Recall (parcha-skills) profile verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

What was checked. I checked the scope, all ten axes, every unit and every finding of the handed memory profile. The checks used the profile type, the record contract and the accepted runtime, memory, epistemic and workflow records. I resolved superseded IDs through the reconciliation's four amendments. For each unit I checked four things:

  • whether its scope is the axis's natural unit;
  • whether the cited records support the value and its basis;
  • whether the unit assessment matches its coverage;
  • whether the axis assessment matches the inventory.

I judged the profile afresh. The previous verification's two blockers and the profile author's answers were treated as arguments, not as established repairs. I read no source file, and the profile records no source reads.

Previous blockers.

  • Engine-ingest update consumer. The previous blocker is corrected. behavioral_authority now has a unit for host transcripts consumed by engine ingest. The unit carries learning at basis claimed and cites RT-RTE-index-maintenance, RT-OBJ-session-transcripts and RT-CLM-ingest-redaction. It is kept apart from the index output's downstream ranking and routing unit, and it matches the engine-ingest unit under trace_learning. claimed is the right basis: the ingest record splits trigger (wired) from ingest (claimed). The rationale answers the memory annotation's "retrievable evidence, not behavior-shaping instructions" with the type's rule that retained trace-derived context qualifies. That answer is consistent with the trace-learning finding.
  • recall_forget decay basis. The previous blocker is corrected. decay is now claimed, and the unit is partial. The unit names the missing service deletion semantics, which would separate decay from invalidate. EPI-RTE-receipt-forget supports that account: the service deletes, the removal is "as declared", and the deletion semantics are absent. A partial unit that keeps a classified finding and names its unresolved alternative matches the type.

Previous limit. The profile now includes host retention deletion (automatic and decay, both claimed). Generated-skill removal is not-determinable under both write agency and curation. The previous limit therefore no longer applies. The scope string does not name these two withdrawal mechanisms in its parenthetical list. The memory member's scope includes host retention deletion as withdrawal, and generated skills are an included object, so the surfaces still agree. I dismissed that wording point. Leaving generated-skill removal not-determinable is cautious but faithful. The records say only that removal is "manual". They do not record a human decision on particular content, and they do not record a route.

Supersessions. The YAML cites only surviving IDs: EPI-OBJ-search-results and EPI-RTE-why-assessment. The rationale lists all four amendments and correctly says which two carry classifications. No value depends on a superseded finding.

Axis dispositions.

  • storage_substrate (known): supported. Each retained object in scope has one substrate unit. files for transcripts and generated skills is wired. The index (sqlite), Brain content (service-object) and cursors (files) are claimed, and their records cite the absent engine and service as the reason. Synced content has a classified substrate but an unresolved form. That is consistent, because substrate and encoding are separate.
  • representational_form (partial): supported. Transcripts are split by consumption path: the model reads text, and the engine parses symbolic JSON. Index structures and model-read snippets are split the same way. Possible dense ranking state and the form of synced Brain content stay not-determinable. That matches RT-CMP-recall-engine and MEM-OBJ-brain-synced-content.
  • lineage (known): supported at the stated bases. Transcripts are imported because the target acquires them from the host. I considered whether sync should be imported, as a raw copy, rather than trace-extracted. The declared sync is from local Recall session history, so trace-extracted at claimed follows the record. The unit also keeps imported for the external connectors. I dismissed the question.
  • behavioral_authority (known): supported. The consumers and effects are:
  • knowledge for retrieved windows and Brain records;
  • instruction for repeat prompts and for generated skills, the latter at afforded through the host loader;
  • learning at both update consumers that read retained material (engine ingest and skill-ify);
  • ranking and routing for index content;
  • routing for cursors.

I considered three objections: - Mixed bases in the knowledge and routing units. Index-backed windows and WHY lines exist only through the claimed engine, yet these units carry wired. The type attaches a basis to the part or mechanism that supports the authority, and here that mechanism is the model's instructed consumption. The procedure wires it (EPI-RTE-why-assessment, RT-RTE-cold-scan). Production remains claimed under ranking and trace_learning. I dismissed this objection. - No learning unit for capture's update consumer. trace_learning counts capture as a qualifying write. The capture records, however, establish only that the model selects decisions or findings (RT-RTE-brain-capture). They do not establish that capture reads retained, read-back material rather than the current session's working context. Behavioral authority classifies retained parts, so this omission is not shown to be a gap. I dismissed it. - Continue playbook classified only as knowledge. RT-BAP-retrieved-sessions says retrieved state can drive new actions after confirmation. The records, however, place the effect in a resume summary that informs the user's decision (EPI-OBJ-resume-summary). They record no standing directions taken from the tail. I dismissed this objection.

The grouping of related output is a limit (see Limits). - write_agency (partial): supported. Capture and direct skill-ify are automatic, because the user starts the workflow but approves no text. That follows the type's rule that starting a workflow is not per-content admission. recall_forget is manual: EPI-RTE-receipt-forget records a user request and exact-receipt selection. Host writing, ingest, host retention deletion and cursor writing are automatic at claimed. Three included mechanisms remain unresolved: the external creator, Brain sync and import, and generated-skill removal. Those unresolved units make the axis partial. - curation_operations (partial): supported. consolidate for capture and recipe separation follows the type's extraction example. The memory annotation's "meaning-changing generalization ... as requested" does not establish synthesize. The record contract says requested behavior does not establish a new claim. EPI-RTE-recipe-separation also classifies the recipe as non-truth-apt. Both decay units are claimed. Engine and service curation are correctly not-determinable, and indexing and redaction establish no controlled operation. - read_back_direction (partial): supported. Pull is classified for the session-material routes under the bounded absence MEM-ABS-push-hook. Generated-skill direction stays not-determinable, as MEM-RTE-generated-skill-load and the reconciliation require. - read_back_signal (not-determinable): supported. The session-material unit is inapplicable on a bounded pull-only warrant. The unresolved generated-skill unit prevents an inapplicable axis. With no controlled findings, not-determinable is the type's value. - trace_learning (partial): supported. Engine ingest (claimed), capture (claimed) and skill-ify (afforded) each record an automatic trace-fed write whose retained output reaches later invocations. Brain sync stays not-determinable because its admission control is unknown. The system "yes" with partial coverage follows the type's example, and no "no" is emitted. - trace_source (partial): supported. Index values follow the declared ingest content. The capture unit is partial, with tool-trace origin unresolved. Skill-ify tool-traces rests on the commands and steps that EPI-RTE-recipe-separation records as taken from the window.

Consistency between axes. Write agency, trace learning and behavioral authority now treat engine ingest, capture and skill-ify consistently. Read-back direction and read-back signal share the same unresolved generated-skill unit. The decay-versus-invalidate question is stated in both the curation unit and the rationale.

Blockers

none

Limits

  • The behavioral_authority unit "Host transcripts and index-backed windows read by the current model in find/verify, continue, cold scan and related" groups the related output under knowledge. Its recorded effect is different. RT-RTE-related-sessions returns sessions that share a project, branch or touched files, as candidates for reading. The epistemic ledger row on that route gives its force as ranking and its operational authority as proposing candidates. That makes the effect closer to selecting which session to open. The system value set is unchanged, because ranking and routing are already supported for index content. The knowledge finding stays supported for windows the model reads. Readers should not take this unit as a per-route classification of related output, or compare query-free retrieval authority across systems from it.