Recall (parcha-skills) synthesis verification

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

Verification

Scope and basis. This verification judges the handed synthesis against the boundary and the accepted runtime, memory, epistemic and workflow records. Superseded IDs were resolved through the reconciliation's four amendments. The synthesis cites only surviving IDs. No refusal was handed, so this is a first judgment and no answers are owed. No source file was read: no cited record was ambiguous enough to need a source read to settle a consequential statement. I read the text as a public reader would, without the members' context, and checked each substantive statement against the records it cites or relies on.

Carried limits and conflicts. The reconciliation retains no Unresolved conflict:. Its three named unknowns are rows in Limitations: index lifetime after source deletion, reach of recall_forget, and CLI search under a central or shadow signal. The report verification's single limit, on conditional loading of central-brain.md, is carried as a row and flagged in the body. The profile verification's single limit, on related output grouped under knowledge, is carried with its citations and prevented conclusion. Both carried limits are stated with their consequences.

Statements checked and found supported.

  • Evidence basis and boundary. The eight package files, the instruction-only test, the target class, the boundary kind and the tier match the boundary. Implementation evidence is kept apart from doctrine/design throughout. Operation, activation and benefit are stated as unobserved for every route.
  • Runtime map. Interface selection, including the CLI as a central client, matches RT-RTE-interface-selection. G1 is stated as model-enforced policy with its alternate paths. Single-agent coordination matches WF-ABS-agent-delegation. The tool-grant and host-model statements match RT-ABS-tool-restriction and RT-CMP-host-model.
  • Pull retrieval. The synthesis scopes pull to session-material routes and leaves generated-skill direction host-dependent. This follows MEM-ABS-push-hook, MEM-RTE-generated-skill-load and the reconciliation's scoped convergence. The text rejects a system-wide pull-only claim.
  • Two-stage selection. The engine gate keeps status claimed (EPI-RTE-engine-evidence-gate). The WHY check is model-run (EPI-RTE-why-assessment). The license "these terms occurred in this session" is the epistemic ledger's. The synthesis also states that related output has no described WHY line.
  • Bounded windows, the 80 MB size rationale, the token-free budget and the less bounded remote route match the memory core ideas and read-back section.
  • Declared filters and the cold-scan bypass match runtime G2, RT-CLM-ingest-redaction and RT-RTE-cold-scan. The text leaves the presence of thinking text as a host fact.
  • Trust by evidence form matches EPI-RTE-receipt-support-check, EPI-ABS-self-report-distinction and EPI-ABS-data-instruction-rule.
  • Writes without content approval match the runtime admission fields, the memory annotations and EPI-ABS-acceptance-record. Brain ingest is stated as named only.
  • Export. The producer is stated as absent. The consumer checks and the source_snapshot_stable gap are attributed to recap, matching G6 and the epistemic export row.
  • Workflow. The participants, the two rejection points, the absence of an answer check (WF-ABS-independent-answer-check), the session-file receipt, the four recovery branches and the empty-result wording conflict all match the workflow member and the surviving epistemic records.
  • Scenarios. Continue's branch check is stated as a reading of the command text, not an observed run. Repeat confirmation is stated as requested, not enforced. Single-store provenance in central mode is limited to the instruction's intent on the MCP path.
  • Learning and self-improvement. The runtime report gives no theory account, so withholding theory-builder, learning, reflection and autonomy conclusions is correct, and a Limitations row carries it. The self-improvement bullet follows the epistemic bounded conclusion: link 1 is unestablished, and links 3 and 4 depend on the host loader and on unobserved runs. "Supports at most a marked dispositional claim ... and only if" is a ceiling, conditioned on link 2. It asserts no such claim, and the preceding sentence states that link 1 is unestablished, so it does not strengthen the record. The method-revision finding matches the runtime change-admission account and RT-RTE-package-change-gate.

Objections considered and dismissed.

  • The description says "every check is model-enforced". Read in context, it covers the agent-side procedure. The same sentence names the engine and its ranking as declared only, and the package test is a repository change gate, not a runtime check. Dismissed.
  • EPI-CLM-subagent-lineage is cited for subagent transcripts being indexed as their own sessions. That record claims path lineage. The indexing statement rests on the memory core ideas, and the evidence-basis section frames every engine statement as declared. The citation is imprecise, but it changes no conclusion. Dismissed.
  • "recall_forget removes one Brain record" omits "as declared". The Limitations row on forget reach states the open deletion semantics. Dismissed.

Faults found. Four passages compress a record and widen or drop a condition. Each is a limit, not a blocker. In each case the rest of the synthesis or its Limitations already contains the condition, and stating the limit keeps the account bounded. None changes the assessment of a central mechanism or any emitted comparison value.

Blockers

none

Limits

  • Skill-ify, in Discriminating mechanisms and in the scenario "Building durable capability from past work". The synthesis says the model "writes a standard SKILL.md package" and "may also hand the recipe to an external skill creator". The records give the opposite order. The procedure invokes the harness skill creator when it is installed, and the direct SKILL.md write is the fallback (RT-RTE-skillify-admission progression; the boundary's "only a fallback instruction"). The scenario then says, without condition, "The package is written without evaluation, approval or provenance". EPI-ABS-generated-skill-test and MEM-ABS-skill-provenance support only that the shipped procedure requires none of these. Both records leave open whether an installed creator adds them, and so does the synthesis's own creator row in Limitations. Readers should not conclude that generated skills lack evaluation, approval or provenance when a creator is installed. The finding holds for what the package requires and for the direct-write fallback.
  • Representative workflow: the list of classes "unsearchable through the index" includes "host-deleted transcripts". This repeats the gap field of EPI-RTE-empty-result-report. Memory integration issue 6, concerning RT-OBJ-session-index and RT-OBJ-session-transcripts, leaves open whether index entries or snippets survive deletion of their source transcripts. The synthesis's own Limitations row states the same unknown. Readers should not conclude that host-deleted transcripts are unsearchable. Pi sessions, thinking-only traces and redacted lines remain unsearchable as declared, so the scenario "Proving that something did not happen" still holds.
  • Representative workflow: "Nothing persists the query, the rejected candidates or the answer." The workflow member scopes this to the skill ("The skill writes no record of the answer"). The host transcript retains the session, including the query and the answer. As declared, ingest makes that session retrievable by later invocations, including the Recall output it read back (memory annotation on RT-RTE-index-maintenance; later read-back field of EPI-RTE-answer-receipt). Readers should not conclude that Recall answers are lost after delivery. Only the skill's own writes omit them. They can return through host transcripts and the declared index.
  • Learning, reflection, autonomy and self-improvement: the strongest-contribution bullet says "Retained session experience is automatically turned into durable artifacts", and it lists Brain captures and skill-ify packages among them. Only index ingest is automatic in the ordinary sense. A user request starts each capture and each skill-ify run, and the model authors the content without per-content approval (RT-RTE-brain-capture, RT-RTE-skillify-admission; memory annotation on RT-OBJ-generated-skill: "user-triggered rather than automatic"). Readers should not conclude that Recall captures records or generates skills without a user request. "Automatically" holds there only in the sense that no human approves the written text.