TARDIS

TARDIS proposes a shared knowledge graph with explicit triples and embeddings, an episode log, learned heuristic search or theorem proving, DyGRAIN-inspired continual graph learning and an outer adaptation loop. Protected axioms and a consistency verifier are intended to regulate changes. Doctor Who supplies design provenance. The integrated guarantees remain research proposals.

Read the material

Material	Role
Complete original	Exact 75,780 bytes, 438 lines, three tables and 79 bibliography entries.
Full assessment	All 17 unchanged claims and literal quote contexts.
Current catalog	Accepted historical assessment/review/lead identities.
Historical review guide	Source-specific version, A1/A2 and original path semantics.
Selected evidence	Exact attributed objects and Markdown ranges.
Machine record	Separate statuses, five comparisons and one existing question.
Provenance	Original bases, three whole-copy origins and scoped historical evidence.
Rights	Unknown rights and material roles.
Manifest	Ten member hashes; manifest bound externally by handoff.
Placement	Eleven unique prospective repository targets.

Read each claim with its kind, status, rationale and evidence scope. Seventeen TARDIS claims are distinct from the joint review's 38-claim scope.

Search contract and the two analytical arguments

C09–C10, C15. The source actually prints f(n)=g(n)+h(n) on line 215. A learned estimate with small average error need not be admissible. Consistency and reopening policy matter; ordering states differs from irreversibly pruning them. Valid paths and optimal paths are different claims. A search contract must state termination, bounds and permitted approximation.

Historical A1 uses positive edges s→a=1, a→g=1 and s→g=3, with h(a)=10 and h(g)=0. After s, f(g)=3 precedes f(a)=11. Standard goal-pop termination returns cost 3 although s→a→g costs 2. This is an attributed finite-graph argument with explicit conditions, not a run or a universal impossibility result for neural heuristics. The source's line-305 optimality target needs an applicable guarantee.

Historical A2 concerns line 314: performance divided by task-specific example count has no ordinary real value when that count is exactly zero, including 0/0. Keep the printed metric and its limitation. New fixed-budget reporting would be a separate metric choice. No replacement convention is silently added.

Match A*/IDA*, Manhattan and competent pattern-database heuristics. Count training, inference, preprocessing, memory and total wall time alongside state expansions; fewer expansions alone can conceal added neural cost.

Graph state, supervision and revision

C03–C05, C11. Explicit triples and embeddings need entity identity, relation semantics, confidence, negation, versions and update rules. Vector proximity does not itself establish predicates. A class-property example needs universal versus default inheritance semantics.

DyGRAIN is a real component precedent with ground-truth supervision, a reservoir buffer, influence scoring and loss-discrepancy selection. Informal rarity or importance is a proposed adaptation, not its complete selection rule. Prediction retention differs from preserving arbitrary symbolic facts. Match labels, graph stream, buffer and memory with ordinary replay/distillation.

The planner genuinely restricts actions to information and construction and excludes deleting verified knowledge. Temporary hypotheses, admitted facts, learned embeddings and protected axioms require separate states. Incorrect facts and their consequences need explicit revision, dependency retraction, conflict priorities and transactional visibility; retaining false statements forever would not establish reliability.

Episodes record states, steps, heuristics, tools, outcomes and feedback. Reproducible replay also needs exact input/model/rule versions, randomness and tool results. Storing or retrieving an episode is not by itself learning a new skill or proving held-out generalization.

Adaptation, constraints and explanations

C01–C02, C06–C08, C12, C16. Regeneration may change neural architectures, weights, symbolic rule priorities and agent composition. Define the task distribution, support/query split, objective, discrete search and acceptance contract. Fixed hyperparameters or ordinary task-conditioned fine-tuning provide the matched comparison.

The source permits pretraining, meta-task knowledge and synthetic examples. Count supplied rules, embeddings, labels, generated data and all prior training when discussing zero/few-shot behavior. Generated labels are not independent ground truth.

Constitutional phrases express intended priorities. A checker needs executable predicates, conflict rules, stable interpretation, a faithful state/effect model and complete mediation of effects. Protecting rule text does not prove behavioral alignment or establish authority for external actions. The verifier is described as a periodic neural audit with retraining, not an always-on certified action shield. A logic-consistent heuristic can still overestimate.

Symbolic traces are useful, while the source explicitly allows post-hoc LIME/ SHAP and attention visualization for neural predictions. A local fitted explanation and a path log need faithfulness tests; they do not expose the full neural computation as a symbolic proof. The later direct-transparency claim does not remove that qualification.

Resources, targets and reading routes

C13–C14, C17. Heavy training and regeneration may happen offline. The full graph need not reside in VRAM: the source proposes VRAM/RAM/NVMe graph paging. Retain those strategies while requiring graph/model sizes, optimizer/activation peaks, cross-partition dependencies, versions, misses, transfer volume and query mix. The approximate bandwidth slot has no numerical magnitude. No current hardware specification or measured feasibility is inferred from the old table.

Task-indexed LRU caching is the matched alternative to neural prefetching. Indexed versioned storage, explicit rules and an episode log are the first comparison to the hybrid graph. Together with competent search, matched replay and ordinary fine-tuning, these are the five actual simpler_comparisons.

The 24-month roadmap, 90%-fewer-expansions example, >95% retention over ten tasks and budgeted zero-shot planning criterion are prospective. TS and ECI need operational rubrics, reference solutions and independent judges rather than familiarity with known answers.

One existing optional question, unexecuted: “Can a learned admissible heuristic improve end-to-end 15-puzzle time over a competent pattern database at a fixed training and memory budget?” The other accepted questions remain in the unchanged full assessment.

Historical review and local delivery

The initial/final assessments differ only in review_status and review_records. The recorded metadata reversal reconstructs all initial bytes. Joint review R1 concerns BP-0087-C12, while A1/A2 belong to TARDIS; A3/A4 remain other-source. The incomplete taxonomy label on line 95 remains incomplete.

Five source-ID/prefix/title registry queries return no explicit matches within their pinned scope. A generic TARDIS or time-travel title does not establish a PDF relation. No fresh primary reading, PDF campaign, solver execution, benchmark or scientific experiment occurs here. Rights unknown, known answers exposed, actual repository/platform IDs null. Flynn/bp-0044 is WR-owned and planned. Current independent review and lead adoption remain external stages; the owner handoff is held locally under the WR communication pause.