Epistemic-architecture lens — run AAS-20260820-CCDW-01

Invoked procedure: kb/instructions/analyse-external-system-epistemic-architecture.md, called conditionally at step 7 of analyse-agentic-system. This lens does not own the boundary, the sources, the revision, the evidence tier, or any publication decision. It annotates the orchestrator's canonical records (CMP-*, OBJ-*, RTE-*, BAP-*, CLM-*) by existing ID and returns proposed new records unminted.

Wrapper constraints enforced: no source reacquisition; no boundary widening; no revision change; evidence tier fixed at doc-grounded; SRC-1 treated as doctrine/design; no route recorded as implemented; no parallel ID namespace; no system-wide epistemic grade; no publication decision.

1. Source-and-claim boundary

Field Value
System Claude Code dynamic workflows — the Workflow orchestration facility inside the Claude Code harness. Subsystem-only boundary; no whole-system conclusion about Claude Code is drawn or licensed here.
Reviewed revision / version SRC-1 captured 2026-06-03, documenting v2.1.154–v2.1.160-era behavior. Analysis cutoff 2026-06-03. Unchanged by this lens.
Declared scope As packet §3. Included: CMP-1..CMP-12. Excluded (named): Claude Code's ordinary turn-by-turn loop, skills, agent teams (except the doc's own comparison table cited as doctrine); the sub-agent primitive's internals (/en/sub-agents not in bundle); the /deep-research script's source text; model inference and provider routing; billing internals; the general permission engine beyond the quoted workflow rows.
Analysis question (bounded epistemic subquestion) Within the dynamic-workflows subsystem at the frozen 2026-06-03 documentation boundary, which material routes handle truth-apt content — acquiring, transforming, checking, disposing, retaining, or integrating it — and what does the documented evidence actually license about the system's stated warrant claim that adversarial multi-agent review and cross-checked synthesis yield "a more trustworthy result than a single pass" (CLM-1)?
Assessed route families Deep-research knowledge pipeline (RTE-E1..RTE-E6); pattern-capability quality routes (RTE-E7, RTE-E8); context selection and delivery of truth-apt material (RTE-X1, RTE-X2, RTE-X3, RTE-X6); retention and recovery of truth-apt material (RTE-S5, RTE-C6); artifact retention and installation (RTE-S1, RTE-S2); operational admission of the orchestration (RTE-C2); external action by agents (RTE-S4); direct behavior/policy adaptation with no truth-apt object (RTE-C1, RTE-C5, RTE-S3) — included under the run's direct-adaptation exception because they are material to authority even though they do not themselves trigger this lens.
Unassessed route families Sub-agent-internal reasoning, retrieval, and self-checking (RTE-X1 terminates at the prompt; the agent's own loop is outside the bundle). Search-engine ranking and retrieval behind RTE-E1. Model inference and provider routing. Usage accounting (RTE-S6) and the disable surface (RTE-C8) — inspected by the orchestrator as control, carrying no truth-apt content; not re-analysed here. Command-registry name resolution (RTE-C7) beyond its authority effect. No system-complete epistemic conclusion is drawn, and none is available at this boundary.
Source register SRC-1 → kb/sources/claude-code-dynamic-workflows-docs.md (from https://code.claude.com/docs/en/workflows), captured 2026-06-03, genre tool-announcement — evidence layer doctrine/design (vendor documentation; its permission and limit tables are doctrine about implemented behavior, not inspected implementation). SRC-2 → live-session tool-roster capture, 2026-08-20, host 2.1.237 — evidence layer observed run, scope roster-only / negative; supports no positive finding about the target.
Overall evidence tier doc-grounded, fixed by the orchestrator (packet §5). No implementation inspected, no observed run of the target, no causal experiment. No route in this lens reaches implemented or above.
System knowledge-production / warrant claims CLM-1 (warrant claim, SRC-1#when-to-use); CLM-2 (knowledge-production/operation claim, SRC-1#bundled); CLM-3 (#when-to-use), CLM-4 (#how-it-runs), CLM-5 (#limits, #cost), CLM-6 (#limits), CLM-7 (#when-to-use), CLM-8 (#approve) — the last six are operation, mechanism, control, isolation, design, and authority claims that bound what CLM-1/CLM-2 can mean.

Missing evidence → conclusion prevented

Missing evidence Conclusion it prevents
MG-1 — the /deep-research script text (excluded by boundary; never inspected) Prevents recording RTE-E1..RTE-E6 as implemented; prevents naming the vote's quorum, threshold, tie-handling, or abstention rule; prevents attributing the filter's criterion; prevents any component-level attribution among fan-out, cross-check, vote, filter, and synthesis.
MG-2 — no observed run of any workflow (SRC-2 is roster-only and negative) Prevents every observed candidate state in output 4; prevents any acceptance, rejection, revision, or integration finding for OBJ-9, OBJ-6, OBJ-4, OBJ-11; prevents upgrading CLM-2 from doctrine to reported or observed operation.
MG-3 — no baseline contrast, metric, or interventional comparison anywhere in SRC-1 Prevents any support for the comparative core of CLM-1 ("more trustworthy than a single pass"); prevents attributing trustworthiness to adversarial review, to cross-checking, or to agent count.
MG-4 — /en/sub-agents not in the bundle Prevents any conclusion about what a sub-agent reads, whether agent-internal checking exists, or what warrant OBJ-4 carries when it enters OBJ-3.
MG-5 — no documented claim-individuation step between fetched sources and voted claims Prevents classifying OBJ-9's transformation as non-ampliative reshaping vs. ampliative conjecture; prevents saying source warrant is preserved into OBJ-9. (See RTE-E3/extract and PROPOSED NEW RECORD 1.)
MG-6 — no documented persistence, export, or downstream consumer of OBJ-6 Prevents any retention or lifecycle-integration conclusion about workflow reports. The negative is scoped to documented routes; it does not assert that no informal user-side retention exists.
MG-7 — OBJ-3 persistence past process exit is uninspected (packet Loop C) Prevents any lineage conclusion about intermediate truth-apt results after the run's process ends.
MG-8 — version gap: SRC-1 pinned 2026-06-03 / v2.1.154–160; host under SRC-2 runs 2.1.237; no changelog in bundle Prevents any conclusion about present-day behavior of this subsystem. Preserved as a conflict, not resolved.
MG-9 — no inspected instance of a non-bundled workflow script Prevents generalizing any checking finding from /deep-research (CMP-11) to workflows as a class.

2. Epistemic-object inventory

Orchestrator-owned generic identity (packet §6.2) is reused verbatim by ID; this lens adds the truth-apt, lineage, role, and gap columns.

Object ID System name and description Representational form Source/input and lineage Producer / consumer Candidate truth-apt content or none Claimed role Evidence (source ID + anchor) Gap / limit
OBJ-1 Workflow script text symbolic (JavaScript) authored by CMP-3 from the user prompt (+ optionally an existing orchestrator pointed at) producer CMP-3 / consumers CMP-1, human reader, RTE-C2 gate none — an orchestration procedure, not a proposition. Its phase list summary is a description of itself, not a claim about the world. "the plan moves into code"; "what's repeatable is the orchestration itself" (CLM-7) SRC-1#how-it-runs, #when-to-use What the authoring model read is undocumented (MG-1-adjacent); the shipped /deep-research text is boundary-excluded.
OBJ-2 args global structured data supplied by the user via CMP-12 at invocation producer CMP-12 / consumer OBJ-1 none in the general case (parameters); may transport user-asserted content, unclassified here invocation-time input instead of script editing SRC-1#args No example of truth-apt payload in args documented.
OBJ-3 Script variables / intermediate results in-language values; may hold natural-language agent output written by RTE-X2 from OBJ-4; read by RTE-X1, RTE-E3..RTE-E6 producer/consumer OBJ-1 in CMP-1 heterogeneous container — its truth-apt content is instances of OBJ-4, OBJ-9, OBJ-11; the container itself is not a proposition keeps intermediate results out of CMP-12's context (CLM-3) SRC-1#when-to-use, #how-it-runs MG-7: persistence past process exit uninspected. Container split is honoured by the packet's own OBJ-4/OBJ-9/OBJ-11 rows.
OBJ-4 Agent result (with its prompt and recent tool calls) natural-language + tool-call trace produced by a sub-agent (CMP-4) from its prompt and its own tool use; agent-internal lineage not in bundle (MG-4) producer CMP-4 / consumers OBJ-3 (via RTE-X2), CMP-9 (via RTE-S5), human (via RTE-X6) yes — natural-language findings about a codebase, the web, or a task; truth-apt over the task's domain the unit the runtime tracks; the unit the human drills into SRC-1#watch-the-run, #how-it-runs No documented check applies to OBJ-4 in the general case; checking is documented only inside CMP-11.
OBJ-5 Phase record (agent count, token total, elapsed time) numeric/structured telemetry measured by CMP-1 instrumentation producer CMP-1 / consumer human via CMP-6 yes, non-ampliative — measurements truth-apt over the run itself progress and cost visibility SRC-1#watch-the-run, #cost Instrumentation is uninspected; no accuracy claim, no reconciliation against billing (#cost excluded).
OBJ-6 Final report / answer natural language; cited in the /deep-research case synthesized by RTE-E6 from surviving OBJ-9 over OBJ-10; in the general case, from OBJ-3 by undocumented means producer OBJ-1 / consumer CMP-12 and the human yes — the subsystem's sole delivered truth-apt product "one report at the end instead of a turn-by-turn transcript"; "Claude's context holds only the final answer" SRC-1#run-a-bundled-workflow, #when-to-use MG-6: no documented retention, export, or downstream consumer. Lands in a context window with advisory force (BAP-4) and no documented persistence.
OBJ-7 Saved workflow command file symbolic (JavaScript) + command name a specific OBJ-1 promoted by human keypress s producer human via RTE-S2 / consumer future sessions' resolver (BAP-2) none — a procedure, not a proposition turns a one-off run into a repeatable command SRC-1#save The promotion criterion is stated only as "if the run does what you wanted" — no recorded evaluation artifact.
OBJ-8 Permission consent record configuration entry written by human answer at RTE-C2 producer human / consumer CMP-5 none — an authorization, not a proposition suppresses the launch prompt SRC-1#approve Scope differs by mode (per workflow+project vs. user-wide under Auto); no expiry documented.
OBJ-9 Deep-research claim natural-language proposition with source citation derived from OBJ-10 by an undocumented individuation step (MG-5); then held in OBJ-3, then in OBJ-6 producer: undocumented (see RTE-E3/extract) / consumers RTE-E3/check, RTE-E4, RTE-E5, RTE-E6 yes — the central truth-apt candidate of this subsystem the unit that is cross-checked, voted on, filtered, and cited SRC-1#bundled, #run-a-bundled-workflow MG-5 blocks its transformation classification; MG-1 blocks its check semantics; MG-2 blocks every observed state.
OBJ-10 Fetched web source external document acquired from the open web by RTE-E1→RTE-E2 producer external / consumers RTE-E3, RTE-E6 (citation target) yes, acquired — external truth-apt content entering the system the evidence base claims are checked against SRC-1#run-a-bundled-workflow, #bundled No documented source-quality, authority, recency, or independence assessment. Source warrant on import: unknown.
OBJ-11 Cross-check / vote result on a claim disposition value produced by RTE-E3/check + RTE-E4 over OBJ-9 and OBJ-10 producer sub-agents/script / consumer RTE-E5 yes, second-order — a proposition about OBJ-9's standing across fetched sources decides which claims survive into the report SRC-1#bundled Value space, threshold, and tie rules undocumented (MG-1). Not documented as retained or surfaced to the reader.
OBJ-12 Tool allowlist configuration user settings; inherited by CMP-4 producer user / consumer sub-agent tool executor none — an authorization bounds what agents may call SRC-1#approve Inheritance is stated, not inspected.
OBJ-13 Session effort setting (ultracode) configuration set by user via /effort producer user / consumer CMP-3 planning none — a policy setting makes Claude plan a workflow for every substantive task SRC-1#ultracode Session-scoped; resets on new session.
OBJ-14 Session model selection configuration set by user via /model; overridable per stage by OBJ-1 producer user/script / consumer CMP-4 none — a policy setting controls which model backs each agent SRC-1#cost Per-stage routing is stated as a script capability, no shipped instance documented.

Omitted object families and the conclusions their omission prevents: sub-agent-internal scratch state and retrieved context (not in bundle — prevents any warrant conclusion about how OBJ-4 was produced); search-engine result ranking state (prevents any conclusion about OBJ-10 selection bias); provider-side inference state (prevents any conclusion about model behavior). None of these can be reached without widening the boundary, which this lens must not do.

3. Authority-route ledger

One functional kind per record. Linked rows share a base RTE-* ID with a functional suffix; no new base IDs are minted. Architectural status is recorded independently of function, and — per the fixed doc-grounded tier — no record reaches implemented.

3.1 Deep-research knowledge pipeline (CMP-11)

RTE-E1 — search fan-out - route function: content transformation - architectural status: doctrine only — declared at SRC-1#bundled; no implementation inspected (MG-1), no run observed (MG-2) - object/candidate id: OBJ-10 (candidate source references entering) - content/update relation: truth-apt transformation: acquisition/import — external search results enter the system; source warrant: unknown (no ranking, authority, or independence criterion documented). The angle decomposition that precedes the searches is a non-truth-apt policy/content update: query set derived from the question, undocumented in mechanism. - transition or check target: none — this route acquires, it does not check - evaluator/condition and domain: none documented; the external search engine's ranking is outside the boundary - activation and timing: on /deep-research <question> invocation, first phase of the run - possible or observed result: a set of source references; no instance observed - implemented force: none demonstrable at this tier; within doctrine it seeds every later route - epistemic authority and scope: none — acquisition licenses nothing about truth - operational authority: permits the fetch phase to proceed - behavioral-authority path: consumer OBJ-1 in CMP-1 | channel script control flow | force enforcing (BAP-1) | horizon one run - evidence: SRC-1#bundled, #run-a-bundled-workflow - claim IDs: CLM-2 - mismatch marker: none - gap/limit: MG-1, MG-2. Prevents any conclusion about coverage, recall, or angle independence.

RTE-E2 — source fetch - route function: content transformation - architectural status: doctrine only - object/candidate id: OBJ-10 - content/update relation: truth-apt transformation: acquisition/import — document content enters; source warrant preserved as received, i.e. unknown; no filtering, dating, or provenance check documented - transition or check target: none - evaluator/condition and domain: none documented - activation and timing: after RTE-E1 within the run - possible or observed result: fetched documents; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: none — imported content is acquired, not produced, and carries whatever warrant its origin had, which is unrecorded - operational authority: supplies the material the cross-check operates on - behavioral-authority path: consumer OBJ-1 | channel script variables OBJ-3 | force enforcing (BAP-1) | horizon one run - evidence: SRC-1#bundled, #run-a-bundled-workflow - claim IDs: CLM-2 - mismatch marker: none - gap/limit: MG-1, MG-2. Prevents concluding that fetched sources are independent of one another — the property the later cross-check's value depends on.

RTE-E3/extract — claim individuation from fetched sources (linked row; implied, not declared) - route function: content transformation - architectural status: no route found within boundary — neither inspected implementation, observed operation, nor doctrine establishes this step. SRC-1 speaks of "each claim" and of cross-checking sources "against each other" without ever documenting how a claim is individuated from a source. - object/candidate id: OBJ-9 (produced), from OBJ-10 - content/update relation: indeterminate — classifications still possible: (a) non-ampliative reshaping if a claim is a quotation or faithful extraction of a source's assertion; (b) ampliative conjecture if a model composes a proposition that no single source asserts. Evidence needed to decide: the script text or an observed run artifact showing claim provenance (MG-1, MG-5, MG-2). - transition or check target: none — this is the step that creates the check target - evaluator/condition and domain: none - activation and timing: necessarily between RTE-E2 and RTE-E4; undocumented - possible or observed result: OBJ-9 instances; no instance observed - implemented force: not determinable - epistemic authority and scope: none established — and because this step is unclassified, no warrant can be traced from OBJ-10 into OBJ-9 - operational authority: none recorded - behavioral-authority path: none recorded (no consequential consumer distinct from RTE-E4) - evidence: SRC-1#bundled (presupposition only) - claim IDs: CLM-2 (presupposed by it) - mismatch marker: yes — CLM-2 asserts an operation over "each claim" while the source documents no route that produces claims. This is a doctrine-internal gap, not a doctrine-versus-implementation gap. - gap/limit: MG-5. See PROPOSED NEW RECORD 1.

RTE-E3/check — cross-check of a claim against the fetched sources (linked row) - route function: check/evidence production - architectural status: doctrine only - object/candidate id: OBJ-9 checked; OBJ-11 produced - content/update relation: no content change to OBJ-9; produces a new second-order object OBJ-11 - transition or check target: named before evaluator — the target is an individual OBJ-9 claim, and the domain is agreement among the sources this run fetched, not truth - evaluator/condition and domain: model-backed sub-agents comparing OBJ-10 documents "against each other" (SRC-1#run-a-bundled-workflow); a hybrid model procedure with no documented rubric; domain limited to retrieved-source agreement - activation and timing: mid-run, after fetch, before vote - possible or observed result: a per-claim corroboration signal; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: at most "corroborated across the sources retrieved by this run at this time" — explicitly not veracity, not source independence, not absence of common-origin error, not completeness of the evidence base - operational authority: its result is consumed by RTE-E5 to admit or drop the claim - behavioral-authority path: consumer RTE-E5 within OBJ-1 | channel script variable OBJ-11 | force enforcing on report membership | horizon one run - evidence: SRC-1#bundled, #run-a-bundled-workflow - claim IDs: CLM-2, and it is the closest shipped analogue to the mechanism CLM-1 invokes - mismatch marker: yes, scope mismatch — CLM-1 speaks of trustworthiness; this route's evaluator domain reaches only inter-source agreement. - gap/limit: MG-1, MG-2, MG-4. Prevents naming the rubric, the comparison unit, or the handling of a claim supported by exactly one source.

RTE-E4 — vote on each claim - route function: check/evidence production (aggregation of the cross-check into a per-claim result) - architectural status: doctrine only - object/candidate id: OBJ-9 (target), OBJ-11 (result) - content/update relation: no content change - transition or check target: the individual claim OBJ-9; domain: aggregate agent judgement over the cross-check evidence - evaluator/condition and domain: a multi-agent vote — plural model evaluators; quorum, threshold, weighting, abstention, and tie rules all undocumented - activation and timing: after cross-check, before filtering - possible or observed result: survive / not survive; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: majority-of-model-evaluators agreement over retrieved sources — an aggregation of judgements, not a measurement; it licenses no claim about accuracy, and voting agreement among instances of the same model is not evidence of independence - operational authority: determines RTE-E5's input - behavioral-authority path: consumer RTE-E5 | channel OBJ-11 | force enforcing on report membership | horizon one run - evidence: SRC-1#bundled - claim IDs: CLM-2 - mismatch marker: none within doctrine; unknown against implementation - gap/limit: MG-1. Prevents stating the decision rule, and therefore prevents predicting the filter's behavior on any given claim.

RTE-E5 — filter (survival disposition) - route function: disposition/acceptance — kept strictly separate from the checking rows above - architectural status: doctrine only - object/candidate id: OBJ-9 - content/update relation: no content change to the surviving claim; the run's claim set is reduced - transition or check target: transition of each OBJ-9 from candidate to included-in-report or dropped - evaluator/condition and domain: consumes OBJ-11; the criterion is stated only as "didn't survive cross-checking", i.e. the criterion is named by outcome, not by rule - activation and timing: before synthesis, at report-assembly time - possible or observed result: included / dropped; no instance observed - implemented force: none demonstrable at this tier; within doctrine, enforcing on report membership - epistemic authority and scope: this is the only place in the subsystem where anything resembling an acceptance transition occurs, and it is a weak one. Survival is admission by non-rejection, not an affirmative recorded decision; the doc names no intended use, no reliance scope, and no persisted acceptance artifact. It does not license reliance on a surviving claim beyond "not contradicted by the sources this run happened to fetch". - operational authority: permits a claim to appear in OBJ-6; blocks the rest from appearing - behavioral-authority path: consumer human + CMP-12 reader | channel OBJ-6 via RTE-X3 | force advisory (BAP-4) | horizon session / turn - evidence: SRC-1#bundled, #run-a-bundled-workflow - claim IDs: CLM-2 - mismatch marker: yes — the doc presents the filter as a report-quality property, but records no criterion, no acceptance artifact, and no scope; dropped claims are not documented as retained, so the rejection is also unrecoverable. - gap/limit: MG-1, MG-2. Prevents any statement of the acceptance criterion, and therefore prevents calling a surviving claim "accepted knowledge".

RTE-E6/synth — cited synthesis (linked row) - route function: content transformation - architectural status: doctrine only - object/candidate id: OBJ-6 produced from surviving OBJ-9 - content/update relation: indeterminate — classifications still possible: non-ampliative reshaping (ordering, grouping, and formatting surviving claims) or ampliative conjecture (synthesis prose asserting propositions no surviving claim asserts). "Synthesizes a cited report" does not decide between them. Evidence needed: the script text or an observed report artifact with claim-level attribution coverage (MG-1, MG-2). - transition or check target: none - evaluator/condition and domain: none — no check is documented on the synthesized text itself; the checks upstream apply to claims, not to the synthesis that assembles them - activation and timing: final phase of the run - possible or observed result: one cited report; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: none produced by this route. Any warrant OBJ-6 carries is inherited from RTE-E5 survivors, and the inheritance is unverified because synthesis-introduced content is unchecked. - operational authority: produces the run's sole deliverable - behavioral-authority path: consumer CMP-12 + human | channel RTE-X3 | force advisory (BAP-4) | horizon session / turn - evidence: SRC-1#bundled, #run-a-bundled-workflow - claim IDs: CLM-2 - mismatch marker: yes — the report is presented as carrying the filter's quality property, while the step that composes it is itself unchecked. - gap/limit: MG-1, MG-2. Prevents concluding that the report contains only surviving claims.

RTE-E6/lineage — per-claim citation binding (linked row) - route function: lineage/freshness/recovery - architectural status: doctrine only - object/candidate id: citation pairs binding OBJ-9 to OBJ-10 (see PROPOSED NEW RECORD 2) - content/update relation: no content change — provenance attached alongside content - transition or check target: none - evaluator/condition and domain: none — attribution is asserted, not verified; no citation-accuracy check documented - activation and timing: at synthesis - possible or observed result: a report that "cites the sources each claim came from"; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: this is the subsystem's most consequential epistemic feature and it is reader-side, not system-side. Preserved lineage lets a human re-derive or contest a claim; it issues no system warrant, and it is not itself checked. - operational authority: none — citations do not gate anything downstream - behavioral-authority path: consumer human reader | channel OBJ-6 text | force advisory | horizon session / turn (no documented persistence) - evidence: SRC-1#run-a-bundled-workflow ("It cites the sources each claim came from") - claim IDs: CLM-2; see PROPOSED NEW RECORD 3 - mismatch marker: none - gap/limit: MG-2, MG-6. Prevents concluding that citation coverage is complete or that attributions are correct.

RTE-E7 — adversarial peer review (generic pattern) - route function: check/evidence production - architectural status: doctrine only, and weaker than the deep-research rows — SRC-1#when-to-use states this as something a workflow "can" be written to do. It is a capability of the pattern; no shipped workflow is documented as doing it. The packet's anti-conflation check "capability is not deployment" applies directly. - object/candidate id: unnamed findings (a generic OBJ-4-class object); no shipped instance - content/update relation: no content change documented; would produce review results - transition or check target: another agent's findings, "before they're reported" - evaluator/condition and domain: independent model agents in an adversarial posture; independence is asserted by role assignment, and no mechanism for it is documented - activation and timing: only if a script author writes it; not active in any documented shipped workflow - possible or observed result: none documented; no instance observed - implemented force: none — a capability statement has no consumer - epistemic authority and scope: none. A capability that no named artifact exercises licenses nothing. - operational authority: none - behavioral-authority path: none recorded - evidence: SRC-1#when-to-use - claim IDs: CLM-1 — this is one of the two routes CLM-1's warrant rests on - mismatch marker: yes, and it is the central mismatch of this lens — the subsystem's strongest warrant claim is attached to a route with no named shipped instance. - gap/limit: MG-1, MG-2, MG-9. Prevents any conclusion that adversarial review happens in this subsystem at all.

RTE-E8 — multi-angle drafting and weighing (generic pattern) - route function: check/evidence production - architectural status: doctrine only — same capability status as RTE-E7 - object/candidate id: candidate plans (unnamed; no shipped instance) - content/update relation: no content change documented; the weighing would produce comparative results - transition or check target: several independently drafted plans, weighed "against each other" - evaluator/condition and domain: model agents; no rubric, no criterion, no weighting rule documented - activation and timing: only if a script author writes it - possible or observed result: none documented; no instance observed - implemented force: none - epistemic authority and scope: none - operational authority: none - behavioral-authority path: none recorded - evidence: SRC-1#when-to-use - claim IDs: CLM-1 - mismatch marker: yes — same as RTE-E7 - gap/limit: MG-1, MG-2, MG-9.

3.2 Context, retention, and recovery routes for truth-apt material

RTE-X1 — script → sub-agent prompt - route function: operational admission/selection/consumption - architectural status: doctrine only - object/candidate id: content selected from OBJ-3/OBJ-2 into a prompt consumed by CMP-4 - content/update relation: indeterminate — classifications still possible: non-ampliative reshaping (templated concatenation of prior results) or ampliative conjecture (a composed framing asserting more than its inputs). The selection signal is script-authored code, not model judgement at dispatch — which makes reshaping the more likely class but does not establish it without the script text (MG-1). - transition or check target: none — no check is documented between an agent's result and its reuse in another agent's prompt - evaluator/condition and domain: none in the general case; only CMP-11 documents downstream checking - activation and timing: at each dispatch, per RTE-C4 - possible or observed result: a composed prompt; no instance observed (though RTE-X6 makes prompts inspectable in principle) - implemented force: none demonstrable at this tier - epistemic authority and scope: none — selection into context is not endorsement; unchecked OBJ-4 content can seed a later agent's premises - operational authority: determines what each worker works from - behavioral-authority path: consumer sub-agent CMP-4 | channel spawn prompt | force directive (BAP-3) | horizon that agent's lifetime - evidence: SRC-1#when-to-use, #watch-the-run - claim IDs: CLM-3, CLM-7 - mismatch marker: none - gap/limit: MG-1, MG-4. Prevents any conclusion about warrant propagation between workflow stages.

RTE-X2 — sub-agent result → script variables - route function: retention (pre-acceptance; explicitly not lifecycle integration) - architectural status: doctrine only - object/candidate id: OBJ-4 into OBJ-3 - content/update relation: no content change - transition or check target: none - evaluator/condition and domain: none documented — results are retained unchecked - activation and timing: on each agent completion - possible or observed result: values in OBJ-3; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: none — retention is not acceptance - operational authority: makes results available to later stages - behavioral-authority path: consumer OBJ-1 | channel script variables | force enforcing on control flow (BAP-1) | horizon one run - evidence: SRC-1#when-to-use, #how-it-runs - claim IDs: CLM-3 - mismatch marker: none - gap/limit: MG-7.

RTE-S5 — agent-result tracking - route function: retention - architectural status: doctrine only - object/candidate id: OBJ-4 in CMP-9 - content/update relation: no content change - transition or check target: none - evaluator/condition and domain: none - activation and timing: continuously as the run progresses - possible or observed result: a per-agent result store; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: none — this is the packet's "retaining material is not selecting it into context" line, and it holds: most OBJ-4 content documentedly never reaches CMP-12 - operational authority: enables RTE-C6 resume and RTE-X6 inspection - behavioral-authority path: consumer CMP-1 scheduler and CMP-6 UI | channel runtime store | force enforcing on resume behavior | horizon session-scoped - evidence: SRC-1#how-it-runs - claim IDs: CLM-4 - mismatch marker: none - gap/limit: MG-1, MG-2. Store contents and eviction are uninspected.

RTE-C6 — resume (cached-result replay) - route function: lineage/freshness/recovery - architectural status: doctrine only - object/candidate id: cached OBJ-4 - content/update relation: no content change — a completed agent's earlier result is reused rather than recomputed - transition or check target: none — no freshness or revalidation check is documented on a replayed result - evaluator/condition and domain: none - activation and timing: on resume, within the same session only; exiting Claude Code starts the workflow fresh - possible or observed result: cached results returned, remainder run live; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: none, and this is a freshness limit worth naming — a replayed result is reused without re-evidence, so a resumed run mixes results produced at different times against a possibly changed world. Applicability is not endorsement. - operational authority: permits a partially completed run to finish without redoing work - behavioral-authority path: consumer CMP-1 | channel CMP-9 | force enforcing on what reruns | horizon same session only - evidence: SRC-1#resume, #how-it-runs - claim IDs: CLM-4 - mismatch marker: none - gap/limit: MG-2, MG-8.

RTE-X3 — final answer → main session - route function: operational admission/selection/consumption - architectural status: doctrine only - object/candidate id: OBJ-6 - content/update relation: no content change - transition or check target: none — no check is documented on the report at delivery - evaluator/condition and domain: none in the system; the human reader is an evaluator only informally, with no recorded result - activation and timing: at run completion - possible or observed result: report in CMP-12's context; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: none issued — arrival in a context window is not acceptance and not endorsement - operational authority: the report becomes available to influence the conversation - behavioral-authority path: consumer main-session Claude + user | channel context delivery | force advisory (BAP-4) | horizon session / turn - evidence: SRC-1#when-to-use, #run-a-bundled-workflow - claim IDs: CLM-3 - mismatch marker: none - gap/limit: MG-6. Prevents any retention or integration conclusion about the report.

RTE-X6 — run internals → human - route function: other — evidence surfacing without a recorded evaluation - architectural status: doctrine only - object/candidate id: OBJ-4, OBJ-5 - content/update relation: no content change - transition or check target: none — this route makes evidence available; it does not evaluate - evaluator/condition and domain: the human, informally; no criterion, no recorded verdict, no artifact - activation and timing: any time during or after a run, via /workflows drill-down - possible or observed result: the human reads a prompt, recent tool calls, and a result; no instance observed - implemented force: none by itself — a surfaced result with no recorded consumer has no implemented force; force appears only when the human acts through RTE-C5 - epistemic authority and scope: none issued by the system - operational authority: none directly - behavioral-authority path: consumer human | channel CMP-6 TUI / Desktop pane | force advisory | horizon the human's attention span during the run - evidence: SRC-1#watch-the-run - claim IDs: none - mismatch marker: none - gap/limit: The run does not wait for the human (SRC-1#limits, "No mid-run user input"), so this surface is asynchronous by construction and cannot gate anything.

3.3 Artifact retention, installation, and admission (non-truth-apt objects with epistemic consequence)

RTE-S1 — run-script archive write - route function: retention - architectural status: doctrine only - object/candidate id: OBJ-1 - content/update relation: non-truth-apt policy/content update: the run's orchestration procedure is written to disk under the session directory - transition or check target: none - evaluator/condition and domain: none - activation and timing: every run, at start - possible or observed result: a readable, diffable script file; no instance observed - implemented force: none until relaunched - epistemic authority and scope: none — but note the asymmetry it creates: the orchestration is archived every run while the report is not documented as persisted at all (MG-6) - operational authority: enables reading, diffing, editing, and relaunching - behavioral-authority path: consumer Claude/user on request | channel path delivered by RTE-X4 | force material for relaunch — advisory until relaunched, then BAP-1 (BAP-9) | horizon persists on disk under the session directory - evidence: SRC-1#how-it-runs - claim IDs: CLM-7 - mismatch marker: none - gap/limit: MG-2.

RTE-S2 — save run script as command - route function: retention, linked to operational admission/selection/consumption for future /<name> invocation. This is deliberately not recorded as lifecycle integration: no truth-apt claim was accepted beforehand, so the procedure's post-acceptance condition is not met. - architectural status: doctrine only - object/candidate id: OBJ-7 (from a specific OBJ-1) - content/update relation: non-truth-apt policy/content update: a one-off orchestration becomes an installed, repo-shareable command - transition or check target: the orchestration procedure, not any proposition - evaluator/condition and domain: the human, on the stated criterion "if the run does what you wanted" (SRC-1#ask-in-prompt) — a satisfaction judgement with no recorded artifact, no rubric, and no scope statement - activation and timing: human presses s in /workflows, after a run; save dialog chooses project or personal location - possible or observed result: /name available in future sessions; project entry beats personal on collision; no instance observed - implemented force: none demonstrable at this tier - epistemic authority and scope: none over any truth-apt content. This is the subsystem's only human evaluation that durably changes future behavior, and its target is the procedure, not any claim the procedure produced. Success of the bundle does not license any conclusion about an individual component of it. - operational authority: permits future sessions to run that orchestration by name - behavioral-authority path: consumer future sessions' command resolver + Claude/user | channel CMP-7 registry lookup | force enforcing on invocation (BAP-2) | horizon persistent across sessions; project-scoped (repo-shared) or user-scoped - evidence: SRC-1#save, #ask-in-prompt - claim IDs: CLM-7 - mismatch marker: yes, worth flagging — CLM-7 says what is repeatable is "the orchestration itself"; the ledger confirms that and shows the corollary the doc does not state: nothing the orchestration concluded is made repeatable, retained, or accepted. - gap/limit: MG-2, MG-6.

RTE-C2 — launch approval - route function: operational admission/selection/consumption - architectural status: doctrine only (its permission table is doctrine about implemented behavior) - object/candidate id: OBJ-1 - content/update relation: no content change - transition or check target: admission of a script to execution; not a check on any proposition - evaluator/condition and domain: the human, on an unstated criterion, deciding from a summary (the planned phases) with full script text available only on opt-in (View raw script, Ctrl+G); on Desktop, name + phase list + token caution. Table-driven configuration selects between human decision and automatic start. - activation and timing: pre-execution, per run; Default/accept-edits → every run unless "don't ask again"; Auto → first launch only, and skipped entirely when ultracode is on; Bypass permissions / claude -p / Agent SDK → never - possible or observed result: run / cancel / adjust prompt via Tab; no instance observed - implemented force: within doctrine, enforcing on whether the run starts - epistemic authority and scope: none — approving an orchestration says nothing about the truth of what it will produce - operational authority: permits or blocks the entire run, including all agent file edits downstream - behavioral-authority path: consumer CMP-5 gate | channel OBJ-8 consent record | force permissive — suppresses the prompt (BAP-5) | horizon persistent per workflow+project (default/acceptEdits) or per user (Auto) - evidence: SRC-1#approve - claim IDs: CLM-8 - mismatch marker: none - gap/limit: This is the run's only pre-execution human control point, it decides on a summary by default, and it is removable by mode or by one persistent consent record.

RTE-S4 — agent-mediated external action - route function: operational admission/selection/consumption (action execution admitted by inherited authorization) - architectural status: doctrine only - object/candidate id: none truth-apt; the filesystem/shell is the target - content/update relation: no content change to any truth-apt object; the external world changes - transition or check target: none — no check on OBJ-4 or any claim precedes an action - evaluator/condition and domain: the allowlist OBJ-12 (configuration, not evaluation) plus forced acceptEdits - activation and timing: mid-run, whenever an agent acts; non-allowlisted shell/web/MCP calls can still prompt - possible or observed result: files written, commands run; no instance observed - implemented force: within doctrine, enforcing - epistemic authority and scope: none — and this is the sharpest asymmetry in the subsystem: action is unconditional and mid-run, while checking exists only inside one bundled research workflow and applies only to report content. - operational authority: permits file edits regardless of session mode; the script itself has no filesystem or shell access (CLM-6) - behavioral-authority path: consumer sub-agent tool executor | channel inherited OBJ-12 + forced acceptEdits | force permissive/enforcing (BAP-6) | horizon the run - evidence: SRC-1#limits, #approve - claim IDs: CLM-6, CLM-8 - mismatch marker: none - gap/limit: MG-4. Prevents any conclusion about agent-internal restraint before acting.

3.4 Direct behavior/policy adaptation with no truth-apt object (run's direct-adaptation exception)

These do not by themselves trigger this lens; they are recorded because they are material to authority and would otherwise leave the authority picture incomplete.

RTE-S3 — consent record write - route function: behavior/policy adaptation - architectural status: doctrine only - object/candidate id: OBJ-8 - content/update relation: non-truth-apt policy/content update: a human's one-time approval becomes a standing suppression of the launch prompt - transition or check target: the gate's future activation - evaluator/condition and domain: the human, at one launch, on an unstated criterion - activation and timing: on "Yes, and don't ask again for <name> in <path>"; under Auto, any Yes records consent in user settings - possible or observed result: later launches start without prompting; no instance observed - implemented force: within doctrine, enforcing on gate behavior - epistemic authority and scope: none — one approval is not a warrant for future runs, and the doc records no re-evaluation trigger even if the saved script is later edited - operational authority: removes the only pre-execution human control point for that workflow - behavioral-authority path: consumer CMP-5 | channel project/user settings | force permissive (BAP-5) | horizon persistent, per workflow+project or per user - evidence: SRC-1#approve - claim IDs: CLM-8 - mismatch marker: none - gap/limit: MG-2. Prevents knowing whether an edited script re-triggers the prompt.

RTE-C1 — trigger → script authoring - route function: behavior/policy adaptation - architectural status: doctrine only - object/candidate id: OBJ-13 - content/update relation: non-truth-apt policy/content update: a session setting changes Claude's default planning behavior - transition or check target: the planning path for every substantive task - evaluator/condition and domain: model judgement, keyword- or setting-gated; under /effort ultracode the model decides per task, and one request can become several workflows in a row - activation and timing: per prompt (keyword / natural language) or per session (/effort ultracode); resets on new session; disableable in /config - possible or observed result: a workflow is planned instead of turn-by-turn work; no instance observed - implemented force: within doctrine, directive - epistemic authority and scope: none - operational authority: changes which execution path a task takes, and under Auto+ultracode it also suppresses the launch prompt entirely — two authority reductions composing - behavioral-authority path: consumer main-session Claude's planning (CMP-3) | channel OBJ-13 session setting | force directive default (BAP-8) | horizon current session; resets on new session - evidence: SRC-1#ultracode, #ask-in-prompt, #approve - claim IDs: none - mismatch marker: none - gap/limit: MG-2. What the authoring model reads is uninspected.

RTE-C5 — interactive mid-run control - route function: behavior/policy adaptation - architectural status: doctrine only - object/candidate id: none truth-apt; the target is the run or an agent - content/update relation: non-truth-apt policy/content update: a human judgement from telemetry stops, restarts, or pauses execution - transition or check target: the run's continuation - evaluator/condition and domain: the human, from RTE-X6 telemetry; no criterion, no recorded result, no artifact - activation and timing: any time during a run; asynchronous — the run does not wait - possible or observed result: pause/resume, stop agent, stop run, restart agent, save script; no instance observed - implemented force: within doctrine, enforcing on execution - epistemic authority and scope: none — stopping or letting a run continue is not a verdict on its content. Operational continuation must not be read as epistemic warrant. - operational authority: halts or resumes work; stopping preserves completed work - behavioral-authority path: consumer CMP-1 | channel CMP-6 keypress | force enforcing | horizon the current run - evidence: SRC-1#watch-the-run, #limits - claim IDs: none - mismatch marker: none - gap/limit: Staged human sign-off is impossible within one workflow; it requires splitting into separate workflows (SRC-1#limits).

3.5 Evidenced absences within the boundary

  • No route found within boundary: acceptance of OBJ-6. No documented evaluator, criterion, intended use, or scope attaches to the delivered report. Searched: #run-a-bundled-workflow, #when-to-use, #how-it-runs, #save, #bundled. This absence is scoped to documented routes; it does not assert that no informal user-side acceptance occurs.
  • No route found within boundary: retention or export of OBJ-6. Only OBJ-1/OBJ-7 are documented as persisted. Searched: #how-it-runs, #save, #resume.
  • No route found within boundary: checking of OBJ-4 in the general (non-/deep-research) case. Searched: #when-to-use, #how-it-runs, #limits, #watch-the-run.
  • No route found within boundary: retention of rejected OBJ-9 claims or of OBJ-11 vote results. Filtered claims are described only as removed from the report. Searched: #bundled, #run-a-bundled-workflow.
  • No route found within boundary: any comparison of a workflow's output against a single-pass baseline. Searched: whole file. This is the absence that blocks CLM-1.

4. Per-object lifecycle disposition

Architectural status and observed candidate state are recorded separately throughout. Because the tier is doc-grounded with no observed run of the target (MG-2), every observed candidate state below is no instance observed — doctrine alone cannot establish one, and SRC-2 is roster-only and negative.

OBJ-9 — deep-research claim (primary candidate)

  • candidate object ID: OBJ-9
  • relevant route IDs: RTE-E1, RTE-E2, RTE-E3/extract, RTE-E3/check, RTE-E4, RTE-E5, RTE-E6/synth, RTE-E6/lineage, RTE-X3
  • transformation: indeterminate, not established as ampliative. The discovery-lifecycle schema is therefore applied provisionally and flagged: MG-5 prevents deciding between faithful extraction (non-ampliative reshaping) and model-composed proposition (ampliative conjecture). Classifications still possible: both. Evidence needed to decide: the /deep-research script text or an observed run artifact showing per-claim provenance.
  • observation/anomaly: routes RTE-E1, RTE-E2 | architectural status doctrine only | observed candidate state no instance observed | evidence SRC-1#bundled (doctrine/design). The "observation" here is web retrieval, i.e. acquisition of others' assertions, not observation of a phenomenon by this system.
  • conjecture: route RTE-E3/extract | architectural status no route found within boundary — the step is presupposed by CLM-2 but never declared | observed candidate state no instance observed | evidence SRC-1#bundled (presupposition only).
  • derived consequence: no route found within boundary. SRC-1 documents no step in which a claim's testable consequences are derived before checking; cross-checking compares the claim to retrieved text directly. Architectural status no route found within boundary | observed candidate state no instance observed.
  • test/evidence: routes RTE-E3/check, RTE-E4 | architectural status doctrine only | observed candidate state no instance observed | evidence SRC-1#bundled, #run-a-bundled-workflow (doctrine/design). Test domain is agreement among retrieved sources, not truth; the evaluator is a multi-agent vote whose rule is undocumented (MG-1).
  • acceptance: route RTE-E5 | evaluator: the vote result OBJ-11, consumed by the filter | criterion: stated only by outcome ("didn't survive cross-checking"); no rule, threshold, or rubric documented | intended use: none stated | architectural status doctrine only | observed candidate state no instance observed | accepted scope: none stated by the system; the most the route could license is "not contradicted by the sources this run retrieved, as of the run" | evidence SRC-1#bundled. Finding: this is admission by non-rejection, not a recorded evidence-consuming acceptance decision naming criterion, intended use, and scope. No acceptance transition in the procedure's sense is documented anywhere in this subsystem.
  • lifecycle integration: routes RTE-E6/synth, RTE-X3 | post-acceptance change/consumer: the claim appears in OBJ-6, which lands in CMP-12's context with advisory force (BAP-4) and no documented persistence | architectural status doctrine only | observed candidate state no instance observed; and independently not reached as a matter of doctrine — appearing in a report and being read is use, not integration: nothing is connected to evidence for later reliance, nothing changes the system's organization, and no store is updated | evidence SRC-1#run-a-bundled-workflow, #how-it-runs, and the evidenced absence in §3.5.
  • missing phase/evidence: MG-1, MG-2, MG-5. Missing phases: consequence derivation (no route), acceptance-with-named-scope (route present but criterion and scope absent), integration (no route). The prevented conclusions: that /deep-research produces accepted knowledge; that surviving claims carry any system-issued warrant; that any claim was in fact checked in any run.

OBJ-6 — final report / answer

  • candidate object ID: OBJ-6
  • relevant route IDs: RTE-E6/synth, RTE-E6/lineage, RTE-X3
  • transformation: indeterminate — non-ampliative reshaping of surviving OBJ-9 claims, or ampliative conjecture where synthesis prose asserts more than the claims do. Classifications still possible: both.
  • preserved lineage: per-claim citations to OBJ-10 (RTE-E6/lineage) — asserted, unverified, with no documented coverage guarantee.
  • implemented checks, retention, or use: no check at or after synthesis (§3.5); no documented retention (MG-6); use is advisory delivery into CMP-12 (RTE-X3, BAP-4, horizon session/turn).
  • current warrant limit: whatever warrant survives from RTE-E5, degraded by an unchecked composition step. The system issues no warrant of its own for the report.
  • evidence needed to decide preservation, entailment, or ampliation: an observed report artifact with per-claim attribution, plus the script text (MG-1, MG-2).
  • observed candidate state across all phases: no instance observed.

OBJ-4 — agent result

  • candidate object ID: OBJ-4
  • relevant route IDs: RTE-X1 (as later input), RTE-X2, RTE-S5, RTE-C6, RTE-X6
  • transformation: indeterminate, ampliation not established — an agent's natural-language findings may report what it read (acquisition/reshaping) or assert conclusions beyond it (ampliative conjecture); the producing loop is outside the bundle (MG-4).
  • discovery lifecycle: not applicable as established — ampliation is not established, so the lifecycle schema is not forced onto it. Recorded instead as an indeterminate object.
  • preserved lineage: the agent's prompt and recent tool calls are retained alongside the result and are inspectable via RTE-X6 — a genuine lineage strength at the doctrine layer.
  • implemented checks, retention, or use: no check documented in the general case; retained by RTE-X2 and RTE-S5; replayed unrevalidated by RTE-C6; reused as another agent's premise via RTE-X1; surfaced to the human by RTE-X6 with no recorded verdict.
  • current warrant limit: none established. Retention, replay, and reuse as premises all occur before and without any check.
  • evidence needed: the sub-agent primitive's documentation and an observed run (MG-4, MG-2).
  • observed candidate state: no instance observed.

OBJ-11 — cross-check / vote result

  • candidate object ID: OBJ-11
  • relevant route IDs: RTE-E3/check, RTE-E4, RTE-E5
  • transformation: indeterminate — an aggregated evaluative judgement about OBJ-9. Classifications still possible: entailed derivation if the vote is a deterministic function of recorded per-source comparisons; ampliative conjecture if the models judge holistically. Undecidable without the script text (MG-1).
  • preserved lineage: none documented — the result is not documented as retained or surfaced to the reader.
  • implemented checks, retention, or use: unchecked; consumed once by RTE-E5; not documented as retained.
  • current warrant limit: licenses only inter-source agreement as judged by model evaluators of unknown independence, at the moment of the run.
  • evidence needed: the vote rule and an observed run (MG-1, MG-2).
  • observed candidate state: no instance observed.

OBJ-10 — fetched web source

  • candidate object ID: OBJ-10
  • relevant route IDs: RTE-E1, RTE-E2, RTE-E3/check, RTE-E6/lineage
  • transformation: truth-apt transformation: acquisition/import
  • discovery lifecycle: not applicable — imported content is acquired, not produced by this system.
  • applicable acquisition, lineage, derivation, or update route and warrant: acquired by RTE-E1→RTE-E2 with source warrant unknown — no authority, recency, independence, or quality criterion is documented. Lineage is preserved forward into OBJ-6 by RTE-E6/lineage.
  • missing evidence/limit: MG-1, MG-2. Prevents any conclusion that fetched sources are mutually independent — the property RTE-E3/check's value depends on.

OBJ-5 — phase record (telemetry)

  • candidate object ID: OBJ-5
  • relevant route IDs: RTE-X6, RTE-S6
  • transformation: truth-apt transformation: acquisition/import (instrumented measurement of the run)
  • discovery lifecycle: not applicable — non-ampliative measurement.
  • applicable acquisition, lineage, derivation, or update route and warrant: produced by CMP-1 instrumentation, surfaced to the human; warrant rests on uninspected instrumentation, and no reconciliation against billing is available (that surface is boundary-excluded).
  • missing evidence/limit: MG-1. Prevents any accuracy claim about token totals or timings.

Per-object no-candidate lines

  • No lifecycle record for OBJ-1: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-S1, RTE-S2, RTE-C2, RTE-C3.
  • No lifecycle record for OBJ-2: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-X5, RTE-C7.
  • No lifecycle record for OBJ-3: heterogeneous container with no candidate truth-apt output of its own; its truth-apt content is dispositioned under OBJ-4, OBJ-9, and OBJ-11; relevant routes: RTE-X1, RTE-X2.
  • No lifecycle record for OBJ-7: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-S2, RTE-C7.
  • No lifecycle record for OBJ-8: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-S3, RTE-C2.
  • No lifecycle record for OBJ-12: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-S4.
  • No lifecycle record for OBJ-13: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-C1, RTE-C8.
  • No lifecycle record for OBJ-14: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-C4 (per-stage model routing).

The global no-candidate statement does not apply: the inventory contains candidate truth-apt output (OBJ-4, OBJ-6, OBJ-9, OBJ-10, OBJ-11, OBJ-5).

5. System-claim versus route comparison

Claim ID Claimed operation or warrant Claim source ID/anchor + evidence layer Doctrine/design support Implemented route IDs Observed-run support Causal support and design limits Supported conclusion Mismatch / unknown
CLM-1 Independent agents adversarially reviewing each other's findings, or drafting a plan from several angles and weighing them, gives "a more trustworthy result than a single pass". Warrant claim. SRC-1#when-to-use; doctrine/design Yes, but only as a capability of the pattern (RTE-E7, RTE-E8), phrased with "can"; no shipped workflow is named as doing either none — nothing reaches implemented at this tier (MG-1) none — no run observed; SRC-2 is roster-only and negative (MG-2) none. SRC-1 contains no baseline, no metric of trustworthiness, no comparison, and no experiment (MG-3). No contrast exists, so no causal identification is even attemptable; component attribution among review, multi-angle drafting, and agent count is impossible. Only that Anthropic states this as a design rationale for the workflow pattern. The comparative core — "more trustworthy than a single pass" — has no doctrine-internal argument, no implementation, no run, and no measurement behind it. Within this boundary it is not merely unsupported but untested and unfalsifiable. Mismatch, two layers. (1) The warrant claim attaches to RTE-E7/RTE-E8, which have no named shipped instance; the one shipped instance (CMP-11) implements a different pattern (cross-check → vote → filter), so even doctrine does not connect CLM-1's mechanism to shipped behavior. (2) "Trustworthy" is never operationalized; the nearest shipped evaluator domain (RTE-E3/check) reaches only inter-source agreement.
CLM-2 /deep-research "votes on each claim, and returns a cited report with claims that didn't survive cross-checking filtered out". Knowledge-production / operation claim. SRC-1#bundled; doctrine/design Yes — RTE-E1..RTE-E6 are declared as the shipped behavior of a named bundled workflow, which is stronger doctrine than CLM-1's capability phrasing none (MG-1; the script text is boundary-excluded) none (MG-2) none. No comparison of filtered vs. unfiltered output; no accuracy measurement; nothing licenses attributing report quality to the filter rather than to retrieval, model, or question. That the vendor declares a shipped check-and-disposition pipeline with a named evaluator form (a multi-agent vote) and a named disposition (survival filtering), plus preserved per-claim citation lineage. That is a real architectural commitment and is the strongest epistemic structure in the subsystem. It licenses nothing about: the vote rule (quorum, threshold, ties — all undocumented); whether filtering tracks truth rather than retrieved-source consensus; whether surviving claims are accepted for any named use or scope; whether the synthesis step introduces unchecked content; or whether any run behaved this way. Doctrine-internal gap: the claim quantifies over "each claim" while no route producing claims from sources is documented (RTE-E3/extract, MG-5). Scope unknown: whether "cross-checking" is claim-vs-source or source-vs-source is not resolved by the text.
CLM-3 "A workflow script holds the loop, the branching, and the intermediate results itself, so Claude's context holds only the final answer." SRC-1#when-to-use; doctrine/design Yes; consistent with RTE-X2/RTE-X3 and the packet's anti-conflation check none none none That the declared context economy is OBJ-4 → OBJ-3 (not CMP-12), with only OBJ-6 delivered. Epistemic corollary the doc does not state: the human's default view of the run's evidence is one synthesized report, and the underlying results are reachable only by opt-in drill-down (RTE-X6) that gates nothing. Unknown whether "only the final answer" is exact (the archive path also arrives via RTE-X4).
CLM-4 "The runtime tracks each agent's result as the run progresses, which is what makes a run resumable within the same session." SRC-1#how-it-runs; doctrine/design Yes — RTE-S5 retention, RTE-C6 recovery none none none That retention exists for recovery purposes and is session-bounded. Retention is not acceptance, and replayed results are reused without revalidation (RTE-C6). Unknown whether cached results are revalidated on resume; nothing suggests they are.
CLM-5 The 1,000-agent cap "prevents runaway loops"; caps "bound the cost of a runaway script". SRC-1#limits, #cost; doctrine/design Yes — BAP-10 enforcing ceiling none none none That a ceiling is declared. It bounds count and spend, not correctness; a bounded run can still be wrong. Unknown how the cap interacts with RTE-C6 resume accounting.
CLM-6 "No direct filesystem or shell access from the workflow itself" — agents act, the script coordinates. SRC-1#limits; doctrine/design Yes — RTE-S4, the architectural line that the orchestrator is a pure coordinator none none none That the orchestrator is declared non-acting. This is an isolation property of the script, not a check on action: agents act unconditionally in acceptEdits, and no documented check on OBJ-4 precedes their actions. None internal; the safety reading it invites (that action is constrained) is not what the claim says.
CLM-7 What is repeatable in a workflow is "the orchestration itself"; the plan moves into code. SRC-1#when-to-use; doctrine/design Yes — RTE-S1, RTE-S2, BAP-2 none none none That the procedure is the retained, shareable, repeatable unit. The corollary this lens adds: nothing the procedure concluded is retained, accepted, or made repeatable. The subsystem codifies method, not findings. None internal; a real asymmetry against MG-6.
CLM-8 "Your permission mode controls only the launch prompt" — spawned sub-agents always run in acceptEdits and inherit the allowlist regardless of session mode. SRC-1#approve; doctrine/design Yes — BAP-6, RTE-S4, RTE-C2 none none none That authority is decided once, pre-execution, on a phase summary, and that the session's own mode does not constrain agents thereafter. None internal. Bounds every authority conclusion in §6.

No further consequential public or design claim was found in SRC-1 beyond CLM-1..CLM-8 and the two proposed additions in the returns section.

6. Bounded conclusion

Scoped to the dynamic-workflows subsystem at the 2026-06-03 documentation boundary, on doc-grounded evidence (SRC-1 = doctrine/design; SRC-2 = negative, roster-only). No whole-system conclusion about Claude Code, and no system-wide epistemic grade, is given or available. Every finding below is a finding about a route, at the doctrine layer.

What it retains, retrieves, reshapes, or uses. The subsystem retains two things durably: the orchestration script, archived every run (RTE-S1) and optionally installed as a repo-shareable command (RTE-S2, BAP-2); and per-agent results, retained session-scoped for recovery (RTE-S5, RTE-C6). It retains no findings — OBJ-6 has no documented persistence, export, or downstream consumer (MG-6, §3.5). Retained agent results are reused as later agents' premises (RTE-X1) and replayed on resume without revalidation (RTE-C6), in both cases before and without any check. Retention here is not acceptance and not integration.

What it acquires, and whether source warrant is preserved. RTE-E1→RTE-E2 import external web content (OBJ-10) with source warrant unknown: no authority, recency, independence, or quality criterion is documented. Provenance is preserved forward as per-claim citations in the report (RTE-E6/lineage) — the single most consequential epistemic feature in the subsystem, and it is reader-side: it lets a human re-derive or contest a claim; it issues no system warrant and is itself unverified.

What it derives. Nothing is documented as an entailed derivation from warranted premises within a declared formal domain. OBJ-11 (the vote result) is the only plausible candidate and is indeterminate without the vote rule (MG-1).

What it conjectures, tests, accepts, and integrates. Claims (OBJ-9) appear in the pipeline without a documented step that produces them from sources — RTE-E3/extract is presupposed by CLM-2 and never declared (MG-5), which blocks classifying OBJ-9 as reshaped or conjectured and blocks tracing warrant from OBJ-10 into it. Claims are then tested (RTE-E3/check, RTE-E4) in a domain that is agreement among the sources this run retrieved, not truth, by model evaluators whose independence is asserted by role rather than mechanism. Disposition (RTE-E5) is admission by non-rejection: survival is named by outcome, with no rule, no intended use, no reliance scope, and no persisted acceptance artifact. No acceptance transition in this procedure's sense — a recorded, evidence- consuming decision against a named criterion for a named use and scope — is documented anywhere in this subsystem. Consequently lifecycle integration is not reached as a matter of doctrine and no instance observed as a matter of evidence: a surviving claim's terminus is a report delivered advisorily into a context window (RTE-X3, BAP-4, horizon session/turn). Consequence derivation has no route at all. /deep-research is best described as a filtered acquisition-and-synthesis pipeline with preserved citation lineage — not a knowledge-production loop.

Acceptance criteria, intended use, scope, and authority — the three kept separate. Epistemic authority: the only route issuing anything like a license is RTE-E5, and its honest ceiling is "not contradicted by the sources this run retrieved, as judged by a model vote of undocumented rule, at run time". Nothing else in the subsystem issues epistemic authority; retention (RTE-S5), delivery (RTE-X3), surfacing (RTE-X6), and continuation (RTE-C5) issue none. Operational authority: decided once, pre-execution, by a human reading a phase summary (RTE-C2), with full script text opt-in only; removable by permission mode, by a persistent consent record (RTE-S3), or entirely under Auto+ultracode. Thereafter agents act in forced acceptEdits with the inherited allowlist (RTE-S4, CLM-8), bounded only by the ≤16/≤1000 caps. The asymmetry is the finding: action is unconditional and mid-run, while checking exists only inside one bundled research workflow and applies only to report content, never to the agent output that drives action. Behavioral authority (consumer | channel | force | horizon): enforcing and durable for the procedure — BAP-1 (script is control flow, one run), BAP-2 (saved command, persistent, repo-shared), BAP-9 (archived script, relaunchable), BAP-10 (caps, every run), BAP-7 (org-wide prohibition). Advisory and ephemeral for the findings — BAP-4 (report into the session, horizon session/turn). Permissive and durable for authority reduction — BAP-5 (consent record, persistent per workflow+project or per user), BAP-6 (forced acceptEdits, run), BAP-8 (ultracode directive default, session). Force and horizon both run in favour of method over findings.

Direct behavior/policy adaptation without a truth-apt route. Three routes change behavior on human or model judgement with no truth-apt object and no recorded result: RTE-S3 (one approval becomes standing prompt suppression), RTE-C1 (a session setting makes workflow planning the default), and RTE-C5 (mid-run stop/restart from telemetry). The human evaluation with the largest durable footprint is RTE-S2 — saving a script on the stated criterion "if the run does what you wanted" — and its target is the procedure, not any claim. Bundle satisfaction licenses nothing about any component of the orchestration.

Claims that remain unsupported for want of implementation, run, or causal evidence. CLM-1 is unsupported in its comparative core and, within this boundary, untested: no baseline, no trustworthiness metric, no contrast, and no named shipped workflow implementing RTE-E7/RTE-E8 (MG-1, MG-2, MG-3, MG-9). It is licensed only as a stated design rationale. CLM-2 is supported as doctrine about a shipped pipeline and licenses the existence of a declared check-and-disposition architecture with citation lineage; it licenses nothing about the vote rule, about whether filtering tracks truth rather than retrieved consensus, about acceptance scope, or about any run's behavior. CLM-3..CLM-8 are doctrine-consistent and bound the above; none is implementation- or run-supported. All findings are pinned to v2.1.154–160-era documentation; MG-8 prevents any statement about current behavior.

Returns to the orchestrator (step 7.4)

Material objects by canonical ID. Truth-apt candidates: OBJ-9 (primary), OBJ-6, OBJ-4, OBJ-11, OBJ-10 (acquired), OBJ-5 (measurement). Non-truth-apt but material to authority: OBJ-1, OBJ-7, OBJ-8, OBJ-12, OBJ-13, OBJ-14, OBJ-2; OBJ-3 is a heterogeneous container dispositioned through its contents.

Material routes by canonical ID, with transformation class and route function.

Route Route function Content/update relation
RTE-E1 content transformation acquisition/import (warrant unknown)
RTE-E2 content transformation acquisition/import (warrant unknown)
RTE-E3/extract content transformation indeterminate (reshaping vs. ampliative conjecture)
RTE-E3/check check/evidence production no content change
RTE-E4 check/evidence production no content change
RTE-E5 disposition/acceptance no content change
RTE-E6/synth content transformation indeterminate
RTE-E6/lineage lineage/freshness/recovery no content change
RTE-E7 check/evidence production no content change (capability only)
RTE-E8 check/evidence production no content change (capability only)
RTE-X1 operational admission/selection/consumption indeterminate
RTE-X2 retention no content change
RTE-S5 retention no content change
RTE-C6 lineage/freshness/recovery no content change
RTE-X3 operational admission/selection/consumption no content change
RTE-X6 other — evidence surfacing, no recorded evaluation no content change
RTE-S1 retention non-truth-apt policy/content update
RTE-S2 retention + linked operational admission non-truth-apt policy/content update
RTE-C2 operational admission/selection/consumption no content change
RTE-S4 operational admission/selection/consumption no content change
RTE-S3 behavior/policy adaptation non-truth-apt policy/content update
RTE-C1 behavior/policy adaptation non-truth-apt policy/content update
RTE-C5 behavior/policy adaptation non-truth-apt policy/content update

Architectural status vs. observed candidate state. All routes above are doctrine only except RTE-E3/extract (no route found within boundary) and the four evidenced absences in §3.5. No route is implemented or above (tier fixed). Every observed candidate state is no instance observed (MG-2).

Checking / acceptance / retention-integration, kept separate. Checking: RTE-E3/check, RTE-E4 only, inside CMP-11 only, in the domain of retrieved-source agreement. Acceptance: RTE-E5 only, and it is admission by non-rejection with no criterion, use, scope, or artifact — no acceptance transition in the procedure's sense exists in this subsystem. Retention: RTE-X2, RTE-S5, RTE-S1, RTE-S2 — all pre-acceptance, none integration. Lifecycle integration: not reached by doctrine, no instance observed by evidence.

Three authorities, kept separate. Epistemic: RTE-E5 only, ceiling stated in §6. Operational: RTE-C2, RTE-S3, RTE-S4, BAP-10, BAP-7. Behavioral (consumer | channel | force | horizon): BAP-1..BAP-10 as annotated per route in §3; force and horizon both favour method over findings.

Missing evidence paired with prevented conclusions: MG-1..MG-9 in output 1, cited per route in output 3 and per phase in output 4.

PROPOSED NEW RECORDS (unminted; for the orchestrator to register or reject):

  1. PROPOSED NEW RECORD: route — claim individuation/extraction from fetched sources (OBJ-10 → OBJ-9). Presupposed by CLM-2's quantification over "each claim" and by cross-checking, but never declared as a step in SRC-1; its absence blocks classifying OBJ-9's transformation and blocks tracing source warrant into it. Provisionally annotated in this lens as the linked row RTE-E3/extract. — evidence anchor: SRC-1#bundled, SRC-1#run-a-bundled-workflow.
  2. PROPOSED NEW RECORD: object — per-claim citation binding (the provenance pair linking an OBJ-9 claim to the OBJ-10 source it came from), carried in OBJ-6. Distinct from OBJ-9 (proposition) and OBJ-11 (disposition value) by form (provenance edge), producer (RTE-E6), and consumer (the human reader, not the script). It is the subsystem's only reader-side warrant-transfer mechanism. — evidence anchor: SRC-1#run-a-bundled-workflow ("It cites the sources each claim came from").
  3. PROPOSED NEW RECORD: claim — lineage claim: the delivered report "cites the sources each claim came from". Distinct from CLM-2, which asserts the vote-and-filter operation; this asserts a provenance property of the delivered artifact and is the only claim in SRC-1 that bears on reader-side re-derivation. — evidence anchor: SRC-1#run-a-bundled-workflow.