Dataset First Semantic Embodiment - TsatsuAmable/nemosyne GitHub Wiki

Generated reference. This page is published automatically from the version-controlled repository. Canonical source: docs/architecture/MONETA_DATASET_FIRST_SEMANTIC_EMBODIMENT.md. Do not edit this wiki page directly.

Moneta Dataset-First, Dual-Mode Semantic Embodiment Architecture

Status: proposed MINIMIZE / REMOVE-FIRST scope under RFC 0012; retained ADAPT contracts apply only within that scope Date: 4 October 2026 Reviewed base: 22c64025 (origin/main, including FM2 NIL action bindings and the prior design) Governing vision: ../Nemosyne_Definitive_Vision_and_Roadmap.md Live sequencing authority: ../ROADMAP.md Canonical Full-Moneta target: FULL_MONETA_SEMANTIC_EMBODIMENT_ARCHITECTURE_PLAN.md Accepted identity/admission amendments: A27_0_AUTHORITY_VERSION_DECISION.md, A27_1_DUAL_EPISTEMIC_REASSESSMENT.md, ../rfcs/0010-moneta-semantic-snapshot-and-forma-admission.md, ../rfcs/0011-dual-epistemic-embodiment-and-preservation.md Design inputs: ../work/DATASET_LEVEL_SPATIAL_EMBODIMENT_ARCHITECTURE.md, ../audits/MONETA_DATASET_FIRST_SEMANTIC_EMBODIMENT_AUDIT_2026-09-18.md, ../roadmap/P1_FULL_MONETA_INCREMENTAL_CAPABILITY_PLAN.md, ../review-plans/FULL_MONETA_COMPLETENESS_AUDIT_2026-10-04.md, ../roadmap/P1_R_SEMANTIC_EMBODIMENT_CONVERGENCE.md, ../roadmap/P1_MCR_COMPOSITIONAL_REPRESENTATION_EXPANSION.md, and QCA0-QCA4 evidence under ../work/quest-compute/ Role: detailed dataset-first design subordinate to the governing vision and canonical Full-Moneta target. It reconciles the dataset-level critique, dual-mode Moneta, human-grounded explainability/learning, compositional search and QCA performance evidence into one implementation-ready design proposal.

1. Decision

Scope revision: RFC 0012 — Minimum Nemosyne architecture supersedes this proposal's earlier assumption that every DSE extension belongs in the initial product. Use four product subsystems: data/evidence, Investigation, representation compiler and presentation/input. The RFC's removal/merge/defer/externalize table and reduced sequence control this proposal's scope; the detailed contracts below specify behavior only when the corresponding capability is commissioned. Owner acceptance remains required for the proposed architecture changes. Existing accepted RFCs remain authoritative until explicitly amended.

Adopt the dataset-first objective by adapting the existing semantic snapshot -> RepresentationGraph -> Forma -> perceptual plan -> runtime spine, not by creating a competing analytical authority, a universal DATASET renderer primitive, or a decorative dataset container.

The Dataset becomes the stable semantic subject and independently addressable aggregate root. Rust/WASM supplies governed semantic structures beneath that root. Moneta proposes which structures, perspectives and relationships deserve perceptual existence. Forma admits and compiles those proposals into bounded, explainable perceptual phenotypes. Runtime objects remain disposable projections.

The default overview must no longer scale its live scene with source row count. Point-per-observation presentation is reserved for explicit observation intent or bounded semantic refinement.

The design is dual-mode from the beginning:

  1. Claim-bearing mode compiles only authoritative grounded bindings.
  2. Exploratory-abductive mode may combine the same grounded anchor with explicitly typed imputed, hypothesized or counterfactual proposals.

Both modes use one admission/compiler boundary and one explanation/critique loop. They do not create two Monetas.

1.1 Why ADAPT

  • ADOPT unchanged is rejected because current contracts do not make the Dataset a stable product-level spatial subject, do not provide a complete multi-dataset composition boundary, and do not connect the existing critique/learning contracts through a complete product loop.
  • REJECT is rejected because the governing vision explicitly requires dataset-scale decoupling, semantic LOD, compositional Moneta, reversible explanation and human-grounded learning. QCA0 shows that the observation-first fallback is also physically unsuitable on the reference device.
  • ADAPT preserves the correct existing authorities and landed Full-Moneta machinery while adding the missing dataset subject, working-set, comparison and product-feedback contracts.

2. High-risk adversarial contract

Invariant

A dataset-first representation may reduce transfer, main-thread, scene-object and GPU cost only by compiling a bounded intent-dependent semantic working set. It must not weaken analytical authority, epistemic status, mandatory meaning, provenance, explanation, reversible detail, critique attribution or replay identity.

Authority and production path

canonical Rust dataset capability
  -> governed analytical outputs and evidence
  -> decision-independent grounded SemanticSnapshot
  -> stable Dataset subject + semantic structures
  + committed Investigation purpose / intent / perspective
  -> fixed obligations
  -> RepresentationGraph proposal
  -> one Forma admission/compiler
  -> bounded PerceptualEmbodimentPlan
  -> SpatialEmbodimentPlan backend
  -> RepresentationSurface / desktop / WebXR

The production entry begins at the existing dataset-load/Atlas authority and ends at the real representation surface. A helper-only demonstration is not production evidence.

Principal failure modes

  1. A dataset root is only a label, box or decorative shell around row marks.
  2. A supposedly bounded overview still scans or materializes all rows in TypeScript.
  3. Graph/search code invents semantic dependencies because a signature or row count makes a primitive look plausible.
  4. Resource adaptation removes mandatory meaning or conjectural disclosure.
  5. A model, genome, critique or archive mints evidence or changes dataset identity.
  6. Two datasets collide because one graph/plan silently assumes one global fingerprint.
  7. Human critique acceptance is mistaken for comprehension or an automatic learning label.
  8. Search improves an internal objective while producing no human-useful or device-feasible representation.
  9. A rich phenotype is treated as additional discovered meaning.
  10. Browser/unit evidence is used to claim Quest cadence, comfort or human comprehension.

Cheapest falsifying evidence

  • Delete row access from the overview path; the dataset body must still compile and render.
  • Hold semantic evidence constant while increasing source N; overview plan-element and rendered-object counts remain bounded.
  • Remove required evidence or change epistemic purpose; admission refuses rather than inventing a substitute.
  • Change only a binding, perspective, conjectural status, resource budget or dataset fingerprint; the appropriate derived identities change while unrelated analytical identity stays stable.
  • Select any data-bearing perceptual property and reverse-resolve it to exact binding, primitive, semantic property, evidence/proposal, dataset, context, versions and limitations.
  • Compile two independently identified datasets simultaneously; selection, explanation, refinement and replay remain unambiguous.
  • Reintroduce point-per-row overview, unauthored learning labels or automatic prior updates; focused tests must fail.
  • Compare the final exact-head candidate with the custody-sealed QCA0 baseline on physical Quest hardware.

Non-goals

  • No new statistical estimator or TypeScript analytical fallback.
  • No universal metaphor claimed to work for every dataset or investigator.
  • No claim that VR, 3D or dataset embodiment is inherently superior to simpler representations.
  • No immediate world model.
  • No automatic learning from clicks, dwell, adoption, successful compilation or critique acceptance.
  • No requirement to implement every semantic family, modality, evolutionary operator or research question in one PR.

3. Problem statement and QCA evidence

Moneta's primary subject is a dataset, not a cloud of observations. Existing governed aggregate, distribution, density, cluster and relationship payloads already prove that Rust-owned bounded semantic embodiments are feasible. Existing semantic-detail machinery proves that exact observations can remain available without becoming the universal rendering substrate.

The remaining defect is the product-level ontology and integration path. A user can still encounter a row-addressable world whose apparent dataset representation is largely one mark per observation. The current RepresentationGraph vocabulary contains useful techniques but no explicit stable dataset subject. The current static Forma slice maps semantic nodes into spatial elements but does not yet establish a meaningful independently addressable dataset body. Existing System-1, search, genome, feedback and knowledge contracts are not yet joined around that subject through the real product path.

The QCA0 record is operator-accepted engineering evidence with a custody-sealed physical Quest 3S manifest. Its governed PERF-04/PERF-05 result remains INVALID_RUN because the summary emitted deviceTarget: UNDECLARED. The original load-duration measurements timed dispatch and are excluded. A corrected declared-device, completion-timed physical control run is mandatory for governed qualification; comparison with this historical run cannot change its disposition.

The usable QCA0 observations make the cost visible on Quest 3S at 100,000 rows:

  • CPU frame time was approximately 224 ms p50 and 244 ms p95;
  • XR frame time was approximately 249 ms p95;
  • effective cadence was approximately 4.4 frames per second;
  • the scene carried approximately 2.4 million triangles and 100,000 rendered nodes;
  • roughly 28 draw calls showed that draw-call reduction alone was not the governing bottleneck;
  • a governor LOD scale of 0.4 did not reduce representation cardinality;
  • heap/WASM capacity remained broadly flat during the five-minute run.

This is evidence of a cardinality/render correlation and an unsuitable default workload, not complete causal attribution. QCA4's marker-geometry experiments may reduce triangle cost, but substituting cheaper glyphs cannot correct the observation-first product abstraction. QCA2 still owes corrected production timing and attribution. Physical comparison remains mandatory.

The architectural performance target is therefore:

source N
  -> Rust-owned evidence and analytical hierarchy
  -> bounded semantic structures
  -> intent-dependent semantic working set
  -> bounded perceptual plan
  -> small live scene projection

4. Goals and success definition

4.1 Product goals

  1. A dataset is independently addressable, selectable, explainable and revisitable even when no observation glyph is visible.
  2. The first useful view foregrounds governed dataset structure rather than row cardinality.
  3. An investigator can move dataset -> structure -> subset -> observation and back without losing identity, context or provenance.
  4. The same grounded dataset can support different perspectives and perceptual phenotypes without changing analytical truth.
  5. Exploratory Moneta can propose missing or alternative structure while making its conjectural status impossible to erase accidentally.
  6. Every representation can be inspected, critiqued, varied and compared; qualified human evidence can improve future proposals under governance.
  7. Multiple datasets can coexist without identity/provenance collision and can be compared only through explicit, evidence-compatible relations.

4.2 Performance goals

  1. Default dataset overview transfer, JavaScript traversal, plan size and rendered-object count are bounded by representation complexity rather than source N.
  2. No non-observation renderer receives raw rows.
  3. Resource pressure reduces optional semantic richness, background work and cold residency before mandatory meaning, interaction acknowledgement or conjectural disclosure.
  4. Stronger hardware expands semantic depth, comparisons, prefetch and resident alternatives rather than merely recovering avoidable baseline inefficiency.
  5. The implemented path demonstrates material physical Quest improvement against QCA0 without a scientific or interaction regression.

4.3 Scientific and human goals

  1. Candidate names match the mathematics actually supplied.
  2. Every data-bearing degree of freedom has a typed, reversible semantic binding.
  3. Human recoverability is measured rather than inferred from compiler validity.
  4. Critique, preference, comprehension, task outcome and discovery outcome remain distinct evidence classes.
  5. A representation that is fast but misleading, unexplainable or cognitively ineffective does not succeed.

4.4 Programme success

FM-DSE succeeds only when all of the following are true:

  • two materially different datasets produce distinct, meaningful dataset-level bodies through the production path;
  • removing observation glyphs leaves each body useful and addressable;
  • at least two independently governed semantic phenomena can coexist within one dataset body;
  • default overview plan/render cardinality is bounded independently of source N;
  • exact observation refinement is deliberate, bounded and reversible;
  • grounded and exploratory purposes enforce different admissible binding sets;
  • every embodied property supports structural reverse explanation;
  • a confirmed human critique can instantiate an alternative without mutating the original;
  • retained feedback is attributable, scoped and incapable of self-promoting a production prior;
  • clean replay or safe materialized restoration preserves the experienced representation and epistemic envelope;
  • physical Quest evidence demonstrates improved frame/resource behavior;
  • human evidence demonstrates that the dataset body communicates at least one intended semantic distinction better than, or without material loss relative to, the selected baseline.

5. Terminology and four-layer model

The phrase "dataset representation primitive" is ambiguous. This design separates four layers.

Layer Purpose Examples Authority
Dataset subject Stable identity and root of the represented world dataset fingerprint, root subject, coverage/refusal state Dataset/Rust identity projected through L0
Semantic structures What the evidence supports about the dataset population, aggregate, distribution, density, cluster, topology, relationship, temporal structure, uncertainty, anomaly, missingness Rust/WASM analytical/evidence authority
Representation primitives How supported meanings are selected and composed AGGREGATION, DISTRIBUTION, DENSITY, CLUSTER, GRAPH, HIERARCHY, TEMPORAL, UNCERTAINTY, COMPARISON, DETAIL_EXPANSION Moneta / RepresentationGraph under evidence gates
Perceptual/spatial primitives How admitted meanings become perceivable surface, volume, field, trajectory, graph, glyph, label, colour, texture, motion, audio, interaction Deterministic Forma compiler and qualified backends

POINT_IDENTITY remains a legitimate representation primitive for explicit observation-level intent and bounded detail. A point, particle or instanced glyph may also be an internal spatial technique beneath a dataset-level semantic object, but does not make the semantic object observation-level by itself.

DATASET is not added to RepresentationPrimitiveKind. Dataset is the subject to which representation primitives apply. Making it another primitive would conflate subject and technique and invite decorative shells.

6. Stable Dataset subject and semantic body

6.1 Subject identity

The first tranche should derive subject identity using existing contracts. RFC 0012 removes a new subject object/store as a default requirement. If DSE0 proves a missing identity/replay invariant, the smallest candidate extension is a closed DatasetSubjectRefV1 concept:

DatasetSubjectRefV1
  schemaVersion
  subjectId = H('dataset-subject-v1', { analyticalDatasetFingerprint })
  analyticalDatasetFingerprint

It contains no new analytical fact. It is a deterministic identity projection over the canonical dataset fingerprint. Snapshot, decision, perspective, plan, device and runtime state do not enter the subject identity.

The exact public-format placement requires the DSE0 schema/RFC review. Two implementations are plausible:

  1. derive the subject reference beside SemanticSnapshotV1 and bind it in RepresentationGraph V2; or
  2. introduce a future SemanticSnapshotV2 with a mandatory subject field while preserving V1 unchanged.

The first is the smaller migration. The second may be cleaner if other required semantic-entity/refinement fields also force a version change. Further contract research is required before freezing the wire format. The objective is fixed: one stable dataset subject, no new truth authority, no optional mandatory-in-practice field, and exact legacy replay.

6.2 Meaningful body requirement

A rendered dataset body consists of:

  • one independently addressable dataset root;
  • one or more evidence-bearing semantic structures, or a typed unavailable/refused state;
  • explicit relations among those structures;
  • explanation and refinement affordances;
  • optional qualified perceptual framing that carries no hidden data meaning.

A root plus label, bounding box or shell is insufficient. Admission must reject a claimed meaningful dataset body whose only content is decorative framing or row marks renamed as a dataset.

Example:

Dataset subject
  CONTAINS aggregate/population structure
  CONTAINS distribution structure
  COORDINATES_WITH relationship topology
  OVERLAY uncertainty disclosure
  DETAIL_OF bounded observation refinement

6.3 Semantic working set

The semantic working set is the subset of the dataset's supported semantic world admitted for the current context and resident in the current perceptual plan. It is controlled by explicit budgets such as:

  • maximum semantic objects;
  • maximum plan elements;
  • maximum observation marks;
  • maximum labels and interaction proxies;
  • maximum resident CPU/GPU bytes;
  • maximum concurrent refinement requests;
  • required frame/resource envelope;
  • mandatory obligations that no budget may remove.

Budgets are inputs to variant selection, not permission to reinterpret truth. When no variant preserves obligations, admission or adoption refuses/narrows scope.

7. Dual-mode Moneta

Dual-mode Moneta is an epistemic contract, not the System-1/System-2 distinction and not a renderer flag.

7.1 Independent axes

Axis Values Meaning
Epistemic purpose CLAIM_BEARING, EXPLORATORY_ABDUCTION What kinds of material assertion may enter the representation
Execution permission PRODUCTION, STUDY_ONLY, REFUSED Whether the admitted result may execute under current authorization
Generation regime frozen/pinned, adaptive Whether the proposal generator may change over time

These axes must remain independent. An exploratory representation may be ordinary product use. A frozen exploratory treatment may be a study. PRODUCTION never means scientifically established.

7.2 Claim-bearing path

Dataset
  -> Rust/WASM evidence
  -> SemanticSnapshotV1
  -> GROUNDED bindings only
  -> admitted RepresentationGraph / Forma plan
  -> reproducible dataset body

System-1, System-2 and search may propose representation choices but cannot invent material semantic content. Missing required evidence produces refusal, not a plausible substitute.

7.3 Exploratory-abductive path

grounded SemanticSnapshotV1 -----------------------+
                                                    |
                   System-1 / System-2 / search     |
                                                    v
                                      ConjecturalProposalV1
                                                    |
                       grounded + conjectural Forma bindings
                                                    |
                                      exploratory dataset body

Conjectural properties and relations remain IMPUTED, HYPOTHESIZED or COUNTERFACTUAL at property/relation granularity. Their assumptions, dependencies, generator provenance, uncertainty availability, supporting evidence and contradicting evidence remain inspectable.

Promotion requires a new analytical/evidence event and new decision lineage. Geometry, human preference, model score or repeated use cannot relabel a conjecture as grounded.

7.4 Mode transition

Changing epistemic purpose creates a new committed Investigation context/DAG revision, revokes pending adoption from the previous context and preserves the old branch. The grounded snapshot may remain identical. Road Not Taken, Memory Palace, TechnoCore and Farcaster operate over this shared lineage rather than creating separate stores.

8. Representation generation and proposal sources

8.1 Deterministic reference composer

Every supported context needs a non-stochastic reference composition generated from exact governed semantic availability and fixed obligations. It is the baseline against which learned, synthesized and evolved candidates compete.

The current reference composer is not sufficient for dataset-first authority because signature heuristics can introduce primitives and dependency names without resolving claim-complete evidence. A row-count threshold must not imply density, and a label such as kde_estimate must not substitute for a governed estimator.

The replacement composer must:

  • consume semantic snapshot/source identities rather than raw signature plausibility alone;
  • bind every primitive to exact semantic properties and evidence/proposal references;
  • produce a stable dataset root and bounded working set;
  • emit typed refusal when required meaning is unavailable;
  • avoid raw renderer parameters except through registered qualified recipes;
  • provide deterministic identity and reverse explanation.

8.2 System-1

System-1 is a fast advisory proposal source. It may:

  • retrieve relevant Forma cases/templates;
  • rank registered representation primitives and bindings;
  • propose bounded compositions and resolution variants;
  • seed/prune search order;
  • predict which alternatives deserve expensive evaluation;
  • abstain.

It consumes compact governed features and pinned knowledge identities, never raw rows. It is not the durable store of metaphor knowledge and cannot admit a result, weaken obligations, choose evidence policy or emit unregistered arbitrary renderer parameters.

The landed FM5/FM6 machinery supplies a foundation. Dataset-first qualification still must show measurable value for dataset-body proposals over deterministic retrieval, including held-out dataset groups, abstention and target-device cost.

8.3 System-2 / compositional Moneta

System-2 is the explicit representation reasoning layer. It may:

  • construct and compare multi-phenomenon graphs;
  • reason about information preserved/lost;
  • propose alternative perspectives and metaphors;
  • request additional governed analysis through Atlas;
  • challenge mappings and assumptions;
  • synthesize bounded candidates inside the admitted grammar;
  • expose alternatives through Road Not Taken.

System-2 is not a second evidence or admission authority. It need not call System-1; rules, researcher input, System-1, System-2 and search are peer proposal sources converging on one admission boundary.

8.4 Bounded search

Search explores a typed grammar and returns an inspectable Pareto frontier. Fitness remains vector-valued, including task relevance, information preservation, perceptual recoverability, interaction cost, resource cost, stability and explicit loss. Scientific validity is a feasibility boundary rather than a tradeable objective.

The current FM7 qualification proves bounded graph search mechanics, Pareto behavior, System-1 seeding/bypass, genome guards and replay invariance. It does not establish that the current grammar/objective functions generate human-useful dataset bodies or that loose semantic inputs are claim-complete. Dataset-first search integration therefore remains a new qualification task.

8.5 Evolutionary and genome search

Evolution begins only after deterministic dataset bodies, binding admission, multi-element runtime, explanation, refinement and replay are correct.

Permitted genes may include:

  • selected representation primitives and composition;
  • semantic hierarchy/refinement policy;
  • registered perceptual bindings and metaphor templates;
  • spatial allocation and admitted phenotype variants;
  • interaction/detail/transition policy;
  • resource and working-set policy;
  • search/evaluation strategy identifiers.

Forbidden genes include:

  • dataset fingerprint or subject identity;
  • analytical facts or evidence receipts;
  • grounded/conjectural status;
  • fixed obligations;
  • provenance or oracle answers;
  • arbitrary unregistered analytical transformations.

The external laboratory genome remains untrusted input. Its handoff materializes an ordinary candidate graph, then subjects it to the same binding, evidence, grammar, purpose, explanation and resource gates. Lineage, ancestors, siblings and failure reasons feed Road Not Taken.

Further research is required to establish mutation/recombination operators, diversity preservation, compute allocation and benchmark corpora that improve semantic capability rather than combinatorial novelty. Success means non-dominated candidates that improve attributable human/task outcomes or measured resource cost without weakening meaning, not simply more graph diversity.

8.6 World-model research

A world model remains unscheduled. Reopen only if a named dynamic-prediction or metaphor-affordance problem cannot be solved adequately by rules, cases, explicit System-2 reasoning and bounded search. The objective would be prediction of useful interaction/representation consequences, not replacement of evidence, Forma admission or the durable knowledge base.

9. Forma admission, compilation and semantic resolution

One Forma entry point receives:

  • exact semantic snapshot;
  • stable dataset subject;
  • committed context and epistemic purpose;
  • fixed obligations;
  • candidate RepresentationGraph with typed bindings;
  • any conjectural proposals;
  • pinned knowledge manifest and qualification references;
  • capability/resource envelope.

It returns an immutable purpose-scoped admitted result and bounded perceptual plans, or a typed refusal.

9.1 Binding rules

Every data-bearing property/relation is either:

  • GROUNDED, resolving to exact snapshot source/property/relation and evidence; or
  • CONJECTURAL, resolving to exact proposal element/property/relation and generator lineage.

No free-form encoding string, proximity, model score or graph label may become a material semantic binding.

9.2 Semantic resolution

Semantic LOD changes the represented abstraction, not only mesh complexity:

dataset
  -> population / distribution / topology
  -> region / substructure
  -> bounded observation subset
  -> exact datum / provenance

Each variant declares mandatory coverage, progressive detail, omissions, approximation, interaction affordances and resource estimates. Returning from detail restores the same dataset subject and overview lineage.

9.3 Performance-aware compilation

The compiler should produce backend-ready compact buffers and element descriptions from bounded semantic payloads. Renderer adapters may allocate geometry/materials, apply presentation-only transforms and build hit targets. They may not scan rows to reconstruct analytical structures.

Steady-state runtime work should use instancing, batching, culling, bounded labels, allocation reuse, multiview/foveation and GPU assistance where physical measurement justifies them. These are subordinate optimizations; no one technique is assumed to close the budget.

9.4 Cost model, caching and falsifiable benchmark

Let N be source observations, S the available semantic structure, K the admitted visible working set, E its visible relations, and C the bounded candidate population. Ingestion, fingerprinting and many exact analytical operations remain at least O(N); graph analysis may cost more. Dataset-first design does not make those operations constant-time. The intended steady overview transfer, compilation, hit-target storage and scene traversal are O(K + E), with explicit caps on K, E and label/channel counts. Search cost depends on C, grammar depth and evaluations, with a wall-time/work budget. GPU cost additionally depends on visible pixels, overdraw, material/channel complexity and geometry; bounded element count alone cannot guarantee cadence.

The reference pipeline must acquire a bounded hierarchy cut through Rust-owned queries. Materializing all S nodes in TypeScript and then selecting K would merely move the bottleneck. Retain canonical analytical caches under dataset fingerprint, method/parameters and evidence/profile identities; representation caches additionally bind snapshot, context, obligations, qualification/knowledge manifest and generator versions. Compiled variants additionally bind the exact resource policy and backend. A cache hit never bypasses current evidence disposition or activation checks. Purpose, population or evidence changes invalidate affected entries; a presentation-only change may reuse the authoritative snapshot. Bound cache residency, work queues and concurrent compilation; cancel stale requests and account for overlap peaks during replacement. Incremental analytical algorithms require separate Rust qualification before reuse; this design does not assume they already exist.

DSE0 must freeze a benchmark protocol before optimization:

  • Compare the recorded QCA0 reference, the current low-triangle observation path, and the proposed dataset-body path on the same fixtures/tasks. Preserve legacy observation mode as an explicit control/detail choice; unsupported dataset-body compilation yields a typed unavailable/refusal rather than silently returning to an unbounded observation scene.
  • Vary N with approximately fixed semantic information, then separately vary semantic complexity at fixed N. Include rare exceptions, missingness, uncertainty and a high-complexity case whose required meaning cannot fit; refusal is a valid outcome and must be measured.
  • Measure cold ingest/analysis, first useful view, warm recomposition, steady frames, refinement, return, cancellation and eviction/rebuild separately. Record CPU/XR frame distributions, GPU timing where supported, transfer bytes, scene/interaction counts, geometry/overdraw proxies, peak residency, search work, and total energy/thermal behavior where available. Missing GPU timing stays unavailable.
  • Use repeated, counterbalanced physical runs with pinned device model/browser/build, refresh policy, thermal conditions, datasets, intent and evidence. QCA0 is a historical anchor; rerun the controls when environment drift prevents a fair comparison. Quest 3 and Quest 3S results remain separate.
  • Predeclare acceptable task-information loss, comprehension/error criteria, frame/resource budgets, a practically meaningful improvement margin and analysis/uncertainty protocol. DSE0 must select the numeric thresholds from the target refresh policy and pilot variance; this proposal does not invent them. A flatter warm cost curve fails the overall hypothesis if cold analysis, search, refinement, cognitive cost or peak memory erases the user-visible benefit.

Success requires both bounded complexity at fixed task information and measured task-equivalent physical benefit. If either fails, revise the working-set policy or reject the performance claim.

10. Structural explanation, critique and human-grounded learning

All semantic bodies and metaphors must remain explainable so humans can understand, challenge and teach the mappings.

10.1 Reverse explanation chain

Every selectable data-bearing property must reverse-resolve:

runtime object / property
  -> perceptual plan element and channel
  -> typed Forma binding
  -> RepresentationGraph primitive and relation
  -> semantic object / property
  -> grounded evidence OR conjectural proposal
  -> Dataset subject
  -> committed purpose / intent / perspective
  -> limitations, uncertainty, information loss
  -> compiler, model, template, knowledge and policy versions

For a compound metaphor, TechnoCore must also explain why the metaphor/composition was proposed, which semantics justify it, which forbidden implications were guarded against, which alternatives were considered and why the current candidate survived admission/selection.

Generated prose may summarize this graph but cannot replace it.

10.2 Typed critique

The current feedback contract is a useful minimum but remains too loose for dataset-first use: proposed bindings are strings, phenotype alternatives are narrow, attribution can default, and composition/perspective changes are not a closed grammar.

A versioned extension should support exact targets and typed proposals such as:

  • BINDING_CHANGE;
  • PERSPECTIVE_CHANGE;
  • COMPOSITION_CHANGE;
  • RESOLUTION_CHANGE;
  • INTERACTION_CHANGE;
  • ACCESSIBILITY_CHANGE;
  • EPISTEMIC_CHALLENGE;
  • REQUEST_ADDITIONAL_ANALYSIS.

Every confirmed critique binds exact plan, slice, context, dataset subject, semantic object, binding/primitive and exposure/treatment identity plus attributable author and scope. An LLM may parse a draft; the human must confirm it.

10.3 Alternative instantiation

inspect through TechnoCore
  -> target exact property/binding/composition
  -> propose and confirm critique
  -> compile an alternative through Road Not Taken
  -> compare/use alternatives under synchronized semantic selection
  -> accept/reject/retain as an explicit branch

The original representation remains immutable and replayable. Accepting an alternative is not evidence that its intended meaning was understood.

10.4 Human evidence classes

Keep separate:

  • critique/proposal;
  • mapping preference;
  • perceived meaning;
  • task-comprehension outcome;
  • misleading implication;
  • accessibility/comfort issue;
  • discovery outcome;
  • study result.

Passive telemetry, clicks, dwell and system defaults may inform friction/resource research but cannot self-label comprehension or qualification.

10.5 Learning loop

exact representation exposure
  -> attributable critique and meaning judgment
  -> scoped evidence/study routing
  -> curated case with contraindications and limitations
  -> explicit promotion decision
  -> durable Forma knowledge revision
  -> optional System-1/search prior update
  -> held-out evaluation and rollback

Earlier decisions and exposures are never rewritten when knowledge evolves. Research Mode freezes relevant knowledge/model/search identity. Learned priors may order proposals; they cannot authorize meaning or evidence.

Further research is required to determine which critique and meaning-judgment protocols produce reliable learning signals across domains, expertise levels and accessibility needs. The goal is improved held-out proposal coverage, comprehension/task outcomes and reduced search cost with calibrated abstention—not maximizing user agreement.

11. Multi-dataset composition

Current core graphs/plans are primarily bound to one dataset fingerprint. A first-class dataset design must support simultaneous comparison without weakening that invariant.

The recommended direction is independently admitted per-dataset semantic worlds/plans composed by an Investigation-owned comparison scene:

Dataset A subject -> admitted plan A ----+
                                      comparison context -> coordinated scene
Dataset B subject -> admitted plan B ----+

Cross-dataset relations require explicit compatibility/evidence and cannot be inferred from spatial proximity. Selection, explanation, detail and replay retain the originating dataset subject.

Further schema research is required to decide whether cross-dataset relations belong in RepresentationGraph V2, a bounded scene-composition contract or an Investigation comparison artifact. The objective is fixed: independent fingerprints/receipts, explicit relation authority, simultaneous addressability and no merged shadow dataset.

12. Persistence, replay and lifecycle

Grounded claim-bearing results retain exact regeneration identity. Exploratory results also retain a materialized static capture so an evolving generator cannot erase what the investigator experienced.

Required durable state includes:

  • dataset subject and snapshot references;
  • context/purpose/perspective and obligations;
  • selected/rejected RepresentationGraph candidates and search lineage;
  • grounded/conjectural bindings;
  • exact plan/variant and reverse explanations;
  • generator/model/knowledge/policy identities;
  • critiques, judgments, annotations and alternatives;
  • static spatial organization/viewpoint where required for historical inspection;
  • explicit limitations and current-disposition behavior.

Runtime objects, handles and GPU resources never become durability authorities. ACTIVE/WARM/COLD/EVICTED lifecycle may remove presentation resources while semantic state remains reconstructable. Historical inspection capability remains distinct from fresh scientific execution/adoption capability.

13. Current-state gap analysis

Capability Current state Required change
Stable Dataset subject Fingerprint exists; DATASET semantic kind exists in a decision-coupled graph; no complete product-level root contract Freeze deterministic subject identity and bind it through graph, plan, runtime, explanation and persistence
Governed dataset structures Aggregate/distribution/density/cluster/graph family paths exist Normalize/combine them beneath one subject without replacing family-specific Rust builders
Default bounded overview QCA0 exercised 100k row-addressable marks Make dataset structure the default and observation marks explicit detail
Semantic working set Resource/resolution brokers exist Bind budgets to semantic objects/obligations and prove N-independent overview cardinality
Compositional runtime Multi-element Forma/runtime machinery landed; compatibility path remains Qualify meaningful dataset-root composition rather than generic multi-element rendering
Deterministic reference composition FM7 reference composer exists Remove heuristic semantic invention; bind exact semantic/evidence availability
System-1 Advisory contracts, influence and FM6 knowledge qualification exist Qualify dataset-body proposal value, abstention and device cost
Search/genome FM7 grammar/Pareto/genome mechanics qualified Tighten typed bindings/evidence; demonstrate useful dataset bodies and human/device value
Dual epistemic core Grounded/conjectural contracts landed Bind both modes to the Dataset subject and finish product mode/actions/visible disclosure
Reverse explanation Static Forma traces exist Extend to dataset root, composition, perspective, alternatives and all admitted channels
Human critique V1 critique/judgment records exist Require attribution; typed critique grammar; real TechnoCore -> alternative -> comparison path
Learning FM6 promotion/holdout battery is qualified Acquire attributable dataset-body evidence; prevent automatic/self-labelled updates
Multi-dataset comparison Single-fingerprint graph/plan assumption dominates Define bounded independent-world scene composition and cross-dataset relation authority
Persistence V4 static replay exists for current slice Capture dataset subject, alternatives, search/critique lineage and multi-dataset identity
Physical performance QCA0 baseline captured; QCA4 experiments open Run controlled dataset-first physical comparison and attribution before claims
Human semantic value Architecture requires it Conduct preregistered known-structure and investigator studies; keep claim open meanwhile

13.1 Full Moneta FM0-FM8 integration and gap map

FM-DSE is a cross-cutting application of Full Moneta to the Dataset subject. It does not create FM9, renumber the ladder, reopen mechanically verified tranches, or allow a later FM result to satisfy an earlier authority or product-evidence gate. The current status below is inherited from the live roadmap; the final column states the additional dataset-first proof rather than downgrading already obtained bounded evidence.

Increment Retained contract and current evidence Dataset-first dependency or gap Dataset-first success condition
FM0 — Trustworthy Governed Rust evidence, receipts, identity and clean replay remain the only claim-bearing base; finite evidence closure is still active. Dataset subject and every grounded body property must resolve to the exact snapshot/evidence authority. DSE must not proceed around missing, retracted, wrong-profile or incompatible evidence. A clean replay recovers the same subject, supported semantic body and limitations without a second analytical authority or identity substitution.
FM1 — Question/Perspective-Aware Core context/perspective identity and adoption guards are landed; real researcher-facing intent/perspective actions remain open. Committed purpose, question, task and perspective select the semantic working set. A wording change cannot rewrite facts, and a material purpose change cannot reuse stale authorization. Two material intents over one dataset yield appropriately different bodies or an explicit no-change/refusal through the production action path, with exact context explanation and replay.
FM2 — Alternative-Aware Atomic branch/revisit foundations and NIL alternative/intent bindings (#988) are landed; investigator-visible surface qualification remains open. Dataset-body candidates, critique-derived alternatives and search siblings need shared semantic anchors, immutable originals and promotability guards. A researcher compares and branches between meaningfully different evidence-bound bodies; display, preference or critique cannot activate an ineligible candidate or become truth.
FM3 — Compositional Forma Static Forma, multi-element runtime, reverse explanation and V4 replay foundations are landed; semantic comprehension/usefulness remains open. FM3 is the main embodiment seam: Dataset subject, two or more governed phenomena, typed relations and claim-complete bindings must compile as one meaningful body. A production-reachable body remains meaningful without row glyphs, coordinates multiple phenomena, explains every data-bearing property and survives static replay.
FM4 — Resolution-Adaptive Final-content identity, budget/generation refusal and selected-graph fidelity are landed; cross-device qualification remains open. Semantic working-set selection—not polygon LOD—must bound overview cost. Refinement/collapse, eviction and reconstruction preserve obligations, orientation, epistemic disclosure and selection. Quest and desktop receive different admitted semantic richness but the same core scientific story; source-N growth does not force overview cardinality growth.
FM5 — Intuitive/System-1 Advisory recording, disclosure, bypass, replay, ABSTAIN and comparison probes are merged; measurable held-out and target-device benefit remains open. Dataset-body retrieval/ranking/seeding needs a versioned feature/artifact contract, ABSTAIN and off-critical-path execution. System-1 may reduce proposal/search cost but cannot choose truth or bypass Forma admission. A qualified dataset-body advice lane beats its deterministic/reference baseline on declared held-out value or cost, survives replay and Quest constraints, and has a truthful STOP/REVISE outcome if it does not.
FM6 — Human-Refined Generic qualification mechanics, holdout gates, frozen mode and rollback are verified at their bounded exit. Dataset-body critiques, meaning judgments and outcomes still need attributable product acquisition, exact target/binding lineage, consent/scope and non-self-labelled curation. An eligible dataset/investigator-disjoint corpus improves a declared held-out proposal or comprehension objective without conflating preference, acceptance, discovery or scientific validity.
FM7 — Search/Synthesis Grammar, Pareto, advisory seeding/bypass, genome handoff guards and replay mechanics are verified at their bounded exit. Current mechanics do not prove that candidate bindings are claim-complete or that generated dataset bodies are useful. Operators/objectives must work over admitted semantics, not facts, evidence or renderer novelty. At least one non-reference body is admissible, explainable and measurably useful or cheaper than deterministic/System-1 baselines, with lineage, rejected alternatives and ABSTAIN/STOP preserved.
FM8 — Controlled adaptation Dual-epistemic core and the FMA repair set are landed; integrated product/human/device qualification remains open. The production loop must unite FM0-FM7 around the stable Dataset subject: purpose -> proposal -> admission -> bounded embodiment -> explanation/challenge -> durable learning evidence. An end-to-end dataset investigation is spatial, skeptical, performant, recoverable and human-useful under scoped claims; Research Mode freezes all applicable state and Product Mode cannot silently strengthen claims.

Consequently, DSE0-DSE4 may build and qualify the deterministic dataset-first spine while FM5 value work remains separately gated. DSE5 consumes FM5-FM7 only after their applicable authority contracts are ready. DSE6 contributes the dataset-first physical/human evidence required by FM3, FM4 and FM8; it cannot by itself close residual FM0, FM1, FM2 or FM5 obligations.

The remaining FM5 benefit experiment is not a technical prerequisite for deterministic DSE0-DSE4. Their gates are the applicable evidence, identity, admission and product contracts. Execution still uses one authorized forward tranche unless the roadmap explicitly permits disjoint work; an old linear FM status list must not be mistaken for a requirement to repeat already merged qualification.

13.2 Existing qualification evidence to retain and extend

The historical closure reports for FM1, FM2, FM6, FM7 and FM8 are preserved as initial evidence. Their own status notices subordinate their closure claims to the later completeness audit. Use the repaired production paths and subsequent tests as regression seeds; do not repeat the original defects or treat historical green suites as current product qualification.

Existing artifact Evidence worth retaining Additional proof required for this design
tests/fm1-perspective-product.test.ts, tests/fm2-alternatives-road-not-taken.test.ts and subsequent context/atomicity repairs Context correspondence, analytical invariance, preview/branch/revisit and refusal controls Real intent and alternative actions targeting Dataset subjects, with delayed-result and replay cases
tests/fm3-4-composition-resolution.test.ts plus static Forma/adoption/V4 suites Final-content identity, invalid budget/generation refusal, selected-edge fidelity, context serialization and static restoration Dataset-body semantics, bounded hierarchy cuts, meaningful refinement and human recoverability across budgets
tests/fm5-s1-consumed-record.test.ts, disclosure, fallback, recorded-replay and abstention suites Advice influence/recording, disclosed bypass, stale/foreign replay refusal and ABSTAIN behavior Dataset-body feature/candidate coverage and device cost under the same guards
tests/fm5-advice-measurement.test.ts, fm5-advice-heldout.test.ts, fm5-advice-coverage.test.ts (R5/R7/R8) Reproducible advised-versus-bypassed probes across fixture structures/tasks and frontier coverage Their finite-delta/reproducibility assertions do not establish positive utility, human benefit or physical Quest performance; inspect measured outputs and predeclare meaningful benefit before promotion
tests/fm6-r6-learned-prior-qualification.test.ts (#980) Fixture-based evaluation gate, knowledge influence/identity, frozen mode and rollback mechanics Attributable acquired human evidence and dataset/investigator-disjoint generalization; fixture-created confirmed judgments are not a human study
tests/fm7-r7-search-synthesis-qualification.test.ts (#983) Bounded novel graph construction, Pareto mechanics, advisory seeding/bypass, genome guards and frozen digest Claim-complete dataset-body bindings plus independent task/human/resource evaluation; an objective named perceptual recoverability is not itself measured comprehension
tests/fm8-full-moneta-adaptive-intelligence.test.ts and FMA repair suites Integrated orchestration, ungrounded-input refusal, frozen state, obligations and explanations Actual investigator loop, physical interaction/cadence and human orientation; bounded internal stability scores cannot prove absence of disorientation

This is a source/evidence reconciliation, not a claim that these suites were rerun by the architecture PR. DSE plans must reference and extend the applicable existing tests, preserve their positive and negative controls, and specify which new production or human/device observation closes the residual gap.

14. Research programme

These are capability gaps, not permission to invent behavior. Each item states its objective and success criterion.

R-DSE-1 — dataset-body vocabulary and metaphor fitness

Objective: determine which evidence-bound compositions and metaphors help which tasks, domains and populations.

Method direction: known-structure corpora, conventional and deliberately misleading controls, scoped human studies, critique capture and cross-domain holdout.

Success: qualified mappings improve comprehension, error detection, recall or task efficiency over appropriate simpler baselines and record contraindications/abstention scope.

R-DSE-2 — subject and multi-dataset schema

Objective: freeze the smallest versioned contracts that provide stable subject identity, recursive refinement and independent multi-dataset comparison without duplicating truth.

Success: two datasets can coexist, compare, refine, replay and explain without identity collision or optional authority fields; legacy snapshots remain exact.

R-DSE-3 — semantic working-set policy

Objective: learn deterministic policies for selecting mandatory/progressive/optional semantic detail under device and attention budgets.

Success: bounded plans preserve obligations and human recoverability across budgets; inability to do so refuses rather than silently simplifying.

R-DSE-4 — search and evolutionary operators

Objective: identify genome/operator/objective designs that produce useful, diverse, admissible dataset bodies rather than novelty or renderer parameter churn.

Success: a reproducible search finds non-dominated candidates with measurable held-out human/task or resource benefit over deterministic/System-1 baselines while preserving lineage and abstention.

R-DSE-5 — critique-to-learning evidence

Objective: determine how confirmed critiques, meaning judgments and discovery outcomes can improve proposal priors without conflating preference, comprehension and truth.

Success: curated attributable feedback improves held-out proposal coverage/regret or comprehension outcomes, passes dataset/investigator-disjoint evaluation and supports rollback.

R-DSE-6 — dual-mode perceptual disclosure

Objective: find perceptual/interaction techniques that keep grounded and conjectural content distinguishable without making exploratory work unusable.

Success: users correctly identify epistemic status and assumptions under representative tasks; disclosure survives resolution/resource adaptation and accessibility conditions.

R-DSE-7 — Quest semantic-efficiency frontier

Objective: characterize how source N, semantic complexity, visible working set, refinement, search breadth and hardware budget interact on Quest 3S.

Success: controlled physical runs show stable bounded cardinality and material cadence/resource improvement against QCA0, with no semantic/interaction regression and honest attribution.

R-DSE-8 — adaptive-depth/shared model hypothesis

REMOVED from this programme by proposed RFC 0012. No demonstrated task requires this subsystem. Any future proposal must begin with a named problem, simpler baselines and independent evidence.

15. Implementation sequence

Each code tranche requires a fresh written implementation plan, exact-head verification and the high-risk adversarial evidence required by AGENTS.md. This architecture PR does not combine them.

RFC 0012 narrows the initial sequence to DSE0 -> DSE1 -> DSE2/DSE3, with DSE6 qualification beginning at DSE1. DSE4 and runtime DSE5 are conditional extensions. Their specifications below are retained for future use, not prerequisites for the first complete investigation loop.

DSE0 — decision, schema preflight and falsifiers

  • owner review of RFC 0012's removal-first scope and the retained ADAPT contracts;
  • source/caller/archive deletion inventory and only necessary subject identity/public-format decisions; defer multi-dataset schema;
  • production-path test that a dataset overview cannot enter whole-row rendering;
  • current-path cardinality/row-access instrumentation;
  • canonical two-dataset known-structure fixtures;
  • freeze success metrics and physical/human evidence plans.

Exit: accepted contract, no unresolved authority ambiguity, and falsifiers fail against the old forbidden behavior.

DSE1 — stable subject and grounded bounded overview

  • derive the stable Dataset subject from authoritative fingerprint;
  • build one production-reachable dataset body from at least two governed semantic structures;
  • supply a bounded deterministic composer whose every binding resolves supported semantics and evidence before admission;
  • compile a bounded working set with zero default observation marks;
  • extend reverse explanation to the dataset root and composition;
  • capture confirmed attributable critiques against exact subject/plan/binding identities, with durable replay and no automatic learning; DSE3 adds full alternative instantiation;
  • retain explicit observation-level compatibility through detail transition.

Exit: deleting row access does not break the overview; plan/render cardinality is bounded under N growth; every property reverse-resolves.

DSE2 — semantic refinement and resource variants

  • dataset -> structure -> subset -> observation -> datum traversal and return;
  • admitted semantic-resolution variants under existing resource broker/lifecycle;
  • refusal when budgets cannot preserve obligations;
  • exact selection/context identity across refine/collapse/evict/reconstruct.

Exit: semantic continuity and disclosure survive materially different budgets and lifecycle states.

DSE3 — dual-mode product loop

  • bind grounded and conjectural paths to the same Dataset subject;
  • finish visible epistemic disclosure and mode transition;
  • extend typed critique and meaning judgment targeting;
  • TechnoCore explanation -> confirmed critique -> Road Not Taken alternative -> compare/retain workflow;
  • V4 capture of dataset subject, bindings, alternatives and feedback lineage.

Exit: claim-bearing rejects conjectural material; exploratory admits and visibly preserves it; critiques create alternatives without mutation or automatic learning.

DSE4 — multi-dataset comparison

Conditional extension under RFC 0012; unscheduled pending a named blocking research task.

  • compose two independently admitted dataset worlds;
  • explicit cross-dataset comparison authority and synchronized selection;
  • independent refinement, eviction, explanation and replay.

Exit: two materially different datasets meet the minimum proof in the dataset-level critique without fingerprint/provenance collision.

DSE5 — proposal/search/evolution integration

Conditional runtime extension under RFC 0012. Offline generation/evaluation may produce governed artifacts; the default product requires no inference, search population or genome interpreter.

  • generalize the DSE1 evidence-bound deterministic composer for the broader supported grammar, replacing remaining heuristic paths before their use;
  • qualify System-1 dataset-body proposals;
  • tighten RepresentationGraph grammar/objectives/bindings for dataset subjects;
  • route System-2/search and external genomes through identical admission;
  • retain Pareto/search/evolution lineage in Road Not Taken.

Exit: at least one non-reference proposal demonstrates held-out benefit without bypass; deterministic fallback and ABSTAIN remain usable.

DSE6 — physical and human qualification

  • corrected governed physical control capture and exact-head Quest comparisons, retaining QCA0 as historical engineering evidence and controlled renderer variants;
  • corrected main-thread/Worker/GPU attribution;
  • long-session bounded-resource evidence;
  • known-structure and investigator semantic-recovery studies;
  • accessibility and dual-mode disclosure evaluation;
  • STOP/CONTINUE/REVISE decision.

Exit: physical performance and scoped human value are demonstrated at the applicable evidence tiers. Otherwise retain the truthful deterministic subset and revise/stop unsupported claims.

16. Verification matrix

Claim Required evidence
Stable subject identity deterministic/metamorphic identity tests across decisions, perspectives, plans and devices
No raw-row overview source policy plus real load -> Atlas -> Forma -> RepresentationSurface production-path test
N-independent overview same supported semantic complexity across increasing N; bounded transfer/plan/render counts
Scientific fidelity Rust reference/golden tests, claim-complete receipts and candidate/payload compatibility
Dual-mode safety purpose/binding refusal and positive controls; status survives resolution/export/replay
Explanation property selection reaches exact evidence/proposal and versions with no orphan element
Critique safety attribution/confirmation, immutable original, alternative branch, no automatic prior update
Learning value curated dataset/investigator-disjoint held-out evaluation and rollback
Search/evolution value deterministic baseline, lineage, Pareto frontier, admissibility and held-out utility/resource evidence
Multi-dataset simultaneous product path, independent identity/selection/refinement/replay
Quest performance physical exact-head QCA comparison with CPU/XR/GPU/resource attribution
Human meaning preregistered scoped comprehension/task/disclosure evidence with simpler and misleading controls

Unit tests are necessary but cannot promote product, device or human claims. Mocks may isolate failures but do not substitute for the real authority and rendering path.

17. Source reconciliation and retained decisions

This design preserves the following repository decisions:

  • the Definitive Vision's five ontologies, dual epistemic contracts and Appendix A semantic-efficiency doctrine;
  • Rust/WASM analytical authority and canonical dataset handles;
  • decision-independent grounded snapshot identity from A27-0/RFC 0010;
  • committed purpose, conjectural proposals, typed bindings and materialized preservation from A27-1/RFC 0011;
  • RepresentationGraph as the compositional hypothesis, not analytical truth;
  • Forma as one deterministic admission/compiler authority under Moneta;
  • SpatialEmbodimentPlan as a disposable spatial backend phenotype;
  • System-1 as advisory retrieval/ranking and System-2 as explicit reasoning/search;
  • deterministic composition before evolutionary exploitation;
  • durable human-grounded knowledge outside model weights;
  • critique distinct from meaning judgment and promotion;
  • Memory Palace, Road Not Taken, TechnoCore and Farcaster as views/actions over shared Investigation lineage;
  • physical Quest and attributable human evidence for device/perceptual claims.

It changes or clarifies:

  • Dataset becomes an explicit stable semantic subject rather than an implied fingerprint on a row-shaped scene;
  • dataset body means evidence-bearing composition, not container or layout;
  • the default overview has a bounded semantic working set and no observation marks unless required;
  • existing FM5-FM7 proposal/search qualification does not itself qualify dataset-body semantics;
  • explanation/critique is mandatory in the first dataset-body slice;
  • multi-dataset comparison requires a distinct bounded authority seam;
  • performance success includes semantic/cardinality correction, not only renderer micro-optimization.

18. Review disposition and next decision

This document retains bounded ADAPT contracts inside the MINIMIZE / REMOVE-FIRST proposal in RFC 0012. Owner acceptance is required before DSE0 freezes any changed contract. Acceptance does not promote FM-DSE, physical performance, human comprehension, System-1 value or evolutionary value.

After acceptance:

  1. record any required RFC/ADR amendment for the subject/multi-dataset public seams;
  2. write the DSE0/DSE1 implementation plan against fresh main;
  3. implement one tranche at a time with high-risk falsifiers and independent review;
  4. keep unsupported research claims explicitly open.

19. Agent entry point and implementation handoff

This section is the routing contract for future implementation agents. The roadmap owns current authorization and status; this design owns the target contracts and DSE tranche exits.

  1. Start with AGENTS.md, the Definitive Vision, the live roadmap, and this design. Read the Full Moneta target, Completeness Audit, and linked FM/QCA records only for the tranche being prepared.
  2. Before any source change, verify current remote main, inspect active workstream claims/PRs, acquire the required exclusive lease, and use an isolated branch/worktree. Declare touched files, risk class, authority path, invariants and falsifiers in a tranche-specific implementation plan.
  3. The first eligible work item is DSE0, after owner acceptance or revision of RFC 0012. DSE0 resolves the removal/caller inventory and necessary subject schema/RFC decisions; pins the supported semantic fixtures, row-access and cardinality falsifiers; and freezes benchmark/human-evidence thresholds. It must not silently implement the public schema while that decision is open.
  4. DSE1 is the first runtime tranche only after DSE0 closes. It must implement and test the bounded evidence-bound deterministic composition, Dataset subject, two-structure overview, full reverse explanation and minimum attributable critique capture in the real production path. Extend existing FM3/4, admission, replay and QCA suites; do not treat design assertions or isolated compiler tests as shipped capability evidence.
  5. Later DSE tranches are separate plans and PRs in roadmap order. A tranche may not claim or repair another FM lane by implication. Stop and revise if its falsifiers show hidden O(N) presentation work, unsupported meaning, disclosure/obligation loss, identity collision, unbounded resource cost, or no useful task-equivalent improvement.

The design is therefore discoverable from PROJECT_DOCS_INDEX.md, the canonical architecture reference and Full Moneta target, and the live roadmap row. The per-tranche plans created after owner acceptance become the agent-addressable implementation specifications; they are deliberately not pre-authorized by this architecture proposal.