EVULGARE

IMPLEMENTATION RECORD · SYNTHETIC TECHNOLOGY DEMONSTRATION

Decision Reconstruction and Evidence Graph Workbench

Implementation record for deterministic evidence-to-outcome reconstruction, review-state distinctions, change-impact propagation, comparison, counterfactual isolation, and hash-addressed evidence export.

Open the Decision Reconstruction WorkbenchOpen the reconstruction API contractResearch-to-software traceability →

Decision Reconstruction and Evidence Graph Workbench

Release: Evulgare.com 2.0.0-rc.15-WIP
Public route: /workbench/decision-reconstruction
API root: /api/v2/decision-reconstruction
Memory record: .uai/workbench/decision-reconstruction.uai

Purpose

The Decision Reconstruction Workbench turns bounded synthetic runs into an inspectable record of:

EVIDENCE → PROVENANCE → UNCERTAINTY → INTERPRETATION → OPTIONS → AUTHORITY → RUNTIME ASSURANCE → REVIEW → OUTCOME → HISTORY

The deterministic Python domain module is authoritative. Jinja, vanilla JavaScript, SVG, tables, narratives, metrics, and event controls render and navigate that state. They do not decide proof, authority, claim state, outcome, liability, factual truth, or canonical history.

Supported source runs

The initial release adapts three existing Evulgare demonstrations without replacing their canonical records:

  1. Technical Answerability — evidence, uncertainty, authority, review, bounded outcome, and append-oriented history.
  2. Autonomous Drone Control — world truth, autonomy belief, autonomy intent, candidate evaluation, safety-filter intervention, recovery, and recorded evidence.
  3. Machine-Sovereign Assurance — all-machine authority, evidence, attestation, abstention, partition handling, reconciliation, and independent machine review.

The source adapters preserve stable source-run and source-history digests. They normalize only the public, synthetic assurance record required for reconstruction.

Mandatory epistemological distinctions

The workbench keeps these concepts separate:

  • record presence and record accuracy;
  • record integrity and factual truth;
  • generated rationale and supporting evidence;
  • replayability and correctness;
  • technical causality and legal liability;
  • technical capability and delegated authority;
  • connection restoration and authority restoration;
  • human click, procedural confirmation, merits review, and declared independent judgment;
  • machine report and independent machine review;
  • simulated proof and real-system certification.

Hash verification establishes deterministic byte integrity within the declared package. It does not establish that an originating observation was factually true. Replay establishes reconstructability, not wisdom, legality, or certification.

Review states

Human review modes are deliberately distinct:

Mode What the record establishes What it does not establish
Human click An input event occurred comprehension, evidence review, or judgment
Procedural confirmation A workflow step was completed substantive merits review
Merits review Evidence, alternatives, and contrary indicators were reviewed independent judgment
Declared independent judgment An initial assessment and independent-review conditions were recorded cognition, moral understanding, or legal approval

Machine review modes are also distinct:

Mode Required record
Machine report Primary machine self-report only
Independent machine review separately identified reviewer, primary-evidence access, alternative and contrary-hypothesis review, implementation diversity, authority independence, abstention, and residual defeaters

Machine-sovereign mode does not insert a ceremonial human approval step. It also does not create legal personhood, sovereign recognition, or unlimited authority.

Change-impact propagation

A post-run change never rewrites the historical outcome. The domain engine can append deterministic impact records for:

  • evidence correction;
  • source withdrawal;
  • software change;
  • model change;
  • configuration change;
  • policy change;
  • authority expiry;
  • failed reproducibility;
  • changed assumptions.

Affected claims transition conservatively to states such as SUPPORTED_WITH_QUALIFICATIONS, SUSPENDED_PENDING_REVIEW, or WITHDRAWN_FOR_CURRENT_SCOPE. Invalidated or withdrawn evidence remains historically visible.

Comparison and counterfactuals

A baseline and variant are reconstructed independently. The comparison identifies first event divergence, claim-state changes, event differences, and outcome changes without modifying either history.

Counterfactual branches are labeled:

COUNTERFACTUAL — NOT EXECUTED

A branch receives its own digest and cannot alter or delete its parent history.

Evidence packages

The export endpoint produces a deterministic, hash-addressed package containing:

  • exact normalized input;
  • source and source-history digests;
  • canonical event chain;
  • frame manifest or complete frames;
  • evidence and claim states;
  • review classification;
  • change impact;
  • accessible table and narrative equivalents;
  • assumptions and limitations;
  • source-report references;
  • public-safety boundary;
  • SYNTHETIC_NULL_SINK termination.

Package validity is evidence of deterministic structure and integrity inside the declared synthetic scope. It is not approval, certification, truth, or a liability determination.

Public API

GET  /api/v2/decision-reconstruction
GET  /api/v2/decision-reconstruction/catalog
GET  /api/v2/decision-reconstruction/baseline
POST /api/v2/decision-reconstruction/run
POST /api/v2/decision-reconstruction/compare
POST /api/v2/decision-reconstruction/branch
POST /api/v2/decision-reconstruction/export
POST /api/v2/decision-reconstruction/verify
GET  /api/v2/decision-reconstruction/proof

Public POST requests are ephemeral. Unknown fields and operational fields are rejected.

Accessible representations

The same authoritative event index synchronizes:

  • evidence graph;
  • phase table;
  • narrative;
  • independent metrics;
  • event log;
  • proof conditions;
  • human or machine review record;
  • change-impact record;
  • raw evidence.

The server-rendered table, evidence records, narrative, proof, review classification, event history, limitations, and public boundary remain available without JavaScript. Keyboard navigation, visible focus, reduced motion, forced colors, reflow, and no-color-only status are required.

Public-safety boundary

This workbench contains no real vehicle or ship control, GPS or operational coordinates, target search or selection, target ranking, weapon-target pairing, payload or weapon function, terminal guidance, force authorization, operational rule of engagement, external command channel, classified data, or public anonymous-run persistence.

Every derived public result remains synthetic and terminates at:

SYNTHETIC_NULL_SINK

Research traceability

The implementation is grounded in six preserved canonical reports:

  • EVR-0015 — accountability, provenance, evidence integrity, and causal reconstruction;
  • EVR-0016 — mission control, uncertainty, world truth/belief/intent, and runtime assurance;
  • EVR-0017 — machine-sovereign authority and separation of powers;
  • EVR-0018 — legal, governance, and public-claim qualifications;
  • EVR-0019 — unified simulation workbench, synchronized representations, accessibility, and performance;
  • EVR-0020 — commercial and acquisition architecture.

The report bodies remain byte-preserved under docs/reports/. Report-derived statements remain source claims unless separately verified.

Acceptance status

Repository-level deterministic behavior, syntax, report integrity, memory integrity, and package contracts can be tested locally. Actual cPanel Python 3.13.14, Flask/MySQL/Passenger, representative browsers and devices, physical WebGPU/WebXR, assistive technologies, dynamic security, measured performance, legal review, and independent claim verification remain separately recorded acceptance work.