Duckstad

Duckstad proposes hierarchical Soar-inspired agents sharing a world-state store, with impasse specialists, dependency-based chunking, temporal memory, probabilistic constraints, VSA encodings and attribute-described unseen classes. Character roles motivate the design. The composition remains an unvalidated research proposal.

Read the material

Material	Role
Complete original	Exact 93,503 bytes, 436 lines, two damaged tables and 138 bibliography entries.
Full assessment	All 26 accepted claims and 27 literal passages.
Current catalog	Historical accepted assessment/review/lead identities.
Historical review guide	Source-specific CITIES-R1 and exact version/path semantics.
Selected evidence	Exact historical objects, original rationale and source ranges.
Machine record	Separate statuses, five comparisons and one existing question.
Provenance	Three full-copy origins, old/current tool roles and bases.
Rights	Unknown rights and separate material roles.
Manifest	Ten member hashes, manifest externally bound by handoff.
Placement	Eleven unique prospective repository targets and roles.

Read each claim with its kind, status, rationale, scope and primary limits.

Agent control, chunking and retention

C02–C08, C18, C23. The source supplies goal/resource orchestration, specialist search, an executor, indexed handbook facts and simulated testing. Guus represents exogenous stochastic world effects rather than an active agent. Distinguishing action failure from external randomness needs an actual transition/noise model and observations; frustration alone does not identify the cause.

A shared store needs transaction order, writer ownership, versions, complete logs and local/global state semantics. A hierarchy does not automatically guarantee global predictability or eliminate coupled search. An impasse creates a subproblem; operator knowledge, preferences and termination remain necessary.

Soar chunking is a real conditional compilation mechanism. It generalizes dependencies of a resolved subgoal and produces a result that may be a working-memory element or preference, rather than copying an external action sequence. Preserve variables, negative conditions, dependency support and future applicability. Compare total solve, compile, match and interference cost against ordinary planning/caching and validated macros.

Historical C06 observes that an unchanged old rule can coexist with a new matching rule that changes preferences and behavior. Additive storage retains records, without guaranteeing competence. Retrieval growth, memory limits, contradictory specific rules and changing perception matter. This is a non-entailment argument, not an executed forgetting experiment.

The five source stages progress from executing a supplied PDDL plan to learning, VSA graphs, specialist coordination and defensive adaptation. Executing a prewritten plan is distinct from planning and learning. The later source explicitly admits search explosion, drift and fallible abduction; keep those qualifications alongside the opening guarantees.

Temporal facts, probability and meaning

C09–C16. A timestamped fact can support indexed history when entity/event identity, intervals, updates and retractions are defined. It does not itself predict a future fact. TLogic's checked abstract explicitly targets future links; ChronoR's abstract does not establish unseen-time extrapolation. RE-GCN and the full algorithms remain unverified in the historical scope.

An MLN uses formulas with real log-linear weights and a normalized distribution over finite grounded interpretations. A weight is not a standalone probability that a rule holds; other formulas and domain grounding affect marginals. Choose fitting/inference policy and semantics before making probability claims.

VSA provides encoding operations. Select algebra, dimension, dtype, codebook, roles, normalization, similarity thresholds, decoding and capacity. Bundling does not normalize automatically; inversion and crosstalk depend on the model.

Historical C12 restricts its commutativity observation to the named componentwise multiplication or ordinary circular convolution: binding ab equals ba. The printed Dagobert/first-dime pair therefore lacks ownership relation and ordered argument roles. Role tags/permutations would be a new explicit encoding, not a recovered missing expression or a criticism of every VSA model.

Independent random near-orthogonal item codes do not supply thematic similarity. Designed shared attributes or learned alignment can make such retrieval useful. A greed/failure query and an event-time query need different constructions.

A neural scene-to-vector encoder needs training, labels or other supervision, identity tracking, loss and calibrated decoding. Approximate matching needs thresholds, rejection and an applicability test; it is not exact Soar matching. Attribute transfer uses prior trained detectors and supplied target-class descriptions. Adding a description vector is representation, not acquisition of new perceptual meaning from no information.

Verification, explanation and evaluation

C17, C21–C25. The source proposes logs of working memory, rules, facts and preferences. Reproduction also needs model/stochastic state, thresholds, concurrency and external inputs. An accurate execution trace differs from truthful perception, formal validity and authorized effects.

The accepted CITIES-R1 correction permits general formal verification or individual proof checking for infinite-state systems and undecidable theories. Properties, checking methods and environment assumptions must be explicit. Guaranteed terminating exhaustive checking needs a suitable decidable fragment, finite abstraction or other justified termination bound. No universal complete checker or blanket impossibility of neural verification is established.

Narrative coherence measures supplied stylistic criteria and can miss errors in its summary. GEI has no formula, estimator or normalization; entropy alone can reward ineffective randomness. Historical C22 distinguishes coordinates from meaning: a joint orthogonal rotation preserves cosine relations while moving coordinates; unchanged item codes can be used differently. A semantic stability test therefore needs its own invariant/readout.

C24 remains a defensive limitation: false but ontology-consistent observations need not trigger a contradiction. Consistency alone cannot guarantee detection or distinguish an attack from ordinary error. Detection and false-positive rates require a declared distribution and measurement; no attack was run here.

The length-10 execution, 95% tie resolution, depth-3 query 99% accuracy with latency<50ms, 90% task completion, communication overhead<15%, 98% detection and 80% mitigation figures are prospective targets. Fix distributions, denominators, independent oracles, task novelty and total costs. The three-agent requirement must come from the task rather than favoring an architecture.

Incomplete engineering and comparison routes

C15, C19–C20, C26. Table5.1 retains criterion rows without representation columns or comparison values. Table7.1 lists components and Total without allocations. Keep those gaps; do not invent vector dimensions or GPU budgets. Kernel mapping, memory traffic, synchronization, graph access and precision sensitivity determine end-to-end performance, beyond elementwise arithmetic.

ACT-R/pyactr references do not supply Soar chunking semantics. Torchhd operations, graph stores and orchestration are building blocks rather than the proposed composition. The checked Pyperplan component supports STRIPS without action costs; luck, temporal facts and utilities need separate compilation/modeling. The source's anticipated hardware values remain historical design premises.

All five actual competent_simpler_baselines remain addressable: ordinary planning and recurring-impasse caching; indexed temporal facts and matched forecasting; the same attribute detector with direct prototypes; a single controller using identical specialist routines; and matched defensive validation/provenance.

One existing optional question, unexecuted: “When does approximate VSA rule matching improve noisy-input task performance without invalidating chunk applicability?” The other four questions stay in the unchanged accepted assessment.

Historical preservation and local delivery

Only C17.rationale changed inside historical claims; review_status and the disposition closure sentence also changed. Exact reversal reproduces all 41,254 initial bytes. The cities review covers 48 claims across two sources; the wider lane has six. Duckstad's CITIES-R1 and other-source corrections remain separate. The original CTT passage explicitly denies uncomputable functions; it supplies no proof of intelligence from interactive modeling convenience.

Five literal source/prefix/title registry queries return no matching fields within their pinned scope. No Discworld scan or other PDF evidence transfers to Duckstad. Historical and current validator identities are separately bound; neither is run. Rights unknown, known answers exposed, actual repository, commit and platform IDs null. Flynn/bp-0047 is WR-owned and planned as a private repository. Current review and root adoption are external stages; owner delivery is held locally. No new research, witness/PDF campaign, solver, model, source execution or repo work.