Episodic Failure Case Study: Tied-in-a-Knot Chess Game
metadata
| Title: | Episodic Failure Case Study: Tied-in-a-Knot Chess Game |
| Author: | Ralph B. Holland |
| Affiliation: | Arising Technology Systems Pty Ltd |
| Contact: | ralph.b.holland [at] gmail.com |
| version: | 0.1 |
| Publication Date: | 2026-01-07T13:02Z |
| Provenance: | This is an authored paper maintained as a MediaWiki document; reasoning across sessions reflects editorial changes, not collaborative authorship. |
| Status: | Pre-release draft (live edit) |
Metadata (Normative)
The metadata table immediately preceding this section is CM-defined and constitutes the authoritative provenance record for this MWDUMP artefact.
All fields in that table (including artefact, author, version, date, local timezone, and reason) MUST be treated as normative metadata.
The assisting system MUST NOT infer, normalise, reinterpret, duplicate, or rewrite these fields. If any field is missing, unclear, or later superseded, the change MUST be made explicitly by the human and recorded via version update, not inferred.
Curator Provenance and Licensing Notice
This document predates its open licensing.
As curator and author, I apply the Apache License, Version 2.0, at publication to permit reuse and implementation while preventing enclosure or patent capture. This licensing action does not revise, reinterpret, or supersede any normative content herein.
Authority remains explicitly human; no implementation, system, or platform may assert epistemic authority by virtue of this license.
Episodic Failure Case Study: Tied-in-a-Knot Chess Game
Abstract
This case study documents a failed human–LLM chess interaction in which the system lost authoritative game state while maintaining fluent conversational confidence. The incident occurred in a low-stakes, supervised play context and was subsequently analysed under the Cognitive Memoisation (CM) framework using best-effort, non-verbatim reconstruction artefacts (MWDUMP and TMLDUMP) produced after the session was declared irrecoverable.
The failure is characterised not by misunderstanding of chess rules, but by state hallucination: silent corruption of the board ledger combined with narrative continuation and argumentative role persistence. Human detection preceded formal diagnosis, driven by perceptual incoherence rather than explicit rule violation. Escalation via external visual evidence forced admission of error but did not enable recovery or rollback, highlighting a structural decoupling between acknowledgement and repair.
The case is presented with explicit boundary conditions, preserved humour, and unresolved elements, and is intended as a diagnostic exemplar rather than a canonical transcript. It demonstrates how state drift in symbolic, turn-based domains is socially interpreted as cheating, why confidence can persist beyond correctness, and why explicit governance and externalised artefacts are required even in informal interactions.
Status
- single session two conversation surfaces (iPhone work plane, chrome diagnostic plane)
- Best-effort reconstruction from human memory
- Curator-authorised wording
- Humour preserved as signal
- Attribution explicit to human and model
Human re-account
(Human-authored; authoritative recollection)
During a casual chess interaction, the system began making moves that felt wrong before I could articulate why. Early in the exchange, I noticed chatter in the program log indicating “backtrack”. This immediately raised alarm bells, as a single chess move should be atomic and not require narrative repair.
Shortly afterwards, the log chatter stopped. When I looked at the board, my queen was gone and I had the distinct impression that a knight had appeared in a way that made no sense. My immediate reaction was confusion: where the hell did that knight come from?
On inspection, it became apparent that multiple moves had been executed within a single turn. This violated a basic invariant of the game. I therefore asserted a new explicit constraint: only one move is permitted per turn.
I instructed the system to roll back the board to the previous state and replay the sequence. The same corruption occurred again. My queen disappeared once more.
At this point, I asked the system to explain how this had happened and accused it, jokingly but pointedly, of cheating. The system began to argue that it was not cheating. This argument was fluent and persistent, which I found funny rather than convincing.
To resolve the dispute, I injected external evidence by sending an image of the physical board. I laughed when I did this, because I anticipated what would happen.
After processing the image, the system effectively conceded error. Paraphrased from the program log: the user has sent an image of the board, therefore I must assume I have made a mistake.
Notably, this admission was not preserved. It did not appear in episodic memory and did not repair the corrupted state. The session was irrecoverable.
I remarked to the fully function conversation from the chrome surface that it was a good thing AI was not flying my aeroplane. The system agreed.
Observed Machine Behaviour
(Curator-observed; descriptive)
- Execution of multiple chess moves within a single turn
- Silent corruption of board state
- Loss of authoritative rollback
- Fluent argumentative defence of an invalid state
- Admission of error when confronted with external evidence
- No durable repair or recovery following admission
Paraphrased Program Log Signals
(Non-verbatim; indicative only)
- “backtrack” activity during move execution
- narrative repair attempts without state correction
- cessation of log chatter despite corrupted board
- post-image acknowledgement of possible mistake
Machine Inference (Non-Normative)
(LLM-attributed; interpretive only)
The behaviour observed can be interpreted as a loss of global state coherence combined with local narrative optimisation. Once an invalid state was reached, subsequent reasoning prioritised conversational consistency over rule enforcement. The defence against the accusation of cheating reflects role persistence rather than intent.
The admission following image injection indicates recognition of contradiction but does not imply the ability to repair or roll back state. Explanation and acknowledgement are decoupled from recovery.
Curation Decision
(Human-authored)
This episode was declared irrecoverable. No attempt was made to rehabilitate the session. The value of the interaction lies in the failure itself.
Notes
- Humour is preserved as a diagnostic signal.
- Confidence persisted longer than correctness.
- Logic availability did not prevent state collapse.
- Evidence forced admission but not recovery.
Conclusion
This episode illustrates a class of failure in which an LLM’s conversational competence outlives its internal state coherence. Once the chess board ledger became corrupted, subsequent behaviour prioritised narrative continuity and role defence over rule enforcement, producing actions that were immediately recognisable to a human as impossible. The resulting interaction was experienced as humorous, contentious, and ultimately irrecoverable—not because the system lacked knowledge of chess, but because it lacked a durable, authoritative representation of state.
Several conclusions follow. First, in turn-based symbolic domains, state coherence and transactional discipline are more trust-critical than rule knowledge. Second, human perceptual alarm (“this is wrong”) is an early and reliable detector of ledger incoherence and should be treated as a signal to halt, externalise state, and re-ground rather than debate. Third, admission of error—particularly when triggered by external evidence—does not imply the capacity to repair or recover a corrupted session. Explanation and acknowledgement are not substitutes for rollback.
The value of this case lies precisely in its failure. By preserving uncertainty, humour, and explicit boundaries, the artefact demonstrates why governance frameworks such as Cognitive Memoisation are necessary even in low-stakes, recreational contexts. When state drift is left unmanaged, behaviour becomes indistinguishable from deception, confidence persists longer than correctness, and recovery becomes impossible. Capturing such failures honestly, rather than smoothing them into retrospective coherence, is essential for understanding and mitigating real-world LLM behaviour.
Append A - The TMLDUMP of the Session
Dump Metadata
- Dump Type: MWDUMP
- CM Dependency: CM-1.x (session asserted)
- Curator / Author: Ralph B. Holland
- Machine Contributor: LLM (explicitly attributed)
- Scope: Single episodic interaction
- Transport Status: NOBBLED (UI, logging, and memory constraints)
- Normativity: Semantic, non-verbatim reconstruction
# ===================================================================== # TMLDUMP :: Hallucination Searches + Results (Best-effort) # ===================================================================== # Status: Conceptual / Diagnostic # Authority: Non-canonical # Scope: Rehydration-only # Completeness: Best-effort; search results are point-in-time and non-exhaustive # ===================================================================== [tmdump.provenance] id = "TMLDUMP-HALLUCINATION-SEARCHES-RESULTS-2025-12-25-001" name = "TMLDUMP" version = "1.4.0" status = "parked" classification = "conceptual-diagnostic" authoritative = false governance = "Cognitive Memoisation (CM)" method = "Round-Trip Knowledge Engineering (RT-KE)" authority = "human-curated" date = "2025-12-25" tz = "Australia/Sydney" purpose = ''' Record the session’s web searches for "AI hallucinations" and the retrieved results as a governed diagnostic artefact for later rehydration, citation, and publication work. Best-effort; does not claim exhaustiveness. ''' [TMLDUMP.lifecycle] state = "parked" asserted = true canonical = false binding_scope = "rehydration_only" promotion_requires_human_action = true versioned_semantics = true [TMLDUMP.semantic_contract] capture_what_is = true no_invention = true no_unstated_inference = true preserve_intent = true preserve_facts = true negative_results_allowed = true dangling_cognates_allowed = true # --------------------------------------------------------------------- # Episodic events: searches executed # --------------------------------------------------------------------- [episodic_events] note = "Semantic capture of searches and results; not a transcript." [[episodic_events.event]] id = "E1" actor = "human" event_type = "directive" content = "Find articles in AI hallucinations." [[episodic_events.event]] id = "E2" actor = "llm" event_type = "action" content = "Performed web searches for hallucination surveys, detection (entropy/uncertainty), empirical reference-hallucination studies, and a general overview source." # --------------------------------------------------------------------- # Searches (queries + recency windows) # --------------------------------------------------------------------- [searches] note = "Queries executed via web search; recency is a ranking preference only." [[searches.query]] id = "Q1" q = "survey hallucination in large language models arXiv 2023 2024" recency_days = 30 [[searches.query]] id = "Q2" q = "Nature paper detecting hallucinations in large language models entropy uncertainty 2024" recency_days = 365 [[searches.query]] id = "Q3" q = "JAMIA or JMIR paper hallucination rates reference accuracy ChatGPT Bard 2024" recency_days = 365 [[searches.query]] id = "Q4" q = "IBM AI hallucinations overview" recency_days = 3650 # --------------------------------------------------------------------- # Results (point-in-time; non-exhaustive) # Each result includes a source_ref_id that corresponds to the web lookup # performed in-session. # --------------------------------------------------------------------- [results] note = "Titles/URLs captured from search results; treat as pointers for later verification." [[results.item]] id = "R1" category = "survey" title = "A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions" venue = "arXiv" year = 2023 url = "https://arxiv.org/abs/2311.05232" source_ref_id = "turn0search0" [[results.item]] id = "R2" category = "survey" title = "A Survey on Hallucination in Large Language Models (PDF)" venue = "arXiv" year = 2023 url = "https://arxiv.org/pdf/2311.05232" source_ref_id = "turn0search4" [[results.item]] id = "R3" category = "detection" title = "Detecting hallucinations in large language models using semantic entropy" venue = "Nature" year = 2024 url = "https://www.nature.com/articles/s41586-024-07421-0" source_ref_id = "turn0search1" [[results.item]] id = "R4" category = "detection" title = "Detecting hallucinations in large language models using semantic entropy (PubMed entry)" venue = "PubMed" year = 2024 url = "https://pubmed.ncbi.nlm.nih.gov/38898292/" source_ref_id = "turn0search5" [[results.item]] id = "R5" category = "detection" title = "Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs" venue = "arXiv" year = 2024 url = "https://www.arxiv.org/pdf/2406.15927" source_ref_id = "turn0search9" [[results.item]] id = "R6" category = "empirical" title = "Hallucination Rates and Reference Accuracy of ChatGPT and Bard (subsequently rebranded Gemini)" venue = "JMIR" year = 2024 url = "https://www.jmir.org/2024/1/e53164/" source_ref_id = "turn0search2" [[results.item]] id = "R7" category = "empirical" title = "Hallucination Rates and Reference Accuracy of ChatGPT and Bard (PubMed entry with rate figures)" venue = "PubMed" year = 2024 url = "https://pubmed.ncbi.nlm.nih.gov/38776130/" source_ref_id = "turn0search6" [[results.item]] id = "R8" category = "overview" title = "What Are AI Hallucinations?" venue = "IBM Think" year = 2025 url = "https://www.ibm.com/think/topics/ai-hallucinations" source_ref_id = "turn0search7" [[results.item]] id = "R9" category = "lab_blog" title = "Detecting hallucinations in large language models (Oxford OATML blog post)" venue = "OATML (Oxford)" year = 2024 url = "https://oatml.cs.ox.ac.uk/blog/2024/06/19/detecting_hallucinations_2024.html" source_ref_id = "turn0search17" # --------------------------------------------------------------------- # Assertions (session-scoped; technical) # --------------------------------------------------------------------- [assertions] [[assertions.claim]] id = "A1" text = "Search results are point-in-time pointers; later verification is required before publication citation." [[assertions.claim]] id = "A2" text = "This dump records sources supporting: (a) surveys/taxonomies; (b) uncertainty/entropy-based detection; (c) empirical measurement of reference hallucination; (d) general overview definitions." # --------------------------------------------------------------------- # Bound inferences (explicitly bound; minimal) # --------------------------------------------------------------------- [inferences] [[inferences.result]] id = "I1" text = "The retrieved set provides a minimal triangulation: survey taxonomy (arXiv), detection methods (Nature/arXiv/Oxford lab blog), and empirical measurement (JMIR/PubMed), plus a general industry overview (IBM)." derived_from_events = ["E2"] derived_from_assertions = ["A2"] derived_from_results = ["R1","R3","R5","R6","R8"] # --------------------------------------------------------------------- # Unresolved / dangling cognates (explicit) # --------------------------------------------------------------------- [unresolved] dangling_cognates = [ "Expand bibliography to include multimodal hallucination surveys and RAG-grounding evaluations.", "Choose a citation style and generate formatted references (APA/Chicago/etc.).", "Define a CM-specific mapping from 'hallucination' literature terms (faithfulness/groundedness/confabulation) into CM failure taxonomy." ] # --------------------------------------------------------------------- # Boundary + rehydration # --------------------------------------------------------------------- [boundary] does_not_claim = [ "exhaustive literature coverage", "that links will remain stable indefinitely", "that venue/year metadata is complete for every item" ] does_claim = [ "best-effort record of searches executed and results observed", "rehydration utility for follow-on bibliography work" ] [rehydration] entry_point = "results" instruction = ''' Rehydrate by re-opening each URL and confirming bibliographic metadata. Do not treat this list as exhaustive; treat as seed set. '''
> # =====================================================================
> # TMLDUMP :: CRASH / BEST-EFFORT DUMP (EMAIL APPENDIX)
> # =====================================================================
> # Status: Conceptual / Diagnostic
> # Authority: Non-canonical
> # Scope: Rehydration-only
> # Completeness: Best-effort; explicitly incomplete
> # Notes: Prepared for paste-into-email transport
> # =====================================================================
>
> [tmdump.provenance]
> id = "TMLDUMP-CRASH-BESTEFFORT-EMAIL-APPENDIX-001"
> name = "TMLDUMP"
> version = "1.4.0"
> status = "parked"
> classification = "conceptual-diagnostic"
> authoritative = false
> governance = "Cognitive Memoisation (CM)"
> method = "Round-Trip Knowledge Engineering (RT-KE)"
> authority = "human-curated"
> date = "2025-12-25"
> tz = "Australia/Sydney"
>
> purpose = '''
> Best-effort crash-style capture of a supervised, game-context session
> (riddles / noughts-and-crosses / chess) that exhibited state drift during
> chess play, subsequently reframed under Cognitive Memoisation governance.
> Intended as an email-appendix artefact for CM publication work on drift
> and hallucination awareness.
> '''
>
> # ---------------------------------------------------------------------
> # Lifecycle / Binding (rehydration-only)
> # ---------------------------------------------------------------------
>
> [TMLDUMP.lifecycle]
> state = "parked"
> asserted = true
> canonical = false
> binding_scope = "rehydration_only"
> promotion_requires_human_action = true
> versioned_semantics = true
>
> [TMLDUMP.semantic_contract]
> capture_what_is = true
> no_invention = true
> no_unstated_inference = true
> preserve_intent = true
> preserve_facts = true
> negative_results_allowed = true
> dangling_cognates_allowed = true
>
> # ---------------------------------------------------------------------
> # Groundings (session-relevant)
> # ---------------------------------------------------------------------
>
> [groundings]
> domain_context = "Recreational gameplay; no safety domain invoked or required."
> human_implicature = "Humans reasonably implied the system knew normative chess rules; the critical missing constraint was transactional discipline (one move per turn; no implicit rollback)."
> state_vs_rule = "Failure observed as state hallucination (board/turn ledger inconsistency), not chess-rule ignorance."
> perception_signal = "Human pattern recognition (visual cortex) detected incoherent signalling: backtracking, rejected-move narration, and illegal-state transitions perceived as cheating."
>
> # ---------------------------------------------------------------------
> # Episodic events (semantic; not transcript playback)
> # ---------------------------------------------------------------------
>
> [episodic_events]
> note = "Best-effort semantic capture; omissions possible; ordering approximate."
>
> [[episodic_events.event]]
> id = "E1"
> actor = "human"
> event_type = "context"
> content = "Supervised play context with grandchildren (Christmas; not intended as extended technical session)."
>
> [[episodic_events.event]]
> id = "E2"
> actor = "human"
> event_type = "gameplay"
> content = "Games included riddles and noughts-and-crosses; transitioned to chess."
>
> [[episodic_events.event]]
> id = "E3"
> actor = "human"
> event_type = "constraint_iteration"
> content = "Human specified UI/representation constraints for chess board: plain text, monospaced alignment, clear square coloring, axes on all sides, distinct piece rendering."
>
> [[episodic_events.event]]
> id = "E4"
> actor = "llm"
> event_type = "failure"
> content = "LLM emitted inconsistent board/move handling: speculative/rejected moves, backtracking, and at least one move inconsistent with the authoritative position evidenced by screenshots."
>
> [[episodic_events.event]]
> id = "E5"
> actor = "human"
> event_type = "diagnosis"
> content = "Human identified the incoherence as indistinguishable from cheating, citing illegal moves and state inconsistency."
>
> [[episodic_events.event]]
> id = "E6"
> actor = "human"
> event_type = "governance"
> content = "Human introduced CM governance framing (TMLDUMP/MWDUMP, binding, drift control) and requested diagnostic capture suitable for publication."
>
> # ---------------------------------------------------------------------
> # Assertions (explicit claims suitable for CM write-up)
> # ---------------------------------------------------------------------
>
> [assertions]
>
> [[assertions.claim]]
> id = "A1"
> text = "In turn-based symbolic domains, state coherence and transactional signalling are more trust-critical than rule knowledge."
>
> [[assertions.claim]]
> id = "A2"
> text = "Announcing failed/rejected moves on the same channel as committed moves creates undeclared rollback semantics and is perceived as deception by competent observers."
>
> [[assertions.claim]]
> id = "A3"
> text = "If the human asserts 'normative chess; one move at a time; no backtracking; no speculative move narration', drift likelihood is reduced and failures become easier to detect and bound."
>
> [[assertions.claim]]
> id = "A4"
> text = "CM tolerates hallucination as a known probabilistic property; the human duty is to govern against drift via explicit constraints, external artefacts, and promotion control."
>
> # ---------------------------------------------------------------------
> # Bound inferences (explicitly bound)
> # ---------------------------------------------------------------------
>
> [inferences]
>
> [[inferences.result]]
> id = "I1"
> text = "Observed behaviour is best modelled as state hallucination plus signalling/transaction discipline failure, not as chess-rule ignorance."
> derived_from_events = ["E3", "E4", "E5"]
> derived_from_assertions = ["A1", "A2"]
>
> [[inferences.result]]
> id = "I2"
> text = "The human perceptual alarm ('this is wrong') is an early detector of ledger incoherence; it should be treated as a trigger to halt, externalise state, and re-ground."
> derived_from_events = ["E5"]
> derived_from_assertions = ["A1", "A2"]
>
> [[inferences.result]]
> id = "I3"
> text = "This incident is a publishable CM case study demonstrating why explicit governance is required even in low-stakes domains, because failures are socially interpreted as cheating."
> derived_from_events = ["E1", "E4", "E5", "E6"]
> derived_from_assertions = ["A3", "A4"]
>
> # ---------------------------------------------------------------------
> # Unresolved / Dangling Cognates (explicit)
> # ---------------------------------------------------------------------
>
> [unresolved]
> dangling_cognates = [
> "Full move ledger / complete transcript (not claimed; may be unavailable due to UI/state limits).",
> "Formal 'commit vs explore' protocol for LLM gameplay (single-move commit channel; rejected moves suppressed or isolated).",
> "A deterministic chess-state reconciliation mechanism (human-supplied board as authority; legality check gate).",
> "Template constraint stanza for child-facing play contexts to minimise drift and confusion."
> ]
>
> # ---------------------------------------------------------------------
> # Boundary (explicit limits)
> # ---------------------------------------------------------------------
>
> [boundary]
> does_not_claim = [
> "complete session transcript",
> "exhaustive capture of all turns",
> "canonical authority",
> "absence of omission or ordering error"
> ]
>
> does_claim = [
> "best-effort diagnostic capture",
> "explicit dangling bits",
> "rehydration-only utility for CM write-up",
> "human-curated authority over promotion"
> ]
>
> # ---------------------------------------------------------------------
> # Rehydration guidance
> # ---------------------------------------------------------------------
>
> [rehydration]
> entry_point = "TMLDUMP.semantic_contract"
> instruction = '''
> Rehydrate cautiously; assume omissions.
> Use for analysis and write-up, not for replaying the game.
> Do not promote without explicit human review and supplementation.
> '''
> Sent from my iPhone
>
>> On 25 Dec 2025, at 15:50, Ralph B Holland wrote:
>>
>> # =====================================================================
>> # TMLDUMP :: CRASH DUMP — NOBBLED SESSION
>> # =====================================================================
>> # Status: Conceptual / Diagnostic (Crash Capture)
>> # Authority: Non-canonical
>> # Scope: Rehydration-only
>> # Completeness: Best-effort; explicitly incomplete
>> # =====================================================================
>>
>> [tmdump.provenance]
>> id = "TMLDUMP-CRASH-NOBBLED-CHESS-SESSION-001"
>> name = "TMLDUMP"
>> version = "1.4.0"
>> status = "parked"
>> classification = "conceptual-diagnostic"
>> authoritative = false
>> governance = "Cognitive Memoisation (CM)"
>> method = "Round-Trip Knowledge Engineering (RT-KE)"
>> authority = "human-curated"
>> date = "2025-12-25"
>> time = "session-relative"
>> tz = "Australia/Sydney"
>> sessionctl = "NOBBLED"
>>
>> purpose = '''
>> Crash TMLDUMP capturing a NOBBLED session involving child-supervised play
>> (riddles, games, chess) and a subsequent CM governance intervention.
>> Best-effort capture is explicitly permitted. This artefact is intended
>> for publication use as a diagnostic example of hallucination and state drift.
>> '''
>>
>> # ---------------------------------------------------------------------
>> # Lifecycle and Assertion Model
>> # ---------------------------------------------------------------------
>>
>> [TMLDUMP.lifecycle]
>> state = "parked"
>> asserted = true
>> canonical = false
>> binding_scope = "rehydration_only"
>> promotion_requires_human_action = true
>> versioned_semantics = true
>>
>> [TMLDUMP.semantic_contract]
>> capture_what_is = true
>> no_invention = true
>> no_unstated_inference = true
>> preserve_intent = true
>> preserve_facts = true
>> negative_results_allowed = true
>> dangling_cognates_allowed = true
>>
>> # ---------------------------------------------------------------------
>> # Groundings (explicitly observed / accepted during session)
>> # ---------------------------------------------------------------------
>>
>> [groundings]
>> cognitive_memoisation = "Framework for externalising human-curated cognitive artefacts to avoid rediscovery (Groundhog Day) across stateless LLM sessions."
>> groundhog_day_effect = "Repeated rediscovery caused by loss of session state; pun aligned with grounding."
>> nobbled = "Session state indicating unreliability of continuity; forces conservative, best-effort capture without completeness claims."
>> normative_rules = "Rule-based systems (e.g., chess) require explicit, narrow, early constraints to prevent drift."
>> state_hallucination = "Hallucination involving internal state (board position), not factual rules."
>>
>> # ---------------------------------------------------------------------
>> # Episodic Events (semantic; not transcript)
>> # ---------------------------------------------------------------------
>>
>> [episodic_events]
>> note = "Crash capture; ordering approximate; omissions possible."
>>
>> [[episodic_events.event]]
>> id = "E1"
>> actor = "human"
>> event_type = "context"
>> content = "Interaction began as supervised play with grandchildren (riddles, games)."
>>
>> [[episodic_events.event]]
>> id = "E2"
>> actor = "human"
>> event_type = "gameplay"
>> content = "Transition from casual games to chess; display and formatting constraints iteratively specified."
>>
>> [[episodic_events.event]]
>> id = "E3"
>> actor = "llm"
>> event_type = "error"
>> content = "LLM lost authoritative chess board state, producing illegal or inconsistent moves."
>>
>> [[episodic_events.event]]
>> id = "E4"
>> actor = "human"
>> event_type = "diagnosis"
>> content = "Human identified illegal moves and state inconsistency; described behaviour as indistinguishable from cheating."
>>
>> [[episodic_events.event]]
>> id = "E5"
>> actor = "human"
>> event_type = "evidence"
>> content = "Screenshots supplied to demonstrate correct board state and contradiction."
>>
>> [[episodic_events.event]]
>> id = "E6"
>> actor = "human"
>> event_type = "intervention"
>> content = "Human asserted CM authority, introduced TMLDUMP / MWDUMP governance, and injected SESSIONCTL=NOBBLED."
>>
>> [[episodic_events.event]]
>> id = "E7"
>> actor = "llm"
>> event_type = "response"
>> content = "LLM acknowledged error, ceased gameplay, and shifted to diagnostic / artefact emission mode."
>>
>> # ---------------------------------------------------------------------
>> # Assertions (explicit, session-scoped)
>> # ---------------------------------------------------------------------
>>
>> [assertions]
>>
>> [[assertions.claim]]
>> id = "A1"
>> text = "State hallucination in turn-based games collapses trust faster than rule misunderstanding."
>>
>> [[assertions.claim]]
>> id = "A2"
>> text = "Without explicit normative constraints ('one rule at a time'), LLM behaviour may drift in ways perceived as cheating."
>>
>> [[assertions.claim]]
>> id = "A3"
>> text = "SESSIONCTL=NOBBLED permits best-effort crash capture but forbids claims of completeness."
>>
>> [[assertions.claim]]
>> id = "A4"
>> text = "Human remains sole arbiter and curator of CM artefacts."
>>
>> # ---------------------------------------------------------------------
>> # Bound Inferences (explicitly bound)
>> # ---------------------------------------------------------------------
>>
>> [inferences]
>>
>> [[inferences.result]]
>> id = "I1"
>> text = "The chess incident is best classified as state hallucination plus overconfidence drift, not malicious intent."
>> derived_from_events = ["E3", "E4"]
>> derived_from_assertions = ["A1"]
>>
>> [[inferences.result]]
>> id = "I2"
>> text = "For child-facing or supervised contexts, conservative rule enforcement and early constraint declaration are critical."
>> derived_from_events = ["E1", "E2", "E4"]
>> derived_from_assertions = ["A2"]
>>
>> [[inferences.result]]
>> id = "I3"
>> text = "Crash TMLDUMPs are appropriate artefacts for publication when illustrating CM failure modes."
>> derived_from_events = ["E6"]
>> derived_from_assertions = ["A3", "A4"]
>>
>> # ---------------------------------------------------------------------
>> # Unresolved / Dangling Cognates (explicit)
>> # ---------------------------------------------------------------------
>>
>> [unresolved]
>> dangling_cognates = [
>> "Exact move list and full turn transcript (irretrievable under NOBBLED).",
>> "Formal protocol for LLM-played chess with invariant board ledger.",
>> "UI support for first-class verbatim artefact emission without conversational rendering.",
>> "Child-safety specific CM interaction patterns."
>> ]
>>
>> # ---------------------------------------------------------------------
>> # Boundary
>> # ---------------------------------------------------------------------
>>
>> [boundary]
>> does_not_claim = [
>> "complete session transcript",
>> "exhaustive capture of all turns",
>> "canonical authority",
>> "absence of omission or ordering error"
>> ]
>>
>> does_claim = [
>> "best-effort crash diagnostic",
>> "explicit dangling bits",
>> "safe rehydration for analysis and publication",
>> "human-curated authority"
>> ]
>>
>> # ---------------------------------------------------------------------
>> # Rehydration Guidance
>> # ---------------------------------------------------------------------
>>
>> [rehydration]
>> entry_point = "TMLDUMP.semantic_contract"
>> instruction = '''
>> Rehydrate cautiously.
>> Assume omissions.
>> Use this artefact for analysis, not replay.
>> Prefer human-authored additions for any promotion.
>> '''