validation gate.

Primary specifications, archival research, industrial practice, repository observation, experiments, domain judgment, secondary synthesis, and hypotheses are kept as distinct evidence classes. Programs retain charters, result logs, a shared source ledger, decision records, executable fixtures, and dated packets so that the path from hypothesis to evidence to later architectural decision remains inspectable.

What the evidence currently permits

The repository-wide empirical-validation and architecture-selection gates remain open. The current record supports only bounded findings at the strength recorded by each program.

Checkpoint integrity is established only for the current technical-reference and prose slices: their exact repository checkpoint chains can be re-inspected, and the prose record preserves a date discrepancy instead of normalizing it away. Semantic expressibility remains NARROW. The current vocabulary can be encoded across both slices, but several rich semantic fixtures and validators were authored with knowledge of the hypothesis, so they demonstrate fit by construction; they do not independently establish semantic sufficiency.

Current fixtures support machine-checkable referential, causal, and localized targeting checks, while the broader event-state catchall rewrite remains unresolved in its returned-with-findings record. Raw numerical offsets can become stale under structural edits; stronger targeting strategies recovered some localized mutations or failed explicitly instead of fabricating a match. The shared vocabulary also covers both current research slices, but that paired-domain result remains narrow because both artifacts share an owner and author and the proof was constructed with knowledge of their repository structure.

Evaluation remains downstream

If an editorial system can preserve a trace such as problem → evidence → intervention → outcome, production history may become more useful evaluation material than a simple before-and-after text pair. That opens later work on deterministic checks, human and model grader comparison, benchmark episodes, regression testing, portable policy packs, specialist-agent evaluation, and carefully governed training datasets.

None of those possibilities defines the editorial core. The intended dependency direction is from evaluation and agent infrastructure toward the editorial semantics, not the reverse.

Editorial semanticsReusable editorial runtimeDoc Doctor reference integrationReproducible evaluation infrastructurePortable policies and benchmarksPossible writing-system backend

This sequence is a capability hypothesis, not a release calendar. Each step has its own proof obligations, and the project may stop, split, or change direction when evidence fails to justify the next boundary.

Research record

The references below are pinned to the J-Editorial research snapshot inspected for this report. The repository link remains live.

  1. Bootstrap contractHypotheses, invariants, forbidden assumptions, and implementation hold at this snapshot.
  2. Research charterEvidence standards, program lifecycle, and research-to-decision boundary at this snapshot.
  3. Research programs and resultsBounded program records used by this report.
  4. Product roadmapProvisional capability progression and proof obligations at this snapshot.
  5. J-Editorial repositoryLive source research record.
  6. Doc DoctorIntended first downstream reference product.