# OfOne Architecture: Typed Causal-Geometry Compiler **Status:** Active architecture **Category:** Research Methodology + Knowledge Architecture + Decision Compiler **Last Updated:** 2026-05-14 --- ## Summary OfOne turns a bounded objective into an auditable decision map by compiling domain language into typed causal geometry. The operating geometry is: ```text bounded objective + geometry primitives + adapter projection + scene map + evidence ledger + claim graph + causal / constraint graph + loop map + option moves + decision surface + update triggers + human gates = decision rendering ``` The internal map is structured state. The final answer is an addressable rendering node derived from that state. External sources, reports, benchmark cases, and model critiques are untrusted input until compiled into typed objects. --- ## Core Principle Abstract geometry is primary. Adapters project domain language onto geometry. This makes OfOne portable across domains without forcing a single domain vocabulary. Medicine, proof search, policy, strategy, engineering, art, and ethics each supply different semantics for evidence, proof, mechanism, uncertainty, value, and review thresholds. The compiler layer remains stable. --- ## Primitive Geometry Layer | Primitive | Meaning | |---|---| | Scene | A bounded state-space snapshot: what exists, what changes, what can be observed. | | Subscene | A local decomposition of a scene for evidence acquisition, causal mechanism work, option decisions, proof steps, stakeholder context, or review gates. | | Frame | A coordinate system for interpreting the scene: causal, logical, strategic, normative, temporal, evidential. | | Token | The smallest typed unit: entity, variable, claim, evidence item, uncertainty, constraint, option, trigger. | | Unknown | An addressable null object for missing evidence, missing measurement, unresolved conflict, missing adapter, or unobserved variable. | | Move | A transformation: infer, observe, test, intervene, decide, revise. | | Edge | A typed relation with semantic family: causal, evidential, argumentative, or workflow-state. | | Loop | A recurrent dependency: feedback, control, learning, incentive, measurement, contradiction, review. | | Invariant | A constraint or symmetry that survives across frames or reruns. | | Gate | A condition that blocks movement until evidence, review, permission, or redaction exists. | --- ## Traversal Order The user-facing chain is a traversal through a typed graph: ```text 0 Charter -> 1 Geometry Kernel -> 2 Adapter Projection -> 3 Scene Map -> 4 Evidence Ledger -> 5 Claim Graph -> 6 Causal / Constraint Graph -> 7 Loop Map -> 8 Hypotheses And Kill Tests -> 9 Option Moves -> 10 Tradeoff Surface -> 11 Update / Patch Logic -> 12 Human Gates -> 13 Decision Pack ``` The underlying system is a graph with reverse edges. Evidence can invalidate an adapter. A claim conflict can force boundary revision. A kill test can reveal missing observables. A human gate can reclassify stakes. A regime shift can trigger trunk rewrite. --- ## Movement Economy Every emitted sentence or object must perform at least one job: | Job | Meaning | |---|---| | BOUND | objective, scope, horizon, stake, exclusion | | GROUND | evidence, provenance, reliability, gap | | CLAIM | atomic proposition | | LINK | support, contradiction, causality, dependency, evaluation | | TEST | hypothesis, kill test, counterfactual, falsifier | | MOVE | intervention, query, proof step, experiment, reversible action | | EVALUATE | criteria, tradeoffs, costs, stakes, reversibility | | WARN | uncertainty, dissent, hidden variable, adversarial risk | | TRIGGER | no-op, patch, rerun, trunk rewrite, monitor | | GATE | review, permission, redaction, approval | The compression rule is simple: delete content that does none of these. --- ## Output Modes | Mode | Use When | Output | |---|---|---| | Micro | quick, lower-stakes decision support | charter, adapter, top claims, decisive uncertainty, recommendation or gate | | Map | normal OfOne run | geometry chain, evidence, claims, graph, options, triggers | | Audit | high stakes, handoff, research pack | full schemas, dependencies, dissent, validators, review log | --- ## Layer Boundaries OfOne separates three layers: 1. **Core IR:** charter, adapter projection, scene, evidence, claims, unknowns, kill tests, edges, loops, option moves, triggers, gates, confidence model, and decision rendering. 2. **Decision lifecycle:** artifact identity, criteria, tradeoff surface, actors, temporal model, information value, lenses, council result, review log, review cycle, convergence gate, and benchmark trace. 3. **Domain extensions:** adapter-specific structures such as proof certificates, measurement protocols, competitor maps, rights-impact records, or failure-mode tables. A new object belongs in core only if it improves almost every serious decision map. Otherwise it belongs in the decision lifecycle or in an adapter extension. Research tools belong outside the core IR. Cross-surface research and cross-domain transfer are represented as subscenes and lenses that feed the existing evidence, claim, unknown, information-value, council, trigger, and rendering objects. For volatile or multi-surface decisions, invoke `cross-surface-trend-intelligence` as an evidence-acquisition lens. For Medici-effect style analogical research, invoke `cross-domain-transfer` as a challenge lens. The former asks what current surfaces say; the latter asks what distant domains imply. Neither bypasses evidence grading or dependency closure. ```text cross-surface research -> evidence / claims / unknowns / triggers cross-domain transfer -> hypotheses / kill tests / measurement ideas / options OfOne compiler -> map / confidence / synthesis / decision rendering ``` External Deep Research and frontier-model review add another control loop around the compiler, but they still feed the same map: ```text prepared packet -> launch proof -> active research -> harvested report -> adjudicated findings -> implemented patch -> public resubmission ``` The important invariant is proof of state transition. A prompt file or context packet is only `prepared`. A run is only `launched` after the Chrome extension/plugin verifies an isolated ChatGPT tab, the external system exposes a research plan, Start/countdown has occurred, and the UI shows `Researching...` or an equivalent stop-control state. Browser, Computer Use, coordinate clicking, AppleScript/JXA, and generic desktop automation are not fallback launch paths; if extension control is unavailable, the packet stays `prepared` or `blocked`. Active runs keep run-scoped status ledgers so launch, progress, harvest, and acceptance state do not blur across review cycles. This prevents recursive learning from becoming self-reported progress. The standing loop is intentionally indefinite at the heartbeat level and finite at the cycle level. `research/recursive-improvement-loop.md` controls that boundary: watch, harvest, adjudicate, implement, verify, publish, resubmit, or switch to benchmark/converged/blocked/observe mode. A new outside review cannot be launched merely because a heartbeat fired; accepted findings must first be integrated and made publicly visible. Active research has its own watchdog boundary. If the external surface still shows stop-control, the operator records only material progress changes: plan title, step completion, active step, status text, search/source count, completed-report metadata, error/auth state, or stop-control availability. Unchanged active research inside the normal interval produces no repo edits. Unchanged active research beyond the stall threshold produces a possible-stall note, not a relaunch. --- ## Adapter Algebra Adapters are composable. Hybrid domains assign axes to adapters rather than pretending one adapter controls everything. ```text adapter_mix = { strategic: ["incentives", "leverage", "market moves"], scientific: ["measurement", "causality", "mechanism"], formal: ["rules", "proof", "countermodel"], normative: ["stakes", "legitimacy", "plural values"] } ``` Do not fake numeric precision. State which adapter controls which axes. --- ## Object Model Minimum object families: - `ArtifactIdentity`: case ID, objective head, scope hash, config hash, active evidence hashes, created-at marker, lifecycle status. - `Scene`: scope, state variables, observed variables, hidden variables, horizon. - `Subscene`: local decomposition of a scene for evidence acquisition, causal mechanism work, option decisions, proof steps, stakeholder context, or review gates. - `Frame`: frame type, semantics, adapter ownership, assumptions. - `Token`: entity, variable, evidence, claim, constraint, uncertainty, option, trigger, gate. - `Evidence`: source, span, provenance, reliability, recency, permission, content hash, retrieval marker, source owner, chain of custody, risks. - `Claim`: atomic proposition, type, support, contradiction, dependencies, confidence, status. - `Unknown`: addressable null object for missing evidence, missing measurement, unresolved conflict, missing adapter, or unobserved variable. - `KillTest`: falsifier, countermodel, measurement test, stakeholder objection, or constraint violation tied to a target object. - `Criterion`: explicit decision standard, priority, threshold, owner, and movement jobs. - `TradeoffSurface`: option comparison, criteria basis, dominant option, and reversal conditions. - `Actor`: decision owner, reviewer, affected party, adversary, authority, incentives, exposure, legitimacy basis. - `TemporalModel`: decision horizon, deadline, evidence validity windows, staleness triggers, and update cadence. - `InformationValue`: impact/cost/time/risk-reduction view of which unknown to resolve next. - `Lens`: constrained review axis with questions, examined claims, blind spots, and contention. - `CouncilResult`: coverage, missing lenses, dissent, and decision effect. - `ReviewLog`: auditable gate decision linked to a reviewer actor. - `ReviewCycle`: recursive critique state: source review, round, accepted/rejected/unresolved findings, implemented commits, and stop reason. - `BenchmarkTrace`: benchmark execution boundary: suite, cases, arms, model families, superiority readiness, and diagnostics. - `ReviewSidecar`: recursive-review control boundary: inspected surfaces, source and execution policy, evidence classes, ranked backlog, convergence gate, benchmark handoff, and final mode decision. - `Edge`: typed relation between tokens or claims with evidence and confidence. - `Loop`: recurrent dependency, polarity, delay, gain, control points, observable cues, risk. - `OptionMove`: action/query/proof/intervention, preconditions, reversibility, risks. - `Trigger`: condition, affected objects, transition class. - `Gate`: condition, reviewer, required decision, release status. Executable artifacts: - `schemas/ofone.schema.json`: profile dispatcher for Micro, Map, and Audit schemas. - `schemas/ofone.base.schema.json`: shared object definitions. - `scripts/ofone-validate.mjs`: JSON Schema plus semantic graph validator and closure reporter. - `scripts/ofone-test.mjs`: validator regression harness for valid examples and invalid fixtures. - `scripts/ofone-render.mjs`: decision-native renderer from internal map to Micro, Map, or Audit answer. - `scripts/ofone-patch.mjs`: patch report helper for dependency closure from changed object IDs. - `examples/*.json`: strategy, scientific mechanism, formal proof-search, and hybrid policy examples. --- ## Causal Frame Mechanics Every Map or Audit run should identify: - state variables - observed variables - hidden variables - causal or logical edges - control or query channels - feedback loops - constraints and invariants - regime assumptions - failure modes - counterfactuals or kill tests Loop classes: - reinforcing - balancing - measurement - incentive - learning - contradiction - review - deception - regime Each loop carries polarity, delay, gain, control points, observable cues, and a named failure mode. --- ## Confidence Basis Confidence remains ordinal. Each confidence judgment identifies: - provenance strength - source independence - recency - mechanism fit - contradiction load - hidden-variable risk - adversarial risk - adapter fit This avoids fake decimal precision while making the basis inspectable. --- ## Dependency Closure Updates propagate through typed references: ```text new evidence -> affected claims -> affected edges -> affected loops -> affected options -> affected criteria / tradeoff / temporal / lens / review objects -> affected decision rendering node ``` Unknowns participate in closure. If `U1` blocks `O1` and `R1`, resolving `U1` is a patch event for the option and the visible rendering. Boundary, objective, criteria, adapter projection, ontology mapping, or regime changes bypass patching and become scoped reruns or trunk rewrites. --- ## Idempotency OfOne defines idempotency mechanically: ```text same objective head + same scope hash + same config hash + same active evidence hashes + same active trigger state = no-op ``` Transition classes: - `no_op`: identity tuple unchanged. - `patch`: leaf objects changed inside validated dependency closure. - `scoped_rerun`: affected subgraph, adapter section, or lens set must be recomputed. - `trunk_rewrite`: boundary, objective, evaluation criteria, ontology mapping, adapter projection, or regime assumptions changed. - `human_review`: risk, dissent, provenance, or consequence threshold requires approval. --- ## Human Gates Human review is required for legal, medical, financial, safety, compliance, public-policy, rights, employment, education, health, money, physical safety, reputation, trunk rewrites, external research packs, high-severity dissent, low-provenance evidence in high-consequence claims, adapter overrides, or audit deletion/backfill. --- ## Validator Before a decision rendering is final, OfOne runs JSON Schema validation, then semantic graph validation. The artifact may include `validator_result`, but the validator computes pass/fail and can write the computed result back into the artifact. v0.5 validation also checks artifact identity hashes, criterion ownership, tradeoff-surface references, temporal validity windows, information-value coverage for rendering-blocking unknowns, lens-axis coverage, council contention, review-log coverage for approved Audit gates, and confidence consistency against hidden-variable or contradiction load. - every emitted item performs a movement-economy job - evidence, claims, graph, and rendering stay separate - edge relations are legal for their endpoint object types - evidence carries stable source identity and custody fields - subscenes reference valid scene, frame, and token IDs - unknowns are explicit objects when missing evidence or measurement blocks movement - kill tests target and falsify existing objects - strong claims list support, contradiction or gap, confidence basis, and failure mode - causal edges, hidden variables, loop physics, and regime assumptions are explicit enough for the chosen mode - options do not hide dependence on disputed claims - adapter projection fits the domain language - update triggers and human gates are present - trigger closure reaches the rendering node when the final answer depends on changed objects - output size fits the selected mode --- ## Final Framing ```text OfOne is a typed causal-geometry compiler for high-fidelity decision maps. ```