Dominion

Dominion proposes frame-based skill cards, a skill supply, a semantic-network world state and a five-card working hand. Action/buy/cleanup cycles control selection and acquisition. It separately proposes proof-based macro learning, case adaptation, pretrained perception and a trained symbol translator. The game supplies design provenance; integrated performance remains unvalidated.

Read the material

Material	Role
Complete original	Exact 66,607 bytes, 351 lines, one table, 119 bibliography entries.
Full assessment	All 24 claims, 28 literal main/supplementary passages and historical scopes.
Current catalog	Accepted historical assessment/review/lead identities.
Historical review guide	Joint-review attribution, exact versions and bounded PDF context.
Selected evidence	Exact source-specific and explicitly shared values/ranges.
Machine record	Separate statuses, four comparisons and one existing question.
Provenance	Copy origins, original bases, historical versions and scan limits.
Rights	Unknown rights and separate material roles.
Manifest	Ten member hashes; external handoff binds manifest.
Placement	Eleven unique prospective target/role pairs.

Read each claim with its kind, rationale, status and declared evidence scope. The original joint review has 51 claims across two sources; 24 belong to Dominion.

Cards, states and information inputs

C02–C08, C17. Frames organize slots, defaults and attachments; a graph stores world facts. Action cards specify logical preconditions and add/delete effects plus card/action/buy increments. A working implementation needs typed identifiers, finite variable domains, exception/conflict rules, atomic transitions, resource units and agreement between attached procedures and declared effects.

Coins represent abstract purchasing power, not a measured compute or information unit. A goal card is a goal representation, not proof that an external goal was achieved. A harmful belief differs from a mandatory constraint. A five-skill hand and storage partitions are design choices, without neurological validation.

The optional neural score, greedy selection and complete-plan search are distinct. Define frontier, backup, failures, training targets and urgent-skill availability. A modeled precondition check establishes applicability under the encoded state; a perceptual error can leave an externally wrong action legal inside that model.

The phase-one Python/graph/PDDL proposal is concrete engineering. The inspected Pyperplan component supports STRIPS without action costs. Numeric resources need an explicit bounded compilation or another supported planner. Citation 91 points to kbvatral/pyplanning, whereas entries 92–93 are Unified-Planning documentation. No compatibility, installation or engine run is asserted here.

Acquisition, proof-derived learning and grounding

C08, C11–C13, C18, C20. Buying an authored card selects a supplied skill. The library's functions and acquisition information count toward the system's knowledge budget. Learning a purchase policy would need an explicit update.

EBL is a separate meaningful learning proposal: use a domain theory and successful trace, regress the goal through the actions, preserve required conditions and compile a reusable macro. One example still comes with prior action knowledge, proof and search cost. Macro applicability, interactions and added branching need checking; compare total discovery/compilation/matching/retrieval cost.

CBR retrieves and adapts a prior case. Similarity alone does not validate the adapted plan. The source acknowledges overly specific EBL and suggests CBR or broader proof generalization; neither name supplies a finished remedy.

Pretrained perception and a trained embedding-to-symbol translator supply object facts and temporary cards. Count labels, ontology curation, pretrained resources, supervision and model training. Define uncertainty, object identity, false-positive handling and retraction. Symbol insertion is not proof of perceptual truth or learning an unknown concept from one example.

Historical assessor arguments and memory revision

C09, C19, C22. Under ordinary uniform shuffle, each individual card in the shuffled pool is exchangeable. More copies can increase frequency. Recency alone does not supply a weight; conditional availability before reshuffling is a different effect. This is the historical assessor's conditional analysis, not a cognitive experiment or rejection of every possible retrieval policy.

A +Actions effect may replenish the counter. The counter alone therefore is not a fixed decreasing fuel budget. Finite cards and once-per-cycle consumption can bound restricted play. Card generation/reuse and total planner/episode expansions need independent termination and resource contracts. No arbitrary infinite-loop or measured complexity result is inferred.

The source's DQS and OSGI expressions remain concatenated mathematical strings. The accepted PDF scan also found no recoverable fraction layout or orientation. If the OSGI order meant new-time/old-time, a speedup would lower it below one, in tension with the printed higher-is-better statement. Old/new would be a possible attributed repair, not recovered source. DQS needs units and task-outcome meaning. No ratio or operator is silently supplied.

C14–C16. Low utility can identify cache candidates while rare recovery skills remain indispensable. Deleting a macro does not make its previous consequences false. Justification-aware TMS revision tracks reasons; an alternate valid support can preserve a conclusion. Separate eviction, dependency revision, rollback and protected constraints. Investing in known long-term returns also differs from acquiring information through exploration.

Verification, budgets and comparison routes

C01, C18, C21–C24. Formal properties require predicates, transition semantics, environment assumptions, proof obligations and complete action mediation. Model-relative proof does not independently establish truthful perception, intended goals or external authorization. Explicit symbols and traces are useful inspection artifacts, with no general exclusivity or alignment theorem.

The 12-month phases,95%scene/object targets, five solved planning tasks, OSGI>5, rising DQS across 100 games, ten safety constraints and single-GPU footprint are prospective criteria. Reconcile scene/object denominators. Longer chains can be wasteful; known answers, difficulty, memory, uncertainty and all training/search/ macro overhead must be controlled. The source itself recognizes search explosion, inconsistent rules and overspecific generalization; preserve these qualifications.

Four actual competent_simpler_baselines match: symbolic planning with/without validated macros and ordinary caching; indexed skill retrieval at equal capacity; the same perceptual frontend and translator feeding conventional action checks; and a justification store with separate cache eviction.

One existing optional question, unexecuted: “Does random deck cycling improve any independently defined task outcome over indexed retrieval at equal capacity?” The other accepted questions remain in the unchanged assessment.

Historical evidence and local delivery

Only review_status and disposition's closure sentence differ between initial and final assessments. Exact reversal recovers all 35,044 initial bytes. Joint cards review has no findings; the lane's Soar provenance repair belongs to other sources. Historical C09/C19/C22 remain assessor derivations.

Five source-ID/actualbp0045-c/title registry queries identify only the known light-triage record. Candidate sibling identity and later accepted scan retain their distinct stages. Author overview of 23 pages plus zooms 14/15 differs from the independent review sample 1–2 plus zooms 14/15. The bounded no_indication_within_scan outcome is not full correspondence, absence or scientific validation. PDF and derivatives stay reference-only; this task performs no new scan.

Rights unknown, known answers exposed, actual repository/commit/platform IDs null. Flynn/bp-0045 is WR-owned and planned. Current dossier review and root adoption remain separate. Owner delivery is held locally under the WR hold. No source/game/solver execution, training, new primary/PDF study or publication.