Analysis result — Claude Code dynamic workflows

Run / result ID: AAS-20260820-CCDW-01

Produced by the candidate instruction analyse-agentic-system. This file is the entry point to the result; the eleven required logical records are distributed across the files indexed below and are numbered here in the instruction's order.

Logical-record index

# Required logical record Where
1 Run / staging identity this file §1; evidence-packet.md §1
2 System boundary, revision, overall evidence tier evidence-packet.md §2, §3, §4, §5
3 Source register evidence-packet.md §4
4 Shared component / object / route / claim / authority records evidence-packet.md §6, extended by this file §8.1
5 Runtime account evidence-packet.md §7
6 Both lens applicability records lens-dispositions.md
7 Applicable lens outputs lens-memory-context.md, lens-epistemic.md
8 Cross-lens reconciliation this file §8
9 Bounded synthesis this file §9
10 Limitations, each paired with the conclusion it prevents this file §10
11 Verification / blocker report this file §11

Supporting artifact (not a logical record): capture-live-session-tool-roster.md, the SRC-2 capture. Trial apparatus (outside the result): trial-notes.md.

§1. Run and staging identity

Field Value
Run / result ID AAS-20260820-CCDW-01
System Claude Code dynamic workflows — the Workflow orchestration facility inside the Claude Code harness
Staging identity kb/work/analyse-agentic-system/trials/cc-dynamic-workflows/
Publication target none authorized — result retained under the staging identity; publication blocker PB-1 in §11
Boundary subsystem-only (evidence-packet.md §3)
Revision / capture SRC-1 captured 2026-06-03; analysis cutoff 2026-06-03
Overall evidence tier doc-grounded
Run date 2026-08-20

All IDs in this result resolve across all physical parts: SRC-1–SRC-2, CMP-1–CMP-12, OBJ-1–OBJ-16, RTE-* (control C1–C9, context X1–X6, state S1–S7, deep-research E1–E9), BAP-1–BAP-10, CLM-1–CLM-10, ABS-1–ABS-6, MG-1–MG-9, L-1–L-14, PB-1.

§8. Cross-lens reconciliation

8.1 Registration of proposed new records (step 7.4 / step 3 "one canonical ID")

Both lenses returned proposed records unminted, as required. The orchestrator registers them here. These are the only new canonical IDs in the run; they extend the packet's §6 tables and do not replace or rename anything.

New ID Kind Record Proposed by Status Evidence anchor
RTE-S7 state route (candidate) Possible write-back from a mid-run tool-permission prompt (RTE-C9) into the user's tool allowlist OBJ-12. Registered explicitly as uninspected so that OBJ-12's retained-state classification stays visibly provisional. memory uninspected SRC-1#approve, #limits
RTE-X5/structure linked row on RTE-X5 Main-session Claude's conversion of user-named input into the structured args value — a model-performed semantic transformation on the invocation path of every saved workflow, with no preservation check. memory claimed; transformation class indeterminate SRC-1#args
RTE-E9 deep-research route Claim individuation / extraction from fetched sources (OBJ-10 → OBJ-9). Presupposed by CLM-2's quantification over "each claim" and by cross-checking, never declared in SRC-1. Architectural status no route found within boundary. Supersedes the epistemic lens's provisional local label RTE-E3/extract, which is retained in lens-epistemic.md as a synonym. epistemic no route found within boundary SRC-1#bundled, #run-a-bundled-workflow
OBJ-15 operative object The / autocomplete command index derived over CMP-7 plus bundled commands — the discovery surface for the saved-command read-back path, distinct from the registry files; maintenance semantics undocumented. memory claimed (existence); uninspected (maintenance) SRC-1#save, #bundled
OBJ-16 operative object Per-claim citation binding — the provenance pair linking an OBJ-9 claim to the OBJ-10 source it came from, carried inside OBJ-6. Split from OBJ-9 (proposition) and OBJ-11 (disposition value) by form (provenance edge), producer (RTE-E6/lineage), and consumer (the human reader, not the script). epistemic claimed SRC-1#run-a-bundled-workflow
CLM-9 claim Unconditional retention claim: "Every run writes its script to a file under your session's directory in ~/.claude/projects/." Load-bearing premise of the archived-script read-back path and of BAP-9. memory doctrine/design SRC-1#how-it-runs
CLM-10 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. epistemic doctrine/design SRC-1#run-a-bundled-workflow

Evidenced absences promoted to records. Both lenses produced documented absences — findings stronger than "we did not look". The packet had no namespace for these; the orchestrator mints ABS-* (flagged as an improvisation in trial-notes.md, F9). Each names its recorded search boundary, as the absent status requires.

ID Evidenced absence Recorded search boundary Proposed by
ABS-1 No cross-session or cross-run store of workflow results/findings. Doctrine positively asserts a fresh start: "If you exit Claude Code while a workflow is running, the next session starts the workflow fresh." SRC-1#resume, #how-it-runs, #save both
ABS-2 No documented acceptance of OBJ-6 — no evaluator, criterion, intended use, or scope attaches to the delivered report. SRC-1#run-a-bundled-workflow, #when-to-use, #how-it-runs, #save, #bundled epistemic
ABS-3 No documented retention or export of OBJ-6. SRC-1#how-it-runs, #save, #resume epistemic
ABS-4 No documented check on OBJ-4 in the general (non-/deep-research) case. SRC-1#when-to-use, #how-it-runs, #limits, #watch-the-run epistemic
ABS-5 No documented retention of rejected OBJ-9 claims or of OBJ-11 vote results. SRC-1#bundled, #run-a-bundled-workflow epistemic
ABS-6 No documented comparison of a workflow's output against a single-pass baseline — the absence that blocks CLM-1. whole file epistemic

ABS-* records are scoped to documented routes in SRC-1. None asserts that no such mechanism exists in the implementation.

8.2 Duplicate merge and namespace hygiene

  • No duplicate objects or routes were created. Both lenses annotated the packet's existing CMP-*/OBJ-*/RTE-*/BAP-*/CLM-* records by ID, as instructed.
  • Linked-row suffixes (RTE-E3/check, RTE-E4, RTE-E5, RTE-E6/synth, RTE-E6/lineage) are the authorized split of one base route into several functional rows, not a parallel namespace. They resolve to their base IDs.
  • RTE-E3/extract → RTE-E9. The epistemic lens correctly refused to mint an ID for a route it found undeclared, and used a provisional linked label. The orchestrator has now given it a canonical ID with status no route found within boundary. This is the one place where a lens's local label needed promotion rather than mere mapping.
  • RB-1/RB-2/RB-3 (memory lens) are path labels, not new records. They map to existing IDs and are not carried into this result as canonical IDs:
  • RB-1 = RTE-S2 → CMP-7 → RTE-C7, authority BAP-2, material OBJ-7
  • RB-2 = RTE-S1 → CMP-8 → later read via RTE-X4, authority BAP-9 → BAP-1, material OBJ-1
  • RB-3 = RTE-S3 → OBJ-8 → CMP-5, authority BAP-5
  • MG-1..MG-9 (epistemic lens) are missing-evidence records, kept as-is and folded into §10.

8.3 Ownership held

Layer Owner Verified
Complete control and context routes runtime baseline (evidence-packet.md §7) Yes — both lenses annotated RTE-X1..RTE-X6 and the RTE-C*/RTE-S* routes; neither redefined an endpoint or a progression.
Read-back and activation memory lens Yes — the epistemic lens made no read-back or activation claim; it treated RTE-S5/RTE-C6 as retention and recovery, which is its own vocabulary, not a read-back finding.
Transformation, checking, warrant, acceptance, integration, and the two authorities epistemic lens Yes — the memory lens explicitly declined to infer warrant from curation, endorsement from registry membership, or semantic preservation from lineage, and explicitly declined CLM-1.

Boundary, revision, and sources are consistent across every record: one boundary (subsystem-only, evidence-packet.md §3), one capture (SRC-1, 2026-06-03), one cutoff, one register. Both lenses cite SRC-1 for every positive finding and cite SRC-2 for none.

8.4 Conflicts, preserved as conflicts

C-1 — OBJ-6 persistence: uninspected vs absent. The memory lens recorded "no persistence of the report to a store is documented" as uninspected; the epistemic lens recorded it as an evidenced absence with a named search (ABS-3). These are different statuses over the same fact. Resolved on the vocabulary, not by picking the stronger-sounding one: the run's definitions make absent mean "not found within the named, recorded search boundary" and uninspected mean "the evidence was unavailable or not inspected". The epistemic lens named its search; the memory lens did not. The canonical status is therefore absent (ABS-3), and the memory lens's weaker reading is recorded as under-specified rather than wrong. Note that absent here is still scoped to SRC-1 and remains compatible with an undocumented implementation route.

C-2 — CLM-3 ("Claude's context holds only the final answer") is doctrine-internally imprecise. The epistemic lens flagged the exactness of "only" as unknown. The memory lens's RTE-X4 annotation supplies the reason: the archived script's path is also pushed into CMP-12 at run start. The conflict is inside the source, not between the lenses, and it is preserved as such: CLM-3's substantive point (agent results do not land in the context window) stands; its literal universal quantifier does not.

C-3 — OBJ-12 allowlist classification is open. The memory lens classified the allowlist as static configuration provisionally, conditional on whether mid-run prompts write back. The epistemic lens treated OBJ-12 as a non-truth-apt authorization consumed by RTE-S4. These are compatible, but the open question is real and is now registered as RTE-S7 (uninspected). Not resolved. If RTE-S7 turns out to exist, OBJ-12 becomes a fourth accumulation-through-use read-back path with permissive force over tool authority — which would be the most consequential single change to this analysis.

C-4 — Version divergence, carried from the register. SRC-1 is pinned to 2026-06-03 / v2.1.154–v2.1.160; the host observed under SRC-2 runs v2.1.237. Preserved unresolved (MG-8, L-2); the cutoff stays at SRC-1's date.

8.5 Cross-lens agreements that neither lens could reach alone

These are reconciliation products, not lens outputs.

  1. The subsystem retains method and authority, never findings. The memory lens found that every durable retention holds a procedure (OBJ-1/OBJ-7) or an authorization (OBJ-8). The epistemic lens independently found no acceptance transition, no retention or export of OBJ-6 (ABS-3), and lifecycle integration not reached by doctrine. Together these give a single structural finding with two independent derivations — and the doc's own CLM-7 ("what's repeatable: the orchestration itself") states the design intent that matches it.
  2. Force and horizon run in the same direction as retention. BAP-1/BAP-2/BAP-9/ BAP-10 (procedure) are enforcing and durable; BAP-4 (findings) is advisory with a session/turn horizon. The memory lens established the horizons; the epistemic lens established the force asymmetry. Neither alone shows that the two axes align.
  3. The one curation act is ungated in both vocabularies. RTE-S2 is the memory lens's sole curation act, and the epistemic lens finds its stated criterion is "if the run does what you wanted" — a human bundle-satisfaction judgement whose target is the procedure. Per step 8.3, the curation label does not determine any epistemic transformation, and neither lens inferred warrant from it. Bundle satisfaction licenses nothing about any component of the orchestration.
  4. Checking and action are asymmetric, and the memory account explains why the gap persists. Checking exists only inside CMP-11 and applies only to report content (RTE-E3/check, RTE-E4); agent output that drives action has no documented check (ABS-4), and agents act unconditionally in forced acceptEdits (RTE-S4, CLM-8). The memory lens adds that OBJ-4 is retained, replayed on resume without revalidation, and reused as another agent's premise (RTE-X1) — so unchecked output propagates forward through the retention layer as well as outward through the action layer.

Two rules were checked and held. Memory curation labels did not determine epistemic transformation anywhere in this reconciliation. Behavioral influence was not allowed to imply epistemic or operational authority — most visibly at BAP-4, where the report reaching a context window is recorded as advisory delivery and explicitly not as activation, not as acceptance, and not as warrant.

§9. Bounded synthesis

Organized around the deployed system's progression. Scope: the dynamic-workflows subsystem of Claude Code, at the 2026-06-03 documentation boundary, on doc-grounded evidence. This is a subsystem-only boundary; nothing here supports a conclusion about Claude Code as a whole. No system-wide epistemic grade is given, and none is available at this evidence boundary. Every statement below is a doctrine-layer statement about a route unless marked otherwise.

Scheduling — the plan moves into code, and that is the whole design

The defining move is that the orchestration plan is written down as an executable artifact. In the turn-by-turn alternatives the doc compares against, a model decides what runs next at each step; here a JavaScript script decides, and the runtime executes it in an isolated background process (RTE-C3, RTE-C4, BAP-1). Three consequences follow directly, and all three are visible in the runtime account rather than inferred:

  • Scheduling becomes inspectable and repeatable. The script can be read before it runs (RTE-C2), read after (RTE-X4), diffed against a previous run's, edited, relaunched, and installed as a command (RTE-S2). None of that is available for a plan that exists only as model intent across turns.
  • Scale is decoupled from context. Fan-out is bounded by the runtime (≤16 concurrent, ≤1000 per run, BAP-10) rather than by what a context window can hold.
  • The model's judgement moves upstream. It is spent once, on authoring (RTE-C1, Loop A), instead of continuously on dispatch. Under /effort ultracode (BAP-8) even the decision whether to use a workflow becomes a standing model decision for the session.

The trade the doc does not name: a script fixes the plan before any result is seen. The only documented adaptations after launch are human interrupts (RTE-C5) and whatever branching the script's author anticipated.

Context assembly — code selects, and nothing tests what the selection did

The script composes each sub-agent's prompt from its own variables (RTE-X1). Selection is code-determined, not model-determined at dispatch, which is the sharpest difference from a conversational orchestrator. Results return into script variables, not into any context window (RTE-X2), and only the final answer reaches the originating session (RTE-X3, CLM-3 — with the literal "only" qualified by C-2). Invocation input passes through a model-performed structuring step on the way in (RTE-X5/structure) that has no preservation check.

The packet's anti-conflation rule bites here and is the load-bearing observation: the runtime retains far more than any model ever sees. Every agent result is tracked (RTE-S5); most of it is never selected into a prompt and never reaches the session. What the human sees by default is one synthesized report; the underlying per-agent prompts, tool calls, and results are reachable only by opt-in drill-down (RTE-X6) that produces no verdict and gates nothing. No route anywhere in the subsystem tests that delivered material actually changed a receiver's behavior — there is no behavioral-faithfulness test on any context route.

State and action — a pure coordinator delegating unconditional action

The script has no direct filesystem or shell access (CLM-6); sub-agents read, write, and run commands (RTE-S4). This is a genuine architectural line, and it is easy to over-read: it is an isolation property of the orchestrator, not a check on action. The agents doing the acting run in forced acceptEdits regardless of the session's own permission mode, with file edits auto-approved and the user's allowlist inherited (CLM-8, BAP-6). Non-allowlisted shell, web, and MCP calls can still prompt mid-run (RTE-C9) — the only thing besides an explicit pause that can stop a run, since there is otherwise no mid-run user input.

Memory return — three read-back paths, all carrying procedure or permission

Applicable lens; the trigger was met three times over.

  • Saved workflow command (RTE-S2 → CMP-7 → RTE-C7, BAP-2) — a script produced by one run becomes a /<name> command in future sessions, discoverable through OBJ-15, and when saved to .claude/workflows/ it is shared with everyone who clones the repo. Force is enforcing on invocation; horizon is persistent. This is the strongest read-back path in the subsystem, and no quality check, test, or review gates the save.
  • Archived run script (CLM-9, RTE-S1 → CMP-8, BAP-9) — every run archives its script; the path is pushed to the session at run start and the content is pulled on demand. A previous run's file can be re-read, diffed, edited, and relaunched, at which point advisory material becomes control flow. No index, listing, search, retention bound, or staleness check over the archive is documented.
  • Persistent consent (RTE-S3 → OBJ-8 → CMP-5, BAP-5) — an approval decision persists and suppresses later prompts; in Auto mode persistence is an automatic by-product of any Yes. Force is permissive: it removes a check without authorizing anything new. This is the only read-back path for which the documentation asserts the behavior change itself; the other two have uninspected activation.

Deliberately excluded as retained state rather than read-back: resume-from-cache (RTE-C6/CMP-9) is same-run, same-session, and doctrine positively closes the cross-session case; script variables, the final report into the session, the bundled /deep-research script, and the session-scoped ultracode setting likewise.

Nothing in this subsystem retains a finding. No cross-run result store is documented, and doctrine asserts a fresh start in a new session (ABS-1). Raw traces exist (OBJ-4, OBJ-5) but no distillation step turns them into a retained artifact; RTE-S2 copies the procedure, never the results.

Truth-apt and warrant routes — one real pipeline, no acceptance anywhere

Applicable lens; invoked through the epistemic procedure.

The subsystem's one documented knowledge pipeline is the bundled /deep-research workflow: search fan-out → fetch → cross-check → vote → filter → cited synthesis (RTE-E1–RTE-E6). Read at the doctrine layer, it is a filtered acquisition-and-synthesis pipeline with preserved citation lineage — a real architectural commitment, and more epistemic structure than an orchestration layer needs to have.

What it does not amount to, at this boundary:

  • Acquisition warrant is unknown. External sources (OBJ-10) are imported with no documented authority, recency, independence, or quality criterion — and the value of cross-checking depends on exactly the independence property that is undocumented.
  • A step is missing from the doctrine itself. CLM-2 quantifies over "each claim", but no route producing claims from sources is declared (RTE-E9). That gap blocks classifying the claim object's transformation as faithful reshaping or as model conjecture, and blocks tracing source warrant into it.
  • The test domain is agreement, not truth. Cross-checking and voting assess whether retrieved sources agree, by model evaluators whose rule, quorum, threshold, and tie-handling are all undocumented.
  • Disposition is admission by non-rejection. Survival is named only by outcome. No rule, no intended use, no reliance scope, no persisted acceptance artifact — and rejected claims and vote results are not retained (ABS-5). No acceptance transition in the invoked procedure's sense exists anywhere in this subsystem.
  • Integration is not reached. A surviving claim's terminus is a report delivered advisorily into a context window (BAP-4, horizon session/turn), with no documented persistence (ABS-3) and no acceptance (ABS-2).

The genuinely valuable epistemic feature is OBJ-16, the per-claim citation binding (CLM-10): it is the subsystem's only warrant-transfer mechanism, and it is reader-side. It lets a human re-derive or contest a claim. It issues no system warrant, and it is itself unverified.

Outside /deep-research, no check is documented at all (ABS-4). Agent output is retained, replayed on resume without revalidation, reused as another agent's premise, and acted on — all before and without any check. The asymmetry is the finding: action is unconditional and continuous, while checking exists in exactly one shipped workflow and applies only to report content, never to the agent output that drives action.

The doc's own warrant claim (CLM-1 — adversarial review or multi-angle weighing gives "a more trustworthy result than a single pass") is, within this boundary, untested: there is no baseline, no trustworthiness metric, no comparison, and no experiment anywhere in the source (ABS-6, MG-3). It is further mismatched at the doctrine layer — it attaches to capabilities of the pattern (RTE-E7, RTE-E8) with no named shipped instance, while the one shipped instance implements a different pattern. It is licensed only as a stated design rationale.

Governing controls — one pre-execution decision, then a bounded free run

Authority in this subsystem is decided once, before execution, by a human reading a phase summary (RTE-C2), with the full script text available only on request. That decision is removable three ways: by permission mode, by a persistent consent record (RTE-S3), or entirely under Auto plus ultracode. In claude -p and the Agent SDK there is no prompt at all. After launch the controls are: the runtime's agent caps (BAP-10), mid-run permission prompts for non-allowlisted tools (RTE-C9), human stop/pause/restart from telemetry (RTE-C5), and the org-level kill switch (BAP-7) — which is the only hard prohibition.

Observability is the counterweight and is unusually good for a background system: per-phase counts, tokens, and elapsed time, and per-agent prompt, recent tool calls, and result. But it gates nothing automatically. It informs a human who may intervene, and it feeds the one curation decision (RTE-S2) that has a persistent, potentially repo-shared consequence.

Capability versus deployment, and the evidence floor

Two capability/deployment gaps are load-bearing and are preserved rather than smoothed over: RTE-E7/RTE-E8 are stated as things a workflow can be written to do, with no shipped instance named; and per-stage model routing is likewise a script capability with no documented instance. Separately, every finding in this synthesis is doctrine-layer. Nothing was inspected in implementation, nothing was observed in a run, nothing was tested. That floor is not a footnote to the synthesis — it is the reason no claim above rises past "the vendor declares".

§10. Limitations, each paired with the conclusion it prevents

ID Limitation Conclusion it prevents
L-1 Doc-only evidence. Tier doc-grounded; SRC-1 is vendor documentation (doctrine/design). No implementation inspected, no observed run, no causal experiment. Prevents any status above claimed anywhere in the run; prevents concluding that any documented route actually operates, that any write path writes, or that any read-back path returns material.
L-2 Version divergence, unresolved (MG-8, C-4). SRC-1 pinned 2026-06-03 / v2.1.154–160; host observed at v2.1.237; no changelog in the bundle. Prevents any statement about present-day behavior of this subsystem, including whether the archive gained retention limits or whether consent semantics changed.
L-3 No repository or implementation exists to inspect for this system; it is a closed vendor feature. Prevents ever reaching code-grounded for this target by the route this instruction defines, absent an observed run or a leaked artifact.
L-4 No script text was inspected (MG-1) — /deep-research's source is boundary-excluded and SRC-1 contains no example script. Prevents naming the vote's quorum, threshold, tie or abstention rules; prevents attributing the filter's criterion; prevents any conclusion about how RTE-X1 selects material into a prompt (policy, ordering, size discipline, truncation).
L-5 No observed run (MG-2); SRC-2 is roster-only and negative. Prevents every observed candidate state in the epistemic lens; prevents every activation and causal finding in the memory lens; prevents upgrading CLM-2 past doctrine.
L-6 No baseline, metric, or interventional comparison anywhere in SRC-1 (MG-3, ABS-6). Prevents any support for the comparative core of CLM-1; prevents attributing trustworthiness to adversarial review, to cross-checking, or to agent count.
L-7 The sub-agent primitive's internals are outside the bundle (MG-4; /en/sub-agents not fetched, and fetching it would have widened the frozen boundary). Prevents any conclusion about what a sub-agent loads besides its spawn prompt, whether agent-internal checking exists, or what warrant OBJ-4 carries; specifically prevents claiming RTE-X1 is the only context route into a worker.
L-8 No claim-individuation step is documented (MG-5, RTE-E9). Prevents classifying OBJ-9's transformation as non-ampliative reshaping or ampliative conjecture; prevents saying source warrant is preserved into it.
L-9 OBJ-3 persistence past process exit is uninspected (MG-7). Prevents concluding that intermediate results are strictly ephemeral, and prevents any confident statement about a run's data-at-rest footprint.
L-10 CMP-8 archive lifecycle undocumented — no retention bound, deletion, rotation, or index. Prevents concluding the archive is a usable memory surface over time rather than an unbounded write-only log; prevents estimating whether the re-read path degrades with volume.
L-11 Consent-record invalidation undocumented — nothing states whether a changed script under an existing workflow name re-prompts. Prevents concluding that OBJ-8's scope still matches what the human approved; therefore prevents any safety conclusion about the persistent-consent path.
L-12 Allowlist write-back unresolved (RTE-S7, C-3). Prevents settling whether OBJ-12 is static configuration or a fourth accumulation-through-use read-back path with permissive force over tool authority.
L-13 Agent-restart (r) state semantics undocumented. Prevents any conclusion about within-run retry memory; leaves a possible intra-run retained-state route unregistered.
L-14 Only one bundled workflow is documented, and no non-bundled script was inspected (MG-9). Prevents generalizing any checking or epistemic finding from /deep-research to workflows as a class — including to any workflow a user saves.

Conflicting evidence, published as a limitation: C-1 (uninspected vs absent on OBJ-6 persistence, resolved on the vocabulary), C-2 (CLM-3's "only" contradicted by RTE-X4 inside the source itself), C-3 (open, registered as RTE-S7), C-4 (version divergence, unresolved).

Unresolved applicability: none. Both lens dispositions resolved to applicable and both lenses ran.

§11. Verification and blocker report

11.1 Result identity and location

Field Value
Result identity AAS-20260820-CCDW-01
Location kb/work/analyse-agentic-system/trials/cc-dynamic-workflows/ (staging identity; entry point result.md)
Boundary Claude Code dynamic workflows — subsystem-only
Revision / capture SRC-1, captured 2026-06-03; cutoff 2026-06-03
Evidence tier doc-grounded
Memory/context lens applicable, ran
Epistemic lens applicable, ran (external procedure invoked)

11.2 Step 10.1 verification

Check Result
Source anchors and statuses Pass. Two sources registered with kind, identity, capture, evidence layer, inspected scope, anchors, and access gaps. Every positive finding cites SRC-1 with a section anchor; SRC-2 supports no positive finding and is cited as such.
Unique, resolving IDs Pass. ID inventory in §1. Checked for collisions after minting RTE-S7, RTE-X5/structure, RTE-E9, OBJ-15, OBJ-16, CLM-9, CLM-10, ABS-1–ABS-6: none. Lens-local labels (RB-*, RTE-E3/extract) are mapped in §8.2 and are not canonical.
One boundary and revision across all records Pass. Both lens outputs restate the same boundary and cutoff verbatim from the packet; neither widened either.
Mandatory runtime coverage Pass. evidence-packet.md §7 covers five material loops with the full step-4.2 field list, the three anti-conflation checks, and materiality statements for each conditional surface inspected.
Both lens dispositions present as explicit records Pass. lens-dispositions.md; neither is implied by an absent section, and the non-triggering candidates are recorded explicitly rather than omitted.
All applicable lens outputs present Pass. Both lenses ran and both outputs are on disk; the epistemic output contains all six required blocks of the invoked procedure plus its step-7.4 returns (see the integrity note in 11.4).
Prevented conclusions stated for every non-run Pass, vacuously for lens non-runs (no lens was skipped). Prevented conclusions are nonetheless stated for every negative and uncertain finding: L-1–L-14, MG-1–MG-9, and the recorded search boundary on every ABS-*.
Shared-route ownership respected Pass. §8.3.
No forbidden evidence upgrades Pass. §11.3.

11.3 Step 10.2 distinction checks

Distinction Held?
Retention is not read-back Yes. Resume-from-cache, script variables, the report into the session, the bundled script, and the session effort setting are all explicitly classified as retained state with reasons.
Context presence is not activation Yes. The four-way table records presence, wiring, activation, and causal effect as four separate findings; two of three activation cells are uninspected and all four causal cells are uninspected.
Implementation is not deployment Yes. RTE-E7/RTE-E8 and per-stage model routing are recorded as capabilities with no named shipped instance. Nothing reached implemented at all — this run's floor is claimed/doctrine only.
Observation is not causality Yes, vacuously and explicitly. There is no observation to over-read; SRC-2 is negative and roster-only, and every observed candidate state is no instance observed.
Curation is not warrant Yes. RTE-S2 is the sole curation act; both lenses declined to infer validation, correctness, or endorsement from a saved script's presence in the registry.
Use is not acceptance Yes. The report's delivery into a context window is advisory (BAP-4); the surviving-claim filter is recorded as admission by non-rejection, not acceptance; lifecycle integration is not reached.
Behavioral authority is not epistemic or operational authority Yes. Three authority families kept separate throughout, with consumer/channel/force/horizon populated per BAP-* and never collapsed into a family label.

11.4 Run-integrity note

The epistemic lens worker's process terminated on a session usage limit after writing its output file. The orchestrator verified the file rather than assuming it: it contains all six required output blocks of the invoked procedure in order, the per-object lifecycle dispositions, the claim comparison for CLM-1–CLM-8, the bounded conclusion, and the explicit step-7.4 returns section. No content is missing and no lens work was redone. The worker's own last status line was written before its final tool call completed and is misleading; the artifact on disk is authoritative.

11.5 Step 10.3 deterministic validation

No deterministic validation was run, by rule and not by omission. Step 9 forbids improvising a collection contract or reusing the agent-memory review schema, and no result contract exists for this artifact type. Step 10.3 anticipates exactly this and routes to "applicable generic validation plus the semantic checklist"; the semantic checklist is 11.2–11.3 above. No schema or parser was changed to manufacture a validation path.

11.6 Blockers

Publication blocker PB-1 — no authorized target contract. There is no authorized publication target for this run, and no existing collection contract in reach can represent an eleven-record, multi-lens, ID-reconciled system analysis. Per the step-9 rule, the complete logical result is retained under the run's staging identity (kb/work/analyse-agentic-system/trials/cc-dynamic-workflows/) and the blocker is reported here. No collection contract was improvised and the agent-memory review schema was not reused.

No other blockers. Specifically: no missing logical records (all eleven present and indexed); no ID collisions (11.2); no unsupported material claims — every claim in §9 is either doctrine-attributed or an explicitly scoped negative; no failed applicable validation (11.5).

Publishable limitations (each already scoped in §10): doc-only evidence (L-1), inaccessible components (L-3, L-4, L-7), no observed run (L-5), no causal experiment (L-6), conflicting evidence (C-1 through C-4). Unresolved applicability: none.