Temporal Tapestry Architecture

DeLorean's Temporal Tapestry Architecture proposes many small neural-to-symbolic agents exploring hypothetical branches in a layered knowledge graph. The McFly core stores facts, rules, branch IDs and derivations. A Doc Brown supervisor allocates resources, prunes branches and would learn search control through RL. The source seeks weak emergence and explicit consistency within each timeline. Its integration, learning transfer, global consistency and hardware claims remain unvalidated.

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 60,066 bytes, 399 lines, three tables, line-broken ASCII diagram and 33 references.
Full assessment	All 19 claims, main and supplementary anchors, evidence and limits.
Current catalog	Accepted scientific identity and original review/lead links.
Historical review guide	Shared source scope, two assessment versions, argument IDs and bounded PDF scan.
Selected evidence	Exact historical values/ranges, six claim–argument links and PDF scan stages.
Machine record	Separate statuses, seven comparison routes and one existing question.
Provenance	Copy origins, original reference bases and external-only assets.
Rights	Unknown rights and distinct historical/derivative roles.
Manifest	Ten member hashes; handoff binds the manifest itself.
Placement	Eleven unique prospective repository targets and roles.

Read every claim with its kind, status, rationale and actual evidence scope. The 37-claim historical review concerns two distinct sources; DeLorean has 19. The current dossier review and lead adoption are later, separate records.

Branches, states and maintenance

C02–C06, C18. A timeline ID helps isolate hypothetical alternatives. It does not distinguish a branch's earlier and later states, observation from hypothesis, fact-validity intervals, rule versions or inherited versus locally edited data. The source requires branch-scoped queries and assigns merging to custom code. Those constructive proposals remain visible.

Delayed propagation requires a visibility contract. In attributed A6, q has the sole support p→q; retracting p before invalidating q can expose stale support. Pending/unknown labels, snapshots or a commit/read barrier can address that specific interval. This is a conditional deletion case, not an impossibility claim about asynchronous processing or monotonic additions.

A5's two consistent branches add p and not-p separately. Their union conflicts. The source does not prescribe union merging; the example identifies a missing merge policy. Per-timeline checks do not automatically validate concurrent merges. Dependencies, retractions, tombstones, snapshot identity, conflicts, scheduling and quiescence need explicit semantics. Doyle's primary abstract supports reason maintenance and assumption revision, not a full TTA algorithm or global consistency theorem.

Provenance, components and supplied knowledge

C07–C11, C14. A derivation graph records premises and rules, answering why the system believes a statement. A4 separates that from identifying a physical cause or predicting an intervention. Supplied action/mechanism assumptions can support model-relative simulation; no particular causal formalism is required. A trace also needs complete executed premises, versions and retention after pruning and does not explain neural or supervisor internals by itself.

Neo4j GraphRAG supplies graph construction/retrieval components, with LLMs and embeddings in the inspected tutorial. The source assigns temporal and causal layers to custom code. Experta documents patterns, rule actions and fact lifecycle; its pure Python description does not demonstrate GPU matching or parallel safety. PostgreSQL's recursive graph examples refute categorical inability to express traversals, without proving the fastest store for this workload. SWI snapshots and transactions illustrate bounded state/visibility contracts, not universal rollback.

The genesis knowledge base contains supplied facts, logical/ontological axioms and common-sense rules. A neural frontend and relevance/query selector require training, labels, calibration and prior resources. Retrieved task context can omit a needed premise. Solving new combinations with provided rules may legitimately be zero-shot relative to those tasks, while knowledge acquisition and selector training remain costs. No blank-slate or training-free system is demonstrated.

Supervision, resources and recovery

C04, C12–C13, C17, C19. Weak emergence is the explicit aim, with local interactions in principle simulable. It is not a consciousness claim or measured macro-level benefit. Resource allocation, learned pruning and adaptive control still require supervisor state/actions/reward, credit assignment, data splits, update rules, independent evaluation and full training/orchestration cost.

A9 considers a cheap dead end and an initially costly branch containing the only solution. A mistaken relevance or cost score can remove that solution. Bounded search need not promise completeness, but solution loss and success at fixed budgets must be reported separately. More agents or GPU kernels do not validate the scheduler, query workload, graph/model residency, transfer cost or latency.

The source distinguishes immutable guardrails from editable ordinary rules and proposes a circuit breaker for oscillation/deadlock. It would freeze the swarm, restore a known state and possibly change strategy. Enforcement precedence, atomic checks, checkpoint identity, permitted effects and reset scope remain open. Resetting memory does not undo an irreversible external effect without a compensating operation; proof of derivation is not action authority.

Metrics and comparison routes

C15–C16. DER divides correctly solved unseen problems by counted initial facts/rules. Equivalent rule packaging can change that denominator without improving answers. Under A7's explicit same-100-puzzle assumption, N≤100 and strict DER>10 imply K<10. The roadmap may use another DER population, but does not state one unambiguously. This is conditional arithmetic, not an unconditional contradiction or observed benchmark. It does not reconstruct the other source's missing DEQ.

The million-node query below 50 ms, >95% puzzle success, DER>10, 10× Hanoi speedup, ECI>1.5, TS>0.8 and 30% RL improvement after 1000 iterations are proposed example targets. Query shape, competent algorithms, total compute, independent correctness, prior exposure and held-out tasks matter. TS and ECI lack complete numeric scoring definitions. Internal process measures supplement external task validity.

Seven actual simpler_comparisons retain: one engine with explicit branch state; relational/indexed graphs with snapshots; full retrieval or dependency slicing; fixed-budget breadth/best/beam scheduling; ordinary Hanoi/planning controls; dependency ledgers with transaction boundaries; and proof traces versus external causal evidence. Compare matched knowledge, training, expansions, complete costs and lost solutions.

One existing optional question, unexecuted: “Can delayed maintenance preserve declared query semantics under deletions and concurrent branch merges at lower cost than eager closure?” This remains a reading route, with no new simulation or implementation commissioned here.

Historical versions and PDF scope

The initial and final DeLorean assessments differ only in review_status and review_records. Exact reversal restores the initial bytes; all scientific content stays unchanged. Original assessment-directory and lane-directory references resolve separately. A4/A5/A6/A7/A9 are actual JSON argument IDs. The six accepted claim links are preserved without borrowing BP-0020's other-source examples.

PDFT-003-01 selects BP-0040 despite its Back-to-the-Future filename. The earlier candidate inventory and later accepted scan remain separate. The sibling Markdown is byte-identical to this original; title overlap alone does not identify the distinct BP-0020 source. The historical author overviewed 25 pages and enlarged pages 9/10/19. The reviewer sampled overview 1/2/17/18 and the same zooms. Page/line anchors and the diagram/roadmap observations retain those exact scopes.

The accepted outcome is no_indication_within_scan, with no follow-up IDs. The ASCII diagram remains poorly aligned and split. The scan establishes no full character/table equality, source precedence, recovered aligned diagram or science validation. PDF, sibling, page images, zooms and extracts stay external. No new pixels, rendering, extraction, reconciliation, retrieval integration or experiment occurs in this dossier.

Rights are unknown and all answers exposed. Planned private destination Flynn/bp-0040 is WR-owned; actual repository, commit and platform IDs are null. The lead decides adoption/sealing and retains the original owner handoff locally under the separate outgoing WR communication hold.