Cluedo / CDE

CDE proposes a SAT/CSP knowledge base, rule-based query/response agents, a learned query-ranking bridge, an event log and a governor. It uses Cluedo ownership constraints and public/private observations, then proposes a compiler for other formally described puzzles. Iterative elimination as controlled deductive emergence, 100% explainability, consumer-GPU efficiency and alignment remain unqualified architecture claims.

This is the first standalone documentary dossier for BP-0032. The source and accepted historical assessment are preserved study material. No CDE implementation, new solver experiment, PDF investigation or new model score is supplied. Current independent review and root adoption follow the frozen submission in separate version-bound records.

Reading route

Material	Role
Complete original	All 57,095 bytes, 320 lines, tables, walkthrough and 43 bibliography entries.
Full accepted assessment	All 14 claims, exact quotations, rationales, scopes and bounded primary reading.
Catalog	Exact accepted historical version and lead decision.
Historical review guide	Cluedo corrections, shared Butterfly context and PDF closure.
Selected evidence	Exact JSON values, direct initial/final assessment delta and PDF chain.
Machine record	Separate statuses, claim locators and matched baselines.
Provenance	Version identities, original path bases and reference-only assets.
Rights	Unknown rights; historical copies and derivative guide identified separately.
Manifest	Ten member identities; manifest bound externally.
Placement	Eleven unique prospective repository targets.

Read each claim with its kind, status, scope, rationale and actual evidence-read boundary. Historical authors read the selected external scopes; this dossier does not reopen them. All bibliography entries are retained, with no full bibliography audit or replication implied.

Information and deduction contract

Player knowledge versus referee truth — C01, C03–C06. A referee can hold the complete deal. A player reasons from its own hand, exact private card reveals, public suggestions/refutations and declared turn order. The recipient of a shown card can know its identity while other observers learn a disjunction. The source already gives that correct example at lines 74–76, while its table, global update at line 65 and all-agent kernel at lines 115–117 leave routing inconsistent. The diagnosis is assessor reasoning, not an original critique supplied by the blueprint. Giving the reasoner all 18 non-envelope cards changes the information-gathering task.

The walkthrough at lines 235–241 logs Agent 2 refuting and Agent 3 passing. Under a response order with Agent 3 later, the first refutation stops questioning and no pass is observed from Agent 3. An earlier Agent 3 pass could be valid. The source does not state seating order; neither a universal invalidity claim nor a negative fact from an unqueried player follows. Choose a legal transcript and declare the variant, board/movement rules, hand sizes, response order and governor/bridge access. A unique envelope needs less information than a unique complete deal. Nested knowledge also requires an explicit observation/event model; mutual and common knowledge depend on observability assumptions.

Literal Markdown and PDF evidence — C02. The preserved Markdown line 62 says i=j in the exclusion. The accepted assessor derivation is conditional: an allowed self-pair reduces the negative disjunction to one negative literal; doing so for every location contradicts the at-least-one clause. If self-pairs were implicitly excluded, the stated condition instead gives no at-most-one restrictions. Using distinct location pairs is the assessor's explicit reconstruction of the intended constraint, not a repair made to the source or recovered PDF text.

The earlier sibling inventory is only a candidate relation. Accepted light scan PDFT-002-06 compared 22 physical overview pages and selected enlargements, yielding possible_relevant_addition. The subsequently adopted PDFT-002-06-F1 closes the bounded practical investigation: physical PDF page 5 has an additional reproducible glyph before a separate equals sign, absent from Markdown line 62. The existing source-specific mapping establishes a private-use U+E020 representation. Its intended mathematical operator remains unresolved. No equality/inequality is recovered, no technical-content supplement is justified and no accepted scientific status changed. Earlier open/queued wording remains historical. Completed inquiry and unresolved intention are both represented; no universal unrecoverability proof is implied. The PDF, sibling Markdown, render assets and complete shared records stay reference-only. This unit compares their accepted metadata and exact source anchor without fresh extraction, rendering or pixel inspection.

Query semantics — C07, C08. For consistent knowledge K, possibility of H concerns satisfiability of K with H. Entailment requires a distinct test against its negation. A returned model is not an enumeration of all deals. Multiple compatible envelopes do not make a satisfiable hypothesis impossible. Solver interruption or timeout is a separate execution state, whose handling the blueprint leaves unspecified.

Proof, explanation and authority — C09, C12. A proof concerns encoded premises; an internally consistent wrong observation can still yield an externally wrong answer. Proof support and independent certificate checking need explicit handling. A complete event history, a formal certificate, a useful human explanation and diagnosis of a faulty premise convey different information. Explicit constraints and goals do not by themselves validate premises, objectives or authority for an external effect. The bridge explicitly optimizes query choice despite the broad source claim that behaviour is not optimized.

Learning and supplied rules — C10, C11. Query ranking needs a belief/response model, training labels and held-out evaluation. Logical admissibility does not establish information value. Optional perception or induced constraints need separate truth/admission policies. The source's compiler is given formal variables, domains and rules plus one instance. Solving such an instance tests compilation and deductive search, with compiler semantics/coverage still untested; it does not show rule induction from one example.

Comparisons and optional contribution

The accepted competent_simpler_baselines retain four tasks: a single CPU incremental SAT/CSP reasoner on the same legal private/public transcript; the same reasoner with targeted-card and budgeted exact/sampled information-gain policies; a direct ordinary SAT/CSP translation for a new formally specified puzzle; and an event/dependency or proof graph for explanation. Match observations, legal query and response models, solver calls, wall time, training cost and success. Random querying is an additional lower baseline rather than the sole competent comparison.

Optional existing question, unexecuted: “Can learned query ranking beat strong symbolic query policies under the same observation and compute budgets?” No policy was trained or evaluated in this unit.

The 1,000-game accuracy target, 30% query reduction versus random, four approximately three-month roadmap phases and DES/TAS/LSS/OGS are plans. Consistency is not soundness of every derived statement; correctly rejecting a corrupt transcript need not be a solver error. Hardware feasibility has no integrated workload/resource profile. The consumer GPU is the historical source's choice; no current hardware recommendation or independent specification check is made.

Versions and delivery

The shared historical review covers Butterfly and Cluedo. The available initial and final full Cluedo assessments permit a direct eight-leaf comparison: BC-R02 changes C03 kind/scope attribution, BC-R03 corrects timeout wording, BC-R04 corrects CLUE2006 pages to 214–219, and Cluedo's part of BC-R05 replaces the operative source quotations at lines 37 and 218. Two additional leaves close review metadata. BC-R01 and Butterfly's part of BC-R05 are other-source context. All other assessment fields compare equal; all accepted final claims and original source bytes are intact.

The proposed private destination is Flynn/bp-0032, owned by WR. Actual repository, commit and platform IDs remain null. Rights are unknown and public release is not authorized. Included answers are exposed study material. The lead decides adoption and sealing, then holds the exact original owner handoff locally under the active WR communication stop.