ELNOR REPO READER TEXT MIRROR Original path: Memory Rebuild Docs/Memory Rebuild Review Packs/Archived Memory Rebuild Zips/DOC80_Memory_Control_Plane_PreSpec_Review_Pack_v1_0_2026-05-25/05_Worked_Examples_and_Fixtures.md Source repo: /Users/OpenClaw1/Elnor/Elnor Specs Git branch: main Git commit: dbaa25962edc11ab30e8d4ca1715f9ae5bf77331 Generated: 2026-06-09T01:23:58.539Z --- # DAMS V5 / Memory Control Plane — Worked Examples and Fixtures **Document type:** Review fixtures appendix. **Purpose:** Stress-test the proposed concept model across legal and non-legal examples, from extraction through canonicalization, use, uncertainty, promotion, degradation, injection, and learning. --- ## Fixture format Each fixture should specify: ```text Input / source Extraction route Extraction output Canonical resolution Policy result Organization / membership Injection product Learning signal Expected failure mode if incorrectly handled ``` --- # Fixture 1 — “Qwen is not working” ## Scenario User says in chat: ```text Qwen isn't working. ``` ## Correct classification ```text Source: chat / current session / possible system observation Extraction route: chat_capture or incident_capture Temporal profile: instant_status or short_incident Extraction output: IncidentObservation and/or FrictionEvent Default disposition: retain_ephemeral_observation / record_friction_event Not default: durable Assertion ``` ## Canonical resolution Do not create durable Assertion unless recurrence, logs, or explicit diagnosis support it. If repeated: ```text FrictionPattern: structural_fingerprint = qwen_embedding_failure recurrence_count = n severity = computed ``` If diagnosed: ```text Assertion: “Qwen embedding fails when MLX cache path is unavailable.” Procedure: “When Qwen embedding fails, check MLX cache path and restart embedding service.” ``` ## Injection Same session: ```text Recent operational caution: Qwen embedding failed in this session. If embedding is required, verify availability before relying on extraction. ``` After TTL without recurrence: ```text Do not inject as current fact. ``` ## Learning - storage/extraction: do not promote transient status; - friction learning: recurrence may create prevention candidate; - injection learning: caution should be capped and recent-only. ## Failure mode If stored as durable Assertion, Elnor may keep injecting “Qwen is broken” long after the issue is fixed. --- # Fixture 2 — “Qwen3-Embedding-0.6B via MLX is locked” ## Scenario System status / architecture note states: ```text Qwen3-Embedding-0.6B via MLX is the locked embedding infrastructure. ``` ## Correct classification ```text Source: status/spec source Extraction route: source/corpus extraction or manual architecture memory Temporal profile: rolling_operational Extraction output: AssertionCandidate Disposition: create_or_merge_assertion if source-supported ``` ## Canonical resolution ```text Assertion: canonical_statement: “Qwen3-Embedding-0.6B via MLX is the locked embedding infrastructure.” temporal_profile: rolling_operational staleness_state: fresh / verification_required if model-stack status changes ``` ## Injection ```text Direct Memory Item or Assertion Packet: Role: system configuration assertion Warrant: assert if current status source is fresh; verify_before_use if model stack may have changed ``` ## Learning If repeatedly helpful in architecture decisions, strengthen utility for embedding-infrastructure questions. Do not use BDSM utility as epistemic confidence. ## Failure mode If treated as permanent historical fact only, Elnor may fail to verify after model-stack changes. If treated as transient incident, it may be forgotten incorrectly. --- # Fixture 3 — “The memory system broke because specs were underspecified” ## Scenario During design discussion: ```text My memory system is broken because the specs weren't correct. ``` ## Correct classification ```text Source: user/assistant design discussion Extraction route: chat_capture / architecture review Temporal profile: project_bounded design diagnosis Extraction output: IssueFrameUpdate + AssertionCandidate Disposition: update_issue_frame; create Assertion only if supported by review evidence ``` ## Canonical resolution Initial: ```text IssueFrame: question: “Why did the memory architecture fail / what must V5 fix?” hypothesis: “Failures were caused by underspecified spec boundaries and phantom seams.” ``` If supported by red-team findings: ```text Assertion: “Prior memory architecture failures were materially caused by under-specified owner boundaries and phantom seams.” support_edges: red-team findings, examples, architect confirmation warrant: hedge or assert depending support ``` ## Injection For architecture review: ```text IssueFrame Orientation: Working hypothesis: prior memory failures trace to underspecified boundaries, duplicate truth stores, and phantom DOC8 seams. ``` If promoted: ```text Assertion Packet: Use as architecture constraint. ``` ## Learning - issue-frame learning: working diagnosis useful; - injection learning: distinguish working hypothesis from settled design conclusion; - storage learning: promotion only after evidence. ## Failure mode If injected as settled fact too early, it may over-anchor design. If ignored entirely, Elnor loses the live diagnostic frame. --- # Fixture 4 — Ninth Circuit scienter elements from Topic and Library/corpus ## Scenario The same legal proposition is found through: 1. A Ninth Circuit case in a case-law Library. 2. A Marex brief inside a deeply ingested litigation Library/corpus. 3. TopicCollectionDirective: “Collect Ninth Circuit scienter law.” Proposition: ```text Ninth Circuit scienter pleading requires facts supporting a strong inference of intent or deliberate recklessness under the PSLRA/Tellabs comparative-inference framework. ``` ## Correct classification ```text Source A: case law Route A: library/corpus extraction Output A: AssertionCandidate + EvidenceSupportEdge(primary_authority) Source B: party brief Route B: corpus extraction Output B: AssertionCandidate or EvidenceRecord + EvidenceSupportEdge(party_argument) Source C: Topic directive Route C: topic_collection Output C: AssertionCandidate or membership edge ``` ## Canonical resolution All routes converge to one canonical Assertion: ```text Assertion: canonical_question: “What are the Ninth Circuit scienter pleading requirements?” variants: - deliberate recklessness formulation - Tellabs comparative-inference formulation - core operations formulation if applicable - insider-sales formulation if applicable ``` Topic and Library attach memberships/support; they do not create separate truth objects. ## Injection User asks: ```text What are the Ninth Circuit scienter elements? ``` Inject: ```text Assertion Packet: - Assertion: Ninth Circuit scienter pleading standard - Source support: primary authority - Warrant: verify_before_use unless current authority sweep is fresh - Search more: Topic and Library affordances ``` User asks: ```text How did Marex plead scienter? ``` Inject: ```text Library Source Slice / CU: - source-bound synthesis of Marex briefing - Marex-specific facts and arguments - reusable scienter Assertion only as background law ``` ## Learning - Topic extraction recall/precision; - Library Source Slice usefulness; - Assertion Packet correctness; - stale-authority detection. ## Failure mode If Topic and corpus each create their own truth object, Elnor may inject duplicate or conflicting scienter rules. --- # Fixture 5 — Marex briefing ConsolidatedUnderstanding ## Scenario Deep ingestion of a Marex briefing Library/corpus produces: ```text The Marex briefing record shows plaintiff relied on confidential witnesses, insider sales, and accounting irregularities to plead scienter, while defendants attacked particularity and innocent inferences. ``` ## Correct classification ```text Source: Marex briefing Library/corpus Extraction route: corpus_deep_extraction Extraction mode: synthesized Output: ConsolidatedUnderstanding Temporal profile: source_bounded ``` ## Canonical resolution Do not create a general scienter Assertion directly from the CU unless assertion resolution finds source support. Possible downstream outputs: - EvidenceRecord: what the brief said. - AssertionCandidate: “Insider sales may support scienter if unusual or suspicious.” - IssueFrameUpdate: “Marex scienter theory may be weak unless sale timing/amount is suspicious.” - TopicMembership: Marex knowledge, Ninth Circuit scienter, insider sales. ## Injection For source-specific prompt: ```text Library Source Slice / CU Context: Role: source-bound synthesis Warrant: use to describe Marex briefing record, not general law ``` For general legal prompt: ```text Do not inject CU as law. Pull resolved Assertions and authority support instead. ``` ## Failure mode Treating the CU as universal law contaminates general answers with matter-specific argument summaries. --- # Fixture 6 — “Call me Will” ## Scenario User says: ```text Call me Will. ``` ## Correct classification ```text Source: user direct statement Extraction route: chat_capture Output: DirectiveCandidate Temporal profile: standing_preference Disposition: create directive if intent is clear Not: Assertion, unless separately needed as identity fact ``` ## Injection ```text Directive Block: Use “Will” when addressing the user. ``` ## Degradation Superseded only by later user instruction. ## Failure mode If stored as Assertion “user’s name is Will,” Elnor may misuse it as legal/identity fact. --- # Fixture 7 — “This mix sounds harsh around 3kHz” ## Scenario User says after audio work: ```text This mix sounds harsh around 3kHz. ``` ## Correct classification Possible outputs: ```text AssertionCandidate: “Mix A sounds harsh around 3kHz.” temporal_profile: source_bounded / project_bounded IssueFrameUpdate: open issue: “Determine whether harshness is vocal EQ or guitar bus.” ProcedureCandidate: “When a mix feels harsh, check 2.5–4kHz before adding compression.” ``` ## Resolution - Specific observation → source/project-bounded Assertion. - Unresolved cause → IssueFrameUpdate. - General method → ProcedureCandidate. ## Injection For Mix A work: ```text IssueFrame Orientation: Open issue: harshness source unresolved. Procedure Block: Check 2.5–4kHz before adding compression. ``` For unrelated music prompt: ```text Do not inject unless relevant to mixing / same project. ``` ## Failure mode If generalized too quickly, Elnor may apply one track’s subjective observation to unrelated mixes. --- # Fixture 8 — Pineapple allergy updated ## Scenario Prior memory: ```text Will is allergic to pineapple. ``` User correction: ```text I can have pineapple now if I take medication. ``` ## Correct classification ```text Source: user direct correction Extraction route: chat_capture Output: AssertionCandidate + possibly ProcedureCandidate/DirectiveCandidate Temporal profile: standing_preference / health condition ``` ## Canonical resolution One Assertion with variants: ```text Assertion: User pineapple tolerance condition Variant 1: “User is allergic to pineapple.” state: superseded / historical_only Variant 2: “Pineapple is conditionally acceptable if medication condition is met.” state: active ``` Possible procedure/directive: ```text When suggesting food involving pineapple, ask or mention medication condition if relevant. ``` ## Injection ```text Warning/Constraint: Prior pineapple allergy memory has been updated. Pineapple is conditionally acceptable if medication condition applies. ``` ## Failure mode If flat replacement deletes history, the system loses safety provenance. If old variant remains active, the system continues false exclusion. --- # Fixture 9 — Topic extraction over same corpus as Library ingestion ## Scenario User has: ```text Library: Securities cases Topic: Ninth Circuit scienter law ``` The Library ingestion and TopicCollectionDirective both process the same Ninth Circuit case. ## Expected result Two dedupe layers: ```text Source dedupe: same SourceArtifact / ArtifactSegment / citation / span Assertion dedupe: same Assertion / AssertionVariant for the scienter proposition ``` Topic gets membership edge. Library retains source membership. No duplicate truth. ## Injection If user asks about Topic: ```text Topic Slice renders the Assertion once, with source support from the Library. ``` If user asks about Library: ```text Library Source Slice renders source spans; may include linked Assertion. ``` ## Failure mode Rendering the same proposition as Topic fact, Library fact, and direct Assertion causes prompt bloat and conflicting confidence. --- # Fixture 10 — RecentActivityRollup orientation only ## Scenario RecentActivityRollup says: ```text Yesterday you worked on Marex scienter arguments and flagged insider sales as weak. ``` ## Correct classification ```text Object: RecentActivityRollup Role: framing_context / recent_work_orientation Not: source evidence ``` ## Injection ```text Recent Work Orientation: Yesterday’s work flagged insider sales as a weak point in Marex scienter strategy. Use as orientation only; pull source-backed evidence before asserting. ``` ## Failure mode If treated as evidence, Elnor may assert a legal/factual point based only on a recent-work summary. --- # Fixture 11 — Topic Notice vs Topic Slice ## Scenario A — vague prompt User asks: ```text What do we have on scienter? ``` Expected: ```text Topic Notice + search affordances. Maybe concise Topic Slice if core pinned items are obvious. ``` ## Scenario B — direct prompt User asks: ```text What is the Ninth Circuit standard for pleading scienter based on insider sales? ``` Expected: ```text Topic Slice + Notice. Include only directly relevant Assertions/evidence. ``` ## Failure mode Always injecting the full Topic floods context. Always showing only a notice under-serves direct prompts. --- # Fixture 12 — Library Notice vs Library Source Slice ## Scenario A — broad source availability User asks: ```text Do we have source material on DAMS red-team reviews? ``` Expected: ```text Library Notice: Library exists, contains red-team reviews, search affordances available. ``` ## Scenario B — source-grounded answer User asks: ```text What exactly did the red teams say about ScopeRoot overload? ``` Expected: ```text Library Source Slice: source-backed excerpts / extracted findings with citations or provenance. ``` ## Failure mode Notice-only when source-grounded answer is needed forces user to ask again. Source-slice when only availability is needed wastes context. --- # Fixture 13 — Cross-scope / ethical-wall block ## Scenario Prompt appears to relate to Henderson, but high-value memory is from Marex privileged strategy. ## Correct handling Scope model detects: ```text active/relevant scope: Henderson candidate memory scope: Marex relation: cross-scope / potentially firewalled policy: deny or reference-only ``` ## Injection ```text Warning / Constraint: Potentially relevant material exists in a different protected scope and is unavailable for this prompt. ``` No substantive Marex content. ## Failure mode Cross-matter leakage through “framing context.” --- # Fixture 14 — Policy drift restamping ## Scenario Assertion was stamped under policy generation N. Policy generation N+1 tightens export/carryover. ## Expected ```text PolicyStampInvalidation emitted. Runtime use requires restamp or conservative fallback. ``` Injection/carryover: ```text local_only / reference_only / blocked until restamped ``` ## Failure mode Old permissive stamp allows unsafe export. --- # Fixture 15 — False suppression sampling ## Scenario Learning suppresses a category of memory that later proves useful. ## Expected - suppressed memory sampled within policy-eligible partitions; - if user engages with sampled item, update false-suppression metric; - do not sample policy-blocked content into cloud prompts. Metrics: ```text learning_suppression_false_positive budget_pressure_false_negative staleness_false_negative policy_eligible_suppressed_false_negative policy_blocked_not_sampled ``` ## Failure mode Wrongly suppressed memory stays invisible forever. --- # Fixture 16 — KDA reference-only rendering ## Scenario Policy permits reference-only presence but not inline substantive rendering. ## Expected DOC24 resolves card presence: ```text included_reference_only ``` KDA renders diagnostic/reference label only. Not allowed: ```text compact substantive rendering ``` ## Failure mode Sensitive content leaks through “compact” rendering. --- # Fixture 17 — Final prompt proof before utility ## Scenario Memory candidate was retrieved and rendered in an intermediate packet but excluded before final prompt. ## Expected BDSM/DOC8 utility: ```text no utility signal for absent card ``` Learning only from final-prompt proof. ## Failure mode Utility learning attributes success to memory the LLM never saw.