DocMason synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
DocMason is a repo-native knowledge application whose reasoning executor is an external host agent. Python compiles and selects evidence, persists task state, checks artifacts and governs invoked answer closure. Authored skills direct the host's semantic interpretation, enrichment and user communication. This account covers the shipped knowledge application at commit f3970939593194a158ab7ea1d49995f06cbccc86, with applicability cutoff 2026-10-13. Its evidence is inspected implementation and doctrine/design from SRC-1, not execution or experiment. No private corpus, generated KB, host trace or evaluation run was supplied. Host inference, model mutability, permissions and isolation remain external; ancillary release-service deployment is excluded. See the boundary and source register and RT-CMP-host-agent.
From documents to a published evidence surface
Corpus placement or the shipped sample driver supplies inputs. Sync builds staged text, structure and renders, then conservative semantic outputs and reader indexes. The retained substrate is local files; natural-language summaries and claims are distinct from symbolic identifiers, provenance mappings and ranking fields. The host can add semantic overlays for assigned units without rewriting deterministic sidecars. These layers preserve inspectable source representations rather than replacing the corpus with summaries alone (MEM-OBJ-source-evidence, MEM-OBJ-knowledge-text, MEM-OBJ-reader-index, MEM-OBJ-semantic-overlay).
Normal sync admits publication when validation returns valid or warnings; unchanged state can skip it. Validation checks identity, artifact contracts, resolvable citations and nonempty support. Publication switches the current root and marks eligible interaction entries promoted. It does not independently establish that a cited passage entails a generated claim. The helper checks agreement with the copied validation report rather than performing fresh semantic validation; cooperating filesystem leases are best effort. Handled publication-bookkeeping errors have restoration paths, but arbitrary host edits and universal crash durability are not covered (RT-RTE-sync-publish, EPI-RTE-knowledge-check, RT-OBJ-publish-pointer).
The strongest inspected derivation is narrow: a generic conservative template can state that staged evidence includes a particular snippet. That follows from the stipulated staged input, not from the truth of the snippet's business assertion. Host-authored semantic production is requested, while its entailment, conjecture and semantic acceptance remain indeterminate without actual premise/output pairs. Fingerprints withdraw stale overlays' coverage credit, not necessarily all reader access to their text (EPI-RTE-conservative-output, RT-RTE-hybrid-refresh, EPI-RTE-overlay-coverage).
From a question to canonical delivery
The representative workflow is a corpus question through canonical ask and grounded-answer. The host opens a governed turn with stable identities and a thin work brief. The returned answer path, support contract, scope and status bind the inner workflow. Only an execute handoff starts that workflow; a workflow boundary is not evidence of a separately launched agent. Missing publication can route to sync, while usable stale publication can support an answer with a freshness notice. Confirmation and shared-job waits preserve the same turn (RT-RTE-canonical-ask, workflow account).
The host retrieves selected published bundles, inspects relevant text, structure and renders, and drafts a separate answer Markdown file linked by the symbolic turn record. Retrieval is requested read-back, with lexical-plus-graph selection and source/domain constraints. Compact payloads and file-first inspection reduce requested context, but do not establish host-window eviction, isolated side-agent frames or a total token budget (RT-RTE-evidence-tools, EPI-OBJ-answer, RT-OBJ-canonical-turn, RT-BAP-canonical-skills).
Trace evaluates the exact answer file and produces consequential support diagnostics. Its grounding classifier is heuristic, not an entailment oracle: the inspected code deduction shows that adding never to a positively supported sentence can leave vocabulary coverage at the grounding threshold. This is a countercase to a truth-oracle interpretation, not an observed full-gate bypass. Canonical admission separately checks ownership, versions, settled jobs, support bindings and scope. Those checks govern invoked commits; host adherence to final-reply discipline remains a protocol obligation. Public evidence tools and out-of-band host replies are not universally mediated (EPI-RTE-trace-check, RT-CLM-commit-gate).
Recovery separates missing-trace hydration from semantic repair. The wrapper can recover trace bindings and retry completion under its stated conditions. A repairable failure returns prepared state and a repair contract; the host performs the rewrite. The authored protocol permits one rewrite/retrace and another finalize, not an unbounded loop or a demonstrated hard retry counter. Hard-artifact gaps can authorize one narrowed, governed refresh; covered settlement requires post-refresh retrieval and trace. Final business delivery is permitted only by completed or boundary, with clean answer text separate from governance detail. Waiting and blocked states are not delivered ordinary answers (workflow recovery, RT-RTE-hybrid-refresh, RT-RTE-canonical-ask).
These controls suit the inspected scenario of scoped, provenance-linked work over a local corpus. They do not support the broader claim that all answers are KB-grounded or that hallucinated cross-source facts are prevented. External, mixed, model-knowledge and governed-boundary support bases are legal; provenance and closure legality have narrower warrants than factual correctness (EPI-CLM-source-trust, RT-CLM-commit-gate).
What experience changes for later work
DocMason has wired automatic trace-fed writes with later consumers. Native reconciliation retains user text, bounded assistant excerpts, attachments and derived policy metadata. Pending projections can reach current-target readers before publication, conditionally on domain and memory mode; sync can later group entries, produce semantics and publish them. External-factual queries exclude interaction records, as does workspace-corpus minimal mode. Other modes use lexical thresholds and rank priors. Labels such as confirmed can be fallback classifications, not corroboration. Promotion changes participation tier, not truth status (MEM-RTE-native-capture, MEM-RTE-interaction-promotion, EPI-RTE-pending-overlay, MEM-OBJ-interaction-policy).
This is a supported trace-learning pathway in the write-route sense: recorded sessions feed durable context and semantic artifacts that later retrieval can deliver. It does not demonstrate improved capacity for future action. Retained evidence is richer than displayed summaries but still bounded: assistant excerpts and summary fragments can omit reasons. Explicit continuation anchors reload linked task state; optional prompt hooks independently push selected advisory continuation reminders. Hook activation and resulting host behavior are unobserved (MEM-OBJ-interaction-entry, RT-RTE-host-hooks, profile verification).
A second contribution is trace-derived evaluator state. Runtime review groups conversation and tool records into retained reasons and benchmark candidates. Explicit promotion derives and appends selected regression cases; later suites consume cases and baselines. Review itself forbids production mutations. Replay can use supplied answer plans and judge trials, so its existence does not establish autonomous answer generation or causal product improvement (MEM-RTE-review-projection, MEM-OBJ-benchmark-candidates, RT-RTE-operator-eval).
The independent higher-order findings remain unresolved, not a bundled negative. The accepted runtime does not establish theory-builder conditions 1–4 separately: localized theory content, content-dependent consumption, content-directed criticism and retained criticism shaping iteration cannot be inferred from contracts, diagnostics or candidate retention alone. Improved future capacity from criticism of a consumed theory is unestablished. Reflective and autonomous operation are also unestablished at this boundary; descriptions of runtime state and explicit maintenance do not settle those properties. For self-improvement, the review-to-regression chain supplies an improvement-directed evaluation surface, but no supplied subsequent operation establishes exercised, evidence-responsive change to production behavior. Its objective is regression assessment; candidate adoption is explicit and its request control unresolved. Core replacement and adapter projection are separate maintenance channels, not demonstrated evidence-selected successors (RT-RTE-operator-eval, RT-RTE-core-update, RT-RTE-adapter-sync, epistemic bounded conclusion).
Actual source-to-output lineage and host execution traces would change the assessment of authoring, activation and delivery. A proposition-sensitive semantic acceptance mechanism would change the warrant assessment beyond current artifact and lexical checks. Objective-linked comparisons tracing a retained update into subsequent behavior would permit stronger learning or self-improvement attribution. Complete reader invalidation and unit-level withdrawal paths would change the freshness and forgetting assessment. No additional source reads or new evidence were used for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Overlay freshness is coverage-specific. | SRC-1; MEM-OBJ-semantic-overlay, EPI-RTE-overlay-coverage, RT-RTE-hybrid-refresh; report limit | Frozen hybrid fingerprint checks, overlay preservation, generic loading and reader projections. | Universal stale-text quarantine; matching fingerprints also do not establish semantic correctness. | A complete settlement-to-reader invalidation account showing stale text excluded on every included path. |
| Withdrawal does not cover every promoted unit or delivered context. | SRC-1; MEM-OBJ-interaction-entry, MEM-RTE-interaction-promotion, EPI-RTE-orphan-cleanup; report limit | Suppression skips promoted entries; orphan cleanup retains a group with any surviving backing entry. | Unit-level deletion, complete forget or retraction of already-delivered host context. | Traced per-unit edit/deletion propagation and downstream retraction mechanisms, with their actual coverage. |
| Capture, invalidation and later-consumer inventories are incomplete. | SRC-1; MEM-RTE-native-capture, MEM-OBJ-reader-index, MEM-RTE-review-projection; report limit | Selected interaction, retrieval, knowledge and projection paths; individual provider/attachment parsers, incremental cache combinations, worker freshness and answer-history consumers incompletely inspected. | Exhaustive capture, reader invalidation and all-consumer reuse comparisons; no absence follows. | Inspection tracing the named alternatives through their retained outputs and actual consumers. |
| Parser-recovery forgetting is omitted from the partial curation inventory. | SRC-1; MEM-OBJ-native-ledger; profile limit | Transcript-tail compaction keeps records needed to resume the last native turn. | Complete capture-recovery curation inventory or route-level comparison; this is not host-window summarization. | Classification of that bounded recovery-record forgetting operation against its retained parts and consumer. |
| Operation, semantic fidelity and benefit are not observed. | SRC-1; RT-CMP-host-agent, RT-RTE-canonical-ask, MEM-OBJ-source-evidence, RT-RTE-operator-eval; source register | Code/design only; no supplied private corpus, generated KB, host trace or evaluation result; external inference and isolation excluded. | Successful extraction, host compliance, delivered-answer quality, improved capacity and exercised self-improvement; system-complete negatives about opaque reasoning. | Provenanced corpus/output and host traces, plus objective-linked intervention/comparison evidence at the same declared boundary. |
| Initial provenance, control and semantic transformations remain locally unresolved. | SRC-1; MEM-OBJ-source-evidence, MEM-OBJ-knowledge-text, MEM-OBJ-semantic-overlay, RT-RTE-operator-eval | Known deterministic transformations coexist with unavailable private-input provenance, generic evaluation requests and unsupplied host premise/output pairs. | Complete lineage and admission-agency classifications; classification of all host semantic production as reshaping, entailment or ampliative conjecture. | Explicit content-supply/adoption decisions and actual semantic inputs, outputs, checks and dispositions. |