HelixAI

HelixAI proposes a typed knowledge graph with learned concept embeddings and inverse-relation checks, a partitioned GPU store, neural graph queries, theorem-prover/TMS admission, active-inference policy evaluation and a sampler/ planner interface. DNA supplies design provenance. The source also specifies training, curriculum, uncertainty, partition loading and planner feedback. The integrated operational guarantees remain unestablished.

This first standalone documentary dossier preserves the complete original, accepted full assessment and current catalog with scoped historical guidance.

Read the material

Material	Role
Complete original	Exact 79,506 bytes, 433 lines, four tables, 88 references and EOF without newline.
Full assessment	All 25 accepted claim objects, 64 literal main/supplementary passages and evidence limits.
Current catalog	Accepted assessment, review and lead identities.
Historical review guide	Four-field version transition and HelixAI-specific analytical readback.
Selected evidence	Exact historical objects/ranges and unchanged source formula gaps.
Machine record	Separate statuses, four matched baselines and one existing question.
Provenance	Copy origins, original path bases, versions and reading limits.
Rights	Unknown rights and distinct copied/derivative material roles.
Manifest	Ten member hashes; handoff binds the manifest externally.
Placement	Eleven unique prospective repository targets and roles.

Read every claim with its kind, status, rationale and actual primary scope. The historical joint review covers 50 claims; 25 belong to HelixAI. Current dossier review and lead adoption remain separate records.

Encoding and graph integrity

C01–C05, C14. The historical assessor's information bound concerns a declared discrete alphabet: four distinguishable values carry at most two bits per symbol. Three such symbols give at most 64 words, or six bits, before representation overheads. Naming software values after DNA bases does not create extra memory capacity. Physical DNA storage and repair involve different machinery and coding conditions; their reliability or density does not transfer to software.

An enforced inverse pair can certify a representation invariant. Both edges can still encode the same false assertion, and the invariant does not check every ontology axiom or identify causal effects. Subclass-of and instance membership need different types and inverse contracts. The biological analogy does not select which side of a disagreement is correct.

The source's metadata, typed concepts, learned active/inactive embeddings and proofreading stages are meaningful design objects. Entity/category/event identity, versions, confidence and training targets remain unspecified. A provenance chain needs actual inputs, learned scores, selected policies, random choices and observed effects to explain execution.

Probabilistic and causal models

C07–C12, C17. Active inference evaluates beliefs and policies under an explicit generative model. Hidden state, observation likelihoods, action-conditioned transitions, preferences, priors, horizon, costs and approximate inference must be defined. A typed graph is not already a normalized probabilistic model. Changing a descriptive belief and changing a preference are different updates.

The source's expected-free-energy slots at lines 107–112 remain blank. Variational free energy for inference and expected free energy for policy evaluation retain their formulation-specific distinctions. No external expression is inserted into the original. The cited language-model integration is itself forward-looking, with reliability and energy caveats.

CCGM's paired concept embeddings and interventions concern a modeled decision process; its learned structure need not identify the data-generating mechanism. Causal attention supervision likewise intervenes on model attention/predictions. World interventions need additional assumptions and evidence. Two embeddings or a directed edge do not establish environmental causality.

Soft attention changes weights; it need not save dense GNN operations. Hard pruning can remove required premises. If exact logical answers are claimed, preserve a separate exact route and measure both saved computation and missed dependencies. Learned query predictions are not proof-producing deduction.

Admission, proof status and planning

C06, C15–C16, C18–C20. The source combines graph updates, TMS and a theorem prover. Its consistency-expression gap at lines 189–190 remains empty. Refutation, completed satisfiable saturation, resource exhaustion, incomplete search and unknown are distinct outcomes in the inspected older E manual. No contradiction found before timeout is not a consistency certificate; completed supported satisfiable outcomes remain possible. Specify the fragment, proof/ certificate/status handling and atomic admission policy.

TMS records alternative justifications and can revise dependent beliefs after a retraction. This is model-relative justified belief, not factual truth. Changing a premise needs revision priorities, dependency tracking, concurrent visibility and commit/rollback semantics.

PDDLStream conditionally integrates planners with external generators. Output tuples must satisfy their certified predicates. A neural feasibility score or most-likely grasp is not automatically a sound certified fact. Unknown, invalid samples, retries and execution failure require explicit conversion/handling.

The source's feedback categories are useful proposals, but a failed precondition, exhausted search, timeout and proof of unreachable goals are different statuses. Generic tools need not emit the illustrated explanations. Planned or predicted outcomes differ from observations. Model validation, sensing/replanning and external effect authority remain separate obligations; no physical actuation occurred here.

Partitioning, hardware and information budgets

C01–C02, C13, C21–C23. An explicitly enumerated b-ary depth-d frontier has b^d states under that construction. Fixed hardware does not remove that growth; this is not a universal lower bound on every possible algorithm. RT cores accelerate geometric BVH/ray work. A correct reduction from the actual logical query needs graph encoding, filtering, updates, build cost and answer completeness before an acceleration claim follows.

The later on-demand partition loading qualifies the early whole-graph-in-VRAM claim. Include boundary edges, multi-hop closure, stable IDs, stale caches, joins, proof/TMS residents, neural weights, activations/workspaces, staging buffers and host/device transfers. Minimizing cut edges does not remove required dependencies.

PagedAttention manages token KV-cache blocks in its inspected scope. Mutable graphs and repeated message passing need their own tensor inventory and invalidation schedule. The proposed adaptation is not an executed GNN result.

Two budgets matter. Upstream embeddings, perception, GNNs, training tasks, curriculum, supplied rules, human graph knowledge and model priors cost information and computation. Later low-cost fact updates or one-example adoption are separate targets, not evidence of training-free competence. Keep labels, observation signals and independent task oracles visible at each interface.

Criteria and comparison routes

C22–C25. The phased million-node/ten-million-edge storage, five-hop accuracy and latency, fact-admission rates, T-maze, resource and curriculum milestones are prospective. Maze information-seeking depends on preferences, observations, horizon and cost. Acceptance rate alone can be high while wrong facts are admitted.

CTS requires fidelity to executed computation, including neural/policy state. SRT requires an independently observable change and a specified correct behavior. KIL measures trials until behavior adopts an injected fact under existing concepts/rules. KIL=1 does not by itself isolate learning a new category, skill or causal law. Retention, conflicts, held-out tasks, labels and total priors matter.

The four actual competent_simpler_baselines use typed graph/relational storage with inverse constraints and TMS; exact indexed CPU/Datalog queries; a specified POMDP/Bayesian information-gathering controller under the same model/preferences; and standard PDDLStream with sound samplers and explicit sensing/replanning. Match facts, actions, observations, answer semantics and all training/inference resources.

One existing optional question, unexecuted: “Can the graph compile into calibrated predictive distributions and explicit preferences without conflating them?” This remains a reading route, not a compiler or model implementation.

Historical review and local delivery

The available pre-review and final assessments differ in review_status, disposition's pending suffix, review_refs and review_acceptance. The last two fields were newly added; inverse operations remove them rather than assigning null. Exact reconstruction equals the full initial file. Every claim and primary scope remains unchanged; no substantive correction was requested.

Only the historical C01/C03/C14 analytical rows are selected for HelixAI. MandelBrot's complex-number, fixed-point, Jacobian and shaping arguments retain BP-0111 attribution. Shared 50-claim/79-supplement/213-reference counts span both sources. Old batch usage-stop wording is chronology and grants no new research; this bounded task follows the later verified production resumption.

Five actual source-ID/prefix/title register queries return zero within their pinned scope. No PDF or mechanism is attached through the DNA analogy. No new biological, physical encoding, model, solver, GPU, paper/PDF or scientific study is performed. Rights unknown, answers exposed, planned WR-owned Flynn/bp-0043 and null actual repository/commit/platform IDs remain explicit. The lead decides adoption/sealing and holds the owner handoff locally under the separate WR pause.