SwarmWorld agentic-system analysis

Type: types/agentic-system-analysis-result.md

Run identity

Run state: kb/reports/state/agentic-system-analysis/AAS-2026-09-25-swarmworld-01/run-state.md

Generated review: kb/agentic-systems/reviews/swarmworld.md

Memory analysis report: kb/reports/state/agentic-system-analysis/AAS-2026-09-25-swarmworld-01/memory-report.md

Memory analysis report SHA-256: 2ab83fb3dc8102531224fb06ed9a29efb4c3e3b67ff5c0cbf7175b311ee1be9d

Frozen specialist input SHA-256: 1ddda63318f04a3cd137fd88cd5a54a6a97b480b8dcae73cd6e2915d5d96d342. Fresh specialist and coordinator used only primary sources in the declared boundary. This artifact names intended projections; the run state records publication.

Boundary and evidence

Evidence basis: source code and current architecture doctrine in SwarmWorld at 6af7ae9fa36d98b07b0492cf139658e8af1f6eab, inspected 2026-09-25. No target execution, external study artifact or live provider request was admitted. Public name SwarmWorld; historical Python package biofoundry.

This is an enclosing runtime for bounded simulated research episodes, with a complete artifact, partial loop boundary. Included: deterministic world/material/artifact transitions, macroturn scheduling and model API/schema, private and social memory, executable programs and inheritance, CLI/server/PettingZoo entry paths, trace/replay and evaluation interfaces where they affect authority. The biological mechanism and declarative scenario interface are included; every scenario's physical validity is not.

Excluded: external model/provider internals (prevents hidden-reasoning or exact-weight-fixity claims); separately released 8.4GB dataset and paper experiments (prevents observed or causal learning claims); renderer internals (prevents UI completeness judgments); physical/material surrogate validity (prevents real-world scientific warrant); exhaustive atlas/figure and scenario-package coverage (prevents source-wide empirical or scenario equivalence claims). Python packages, provider service/network and credentials are external execution requirements. The purpose is an independent system characterization, not adoption advice or experimental reproduction.

Source register

Source ID	Kind	Identity/location	Revision or capture	Evidence layer	Inspected scope	Citation anchors	Access gaps and conclusion prevented
SRC-1	Git	https://github.com/lamm-mit/SwarmWorld; access root /home/zby/llm/commonplace/related-systems/lamm-mit--SwarmWorld	6af7ae9fa36d98b07b0492cf139658e8af1f6eab	Implementation: runtime and configuration; doctrine/design: README and current architecture. No observed run or causal experiment.	Full material paths and selected entry/evaluation ranges named on records; src/biofoundry/ runtime, named profile, README/current architecture only	Pinned simulator, architecture; local full-path anchors below	External service internals, study executions, all scenarios and physical validity uninspected

Evidence reads used commit-addressed Git only. Worktree/HEAD and historical audits are not evidence. All source-dependent local anchors below resolve within SRC-1. A search miss is not used as a global absence.

Shared records

Components

CMP-1 — Authoritative BioFoundrySimulation, material/world/artifact arrays and ArtifactVM. Implementation conclusion status: wired. Symbolic Python/NumPy transitions own preconditions, numeric outcomes, measurements and persistent simulated artifacts. Sequential step is not a proven all-or-nothing transaction. Source: SRC-1 src/biofoundry/simulation.py:298-770,1275-1320,1526-1587; src/biofoundry/programs.py:14-222; src/biofoundry/artifacts.py:25-64,173-260.

CMP-2 — LLMPolicy scheduler/context/queue/schema. Implementation conclusion status: wired. Symbolic orchestration invokes a model with natural-language instructions, validates output and queues a bounded plan. Source: SRC-1 src/biofoundry/policies/llm.py:236-598; src/biofoundry/structured_output.py:738-785.

CMP-3 — OpenAI-compatible Responses wrapper and external distributed-parametric model. Interface conclusion status: wired. Model identity resolution conclusion status: wired: named profile uses gpt-5.6-luna, a provider identifier rather than immutable weights. Exact model fixity conclusion status: uninspected. Parameter changes during operation conclusion status: uninspected inside the provider; the local request path sends context/schema/sampling parameters, not a weight update. No inference about hidden formulation or criticism follows from this opacity. Source: SRC-1 src/biofoundry/providers/openai_compatible.py:143-273; configs/openai-gpt-5.6-luna-technology-ecology-50.yaml:31-56.

        "store": False,

--- src/biofoundry/providers/openai_compatible.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

CMP-4 — AgentMemory, EmpiricalKnowledge, CulturalArchive, ResearchMission and ProgramLibrary. Implementation conclusion status: wired. In-memory symbolic collections hold natural-language evidence/claims, numeric records, identities and code. Lexical/numeric selectors do not introduce an embedding or parametric judge in the inspected retrieval route. Source: SRC-1 src/biofoundry/memory.py:13-279; src/biofoundry/knowledge.py:18-236; src/biofoundry/program_library.py:53-295.

CMP-5 — CLI/server/PettingZoo plus trace/replay/evaluation interfaces. Implementation conclusion status: wired. These invoke the simulation, record output or supply external policies; they do not replace scientific transition rules. Source: SRC-1 src/biofoundry/cli.py:127-148,355-479,1490-1575; src/biofoundry/server.py:247-380; src/biofoundry/env.py:94-180; src/biofoundry/technology_ecology.py:273-370,435-503.

Operative objects

OBJ-1 — Superseded combined seed; retained evidence and original referent follow. Operative parts are OBJ-9, OBJ-10, OBJ-11. Private MemoryRecord collections and action feedback. Implementation conclusion status: wired. SRC-1 src/biofoundry/memory.py:13-233, src/biofoundry/simulation.py:1022-1063, src/biofoundry/policies/llm.py:342-598. Python deques retain record ID, tick, kind, content, salience and causal-parent IDs. Working memory is clamped to 4–16 entries, episodic memory uses configured capacity, notebook uses at least 32. Under capacity 64, these are 16/64/64; notebook is bounded, not permanent. Records include observations, experiment results, prose action outcomes, messages, teachings, proposals and JSON research state. Action-feedback deques separately retain structured recent outcomes and optional intent. The model consumes these as evidence/self-guidance; existence checks do not verify prose content.

    self.notebook: deque[MemoryRecord] = deque(maxlen=max(capacity, 32))

--- src/biofoundry/memory.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

OBJ-2 — EmpiricalKnowledge sparse local evidence. Implementation conclusion status: wired. SRC-1 src/biofoundry/knowledge.py:18-236, src/biofoundry/simulation.py:1837-1862. Cell-keyed records keep the latest local observation; extraction attempts/successes accumulate by resource. The rendered view aggregates observations and selects nearest remembered sites, with last-seen ticks/mass. It supplies grounded material evidence to the model and contributes to simulator recipe admissibility. Shared depot mass is deliberately excluded from the personal semantic evidence view. Missing entries mean unobserved, not nonexistent.

    self.cells[y * self.width + x] = CellEvidence(

--- src/biofoundry/knowledge.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

        "scope_note": "Unlisted materials are unobserved, not proven absent.",

--- src/biofoundry/knowledge.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

OBJ-3 — Superseded combined seed; retained evidence and original referent follow. Operative parts are OBJ-12, OBJ-13. Public archive, spatial deposits, messages and request/fulfillment records. Implementation conclusion status: wired. SRC-1 src/biofoundry/memory.py:236-279, src/biofoundry/simulation.py:510-560,770-885,1424-1500,1589-1642,1988-2008. Archive contents append by unique ID. Spatial records include author/location and whether published. Communications retain sender, recipient and reply relations; taught records preserve original content/IDs, while free-message teaching can create a fresh authored record. Request fulfillment sends the original requester a durable receipt plus locally resolvable produced records. No semantic proof that the arbitrary successful action actually satisfies the request is implemented in the inspected fulfillment route.

    """Attach a successful arbitrary action to an agent-selected request edge."""

--- src/biofoundry/simulation.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

OBJ-4 — Superseded combined seed; retained evidence and original referent follow. Operative parts are OBJ-14, OBJ-15. Recipe/experiment memory. Implementation conclusion status: wired. SRC-1 src/biofoundry/science.py:402-491, src/biofoundry/simulation.py:1170-1274,1365-1399,2240-2272. ResearchMission retains proposed recipes, staged batches and test records; private views select records for the named agent. Pending batches reveal recipe/identity but not latent measured properties. TEST stores properties and material utility; pass/fail target labels are removed from model-facing records. With experience reuse enabled, measured results retain their input recipe, closing a concrete input–measurement–later-plan route. Without it, results retain recipe IDs but omit that direct recipe payload.

    if self.config.experience_reuse:
        # This is the agent-authored input that produced the measured coupon,
        # not evaluator state. Retaining it closes the experience-reuse loop.
        result["source_recipe"] = json.loads(json.dumps(batch.source_recipe))

--- src/biofoundry/science.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

OBJ-5 — Superseded combined seed; retained evidence and original referent follow. Operative parts are OBJ-16, OBJ-17, OBJ-18. Executable programs, authored artifact claims and program skill records. Implementation conclusion status: wired. SRC-1 src/biofoundry/programs.py:65-222, src/biofoundry/program_library.py:53-295, src/biofoundry/simulation.py:1064-1125,1643-1716. DSL instruction content determines program identity; names, authors and claimed ancestry do not. Program entries retain authors/installations and explicit instruction diffs for forks. Per-agent skill indexes retain sources, up to 16 recent evidence IDs and eight recent measurements, lifetime counts and verification/inheritance status. Model-facing compact skill views retain instructions and a bounded evidence/measurement subset. Installed instructions operate through the VM; library possession only establishes knowledge and fork eligibility. Authored artifact purpose is not inferred to be physically realized.

    payload = json.dumps(
        [dict(item) for item in self.instructions],

--- src/biofoundry/programs.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

OBJ-6 — Superseded combined seed; retained evidence and original referent follow. Operative parts are OBJ-19, OBJ-20. Queued AgentAction plans. Implementation conclusion status: wired. SRC-1 src/biofoundry/policies/llm.py:236-341,489-552. Validated output replaces the per-agent deque; later ticks pop one action. Queues are short-lived operational instructions, not evidence of useful learning by themselves. A configured transient-provider failure rolls back private-memory mutations from the entire scheduled planning batch and preserves action queues; failed/uncommitted model traces must not be counted as adopted research-state writes. Configuration/schema objects themselves are static control, outside the memory comparison.

            simulation.memories[index] = memory_snapshots[agent_id]

--- src/biofoundry/policies/llm.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

OBJ-7 — Recorded prompts, responses, actions, events and snapshots. Implementation conclusion status: wired. SRC-1 src/biofoundry/events.py:39-154, src/biofoundry/counterfactuals.py:58-136, src/biofoundry/cli.py:1490-1575, src/biofoundry/simulation.py:2548-2570. JSONL/gzip is losslessly read through template expansion. Replay reconstructs committed research-state writes before action execution. It does not rerun the LLM or establish provider continuation. Retrieval statistics/last-retrieval metadata are not included in the inspected authoritative-memory serialization or reconstructed by apply_recorded_model_trace; replay equivalence therefore must not be described as complete future LLM-policy equivalence.

OBJ-8 — Proposed access-metadata object: retrieval_stats and last_retrieval. Implementation conclusion status: wired. SRC-1 src/biofoundry/memory.py:30-31,79-158,182-209. These are numeric counters and diagnostics distinct from content records. Every retrieval increments selection counts even when experience attention is disabled. Model citations credit only IDs among the last selection; action outcomes credit previously tracked IDs. Enabled adaptive retrieval consumes these counts. Unlike content deques, the inspected dictionary has no per-entry eviction; turnover/reset replace it with the memory object. Diagnostics describe initial selection before possible prompt trimming, not necessarily delivery.

    used = sorted(selected & cited_record_ids)

--- src/biofoundry/memory.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

    prior_state_record = simulation.memories[index].latest("research_state")

--- src/biofoundry/policies/llm.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

            f"Prior self-authored research state: {prior_state}\n{tail}"

--- src/biofoundry/policies/llm.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

A library entry grants knowledge of a program, never material access or a physics
advantage. ``verified`` means the agent obtained a measured artifact observation;
it is deliberately not an evaluator pass/fail label.

--- src/biofoundry/program_library.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

        "progress": {
            "type": "string",
            "maxLength": 480,
            "description": "What the agent believes it has established so far.",

--- src/biofoundry/structured_output.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

        feedback = 0.5 * citation_rate + 0.5 * (
            causal_success_rate if outcome_attempts else 0.0
        )

--- src/biofoundry/memory.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

        stats["outcome_attempts"] += 1
        stats["successful_outcomes"] += int(success)

--- src/biofoundry/memory.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

experience_attention: bool = False

--- src/biofoundry/config.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

experience_attention: true context_budget_characters: 60000 --- configs/openai-gpt-5.6-luna-technology-ecology-50.yaml @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

OBJ-9 — Private MemoryRecord/deque content. Implementation conclusion status: wired. In-memory natural-language observations, outcomes, intent, proposals and received records have symbolic IDs/ticks/salience/parents. Source and evidence: see superseded OBJ-1, SRC-1 src/biofoundry/memory.py:13-233. Bounded FIFO and turnover give episode persistence; the notebook is not permanent.

OBJ-10 — Self-authored research state. Implementation conclusion status: wired for generation/storage/delivery; meaningful formulation conclusion status: afforded. JSON retains prose goal, hypothesis, progress, checkpoint and collaboration need, plus evidence IDs. Truth-apt hypothesis/progress and non-truth-apt intention fields remain distinct parts. Latest state is read as context and used in retrieval queries. Evidence-ID existence does not certify entailment. Source: SRC-1 src/biofoundry/policies/llm.py:342-598; src/biofoundry/structured_output.py:1-160; quotations retained in the combined operative-object records.

OBJ-11 — Structured action feedback/intent. Implementation conclusion status: wired. Simulator-generated numeric/event summaries and optional model message support later recent-outcome views. Intent is an authored explanation of purpose, not a causal diagnosis. Source: SRC-1 src/biofoundry/simulation.py:1022-1063,2109-2210.

OBJ-12 — Public CulturalArchive and spatial publications. Implementation conclusion status: wired. Authored prose and symbolic provenance persist in memory; location/recency/lexical selectors expose them later. Grounded resource mentions establish source/material provenance, not truth. Source: SRC-1 src/biofoundry/memory.py:236-279; src/biofoundry/simulation.py:1170-1274,1589-1642.

OBJ-13 — Messages, taught records and request/fulfillment receipts. Implementation conclusion status: wired. Prose plus symbolic sender/recipient/reply/record IDs is local or addressed. Copied teachings preserve content; free messages can originate claims. Attaching a successful arbitrary action to a chosen request is not semantic satisfaction proof. Source: SRC-1 src/biofoundry/simulation.py:510-587,770-885,1424-1500; quotation on OBJ-3.

OBJ-14 — Proposed recipes and pending microbatches. Implementation conclusion status: wired. Symbolic recipe inputs/processes and batch IDs represent an executable experiment choice. Pending views expose input and mass, not latent measured properties; no explicit truth proposition follows from a procedure alone. Source: SRC-1 src/biofoundry/science.py:410-467; src/biofoundry/simulation.py:1170-1364.

OBJ-15 — Simulator material and artifact measurements. Implementation conclusion status: wired. Numeric properties, utility/services and identities warrant within-simulator observations under the executed inputs. Experience reuse retains the source recipe. Agent views remove target pass/fail, observer records keep it. Source: SRC-1 src/biofoundry/science.py:435-491; src/biofoundry/simulation.py:1064-1125,1365-1399; quotation on OBJ-4.

OBJ-16 — Artifact specifications, split conceptually into prose claims and numeric geometry. Implementation conclusion status: wired for storage and geometry use. Claimed function/architecture/inspiration/predicted effects are authored natural-language candidates; clipped geometric parameters separately affect physics. Field presence does not test the claim. Source: SRC-1 src/biofoundry/artifacts.py:25-64,231-260; src/biofoundry/simulation.py:1526-1587. Further canonical parts OBJ-24 and OBJ-25 keep these unlike routes separate.

OBJ-17 — ArtifactProgram instructions and content identity. Implementation conclusion status: wired. Symbolic straight-line code executes under VM checks. Content hash excludes name/authorship/claimed parent; instruction diffs and known-parent constraints preserve lineage, not a reason why the program works. Source: SRC-1 src/biofoundry/programs.py:14-222; src/biofoundry/simulation.py:1643-1716; quotation on OBJ-5.

OBJ-18 — Skill evidence/index. Implementation conclusion status: wired. In-memory symbolic program sources, evidence IDs, measurement histories and verification/inheritance flags influence sorting and inheritance. Measurement presence sets verified without performance pass; ordinary view can contain unverified entries. Source: SRC-1 src/biofoundry/program_library.py:53-295; src/biofoundry/simulation.py:1064-1125; quotation on OBJ-8.

OBJ-19 — Queued AgentAction plans. Implementation conclusion status: wired. Symbolic per-agent deques replace the previous plan after committed planning, then pop once per tick or WAIT. Operational state alone is not a learned theory. Source: SRC-1 src/biofoundry/policies/llm.py:245-340.

OBJ-20 — Configuration/scenario capability/schema/prompt controls. Implementation conclusion status: wired. Shipped symbolic configuration and natural-language doctrine govern available actions, contexts and experiment settings; they are excluded from accumulated memory. Raw defaults disable LLM, science, experience attention and the optional skill/turnover mechanisms. The named technology-ecology profile enables model/science/attention/skill reuse but disables economy, so inheritance is not active there. Source: SRC-1 src/biofoundry/config.py:40-196; configs/openai-gpt-5.6-luna-technology-ecology-50.yaml:1-99; src/biofoundry/policies/llm.py:1-157.

OBJ-21 — Model/action trace payloads within OBJ-7. Implementation conclusion status: wired. Symbolic files containing authored prose support operator inspection and reconstruction of committed continuation state/actions. This is importing earlier writes, not a new online synthesis or provider continuation. Source: SRC-1 src/biofoundry/events.py:39-154; src/biofoundry/counterfactuals.py:58-136.

OBJ-22 — Snapshot/state digest within OBJ-7. Implementation conclusion status: wired. Symbolic serialization/digest checks certify only the represented state projection. Retrieval metadata omissions prevent full future-prompt equivalence. Source: SRC-1 src/biofoundry/simulation.py:2548-2570; src/biofoundry/cli.py:1490-1575.

OBJ-23 — Scorecard and frozen-artifact assay results within OBJ-7. Implementation conclusion status: wired. Numeric/boolean outcome records describe material, program, performance, novelty and provenance criteria. Artifact removal and held-out disturbance schedules concern fixed discovery states. Source: SRC-1 src/biofoundry/science.py:530-680; src/biofoundry/technology_ecology.py:24-98,273-370,435-503.

OBJ-24 — Prose artifact claims, the truth-apt part of OBJ-16. Implementation conclusion status: wired for authoring/retention. Claimed function and predicted effects are model claims about an artifact; physical success does not semantically check those words. Source: SRC-1 src/biofoundry/artifacts.py:55-62.

OBJ-25 — Numeric artifact geometry, the non-truth-apt part of OBJ-16. Implementation conclusion status: wired. Clipped numeric fields are behavioral parameters consumed in physical calculations, not inferred geometry truth. Source: SRC-1 src/biofoundry/artifacts.py:29-54,231-260.

OBJ-26 — Research hypothesis/progress claims, the truth-apt part of OBJ-10. Implementation conclusion status: afforded for substantive conjecture; wired for schema/storage/delivery. These fields can state explanations and alleged findings, but no candidate instance was admitted. Source: SRC-1 src/biofoundry/policies/llm.py:342-598.

OBJ-27 — Research goal/checkpoint/collaboration need, the policy part of OBJ-10. Implementation conclusion status: wired for generated self-guidance and later query/context use. Goals can shape future selection without constituting a truth proposition. Source: SRC-1 src/biofoundry/policies/llm.py:342-598.

Routes

RTE-1 — Reset and episode horizon. Implementation conclusion status: wired. SRC-1 src/biofoundry/simulation.py:63-117, src/biofoundry/env.py:131-180. Reset creates a fresh archive, private/empirical stores, research mission and program library. Environment truncation uses configured max ticks. No cross-episode retained-learning import is established beyond replay of an existing episode.

    self.program_library = ProgramLibrary(self.population.agent_ids)

--- src/biofoundry/simulation.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

        if self.manual_actions:
            actions.update(self.manual_actions)
            self.manual_actions.clear()

--- src/biofoundry/server.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-2 — Model continuation write/read loop. Implementation conclusion status: wired. SRC-1 src/biofoundry/policies/llm.py:342-598. Scheduled planning obtains the newest research state, uses it plus local observation and pending requests to retrieve memories, assembles the request, validates output and stores a new notebook state. Evidence IDs are filtered to records actually retained, archive IDs or known programs. This validates references, not evidential entailment. The new state can carry goal, hypothesis, progress, next checkpoint and collaboration need. The next call reads these reasons/expectations as prose, with no explicit diagnosis algorithm or independent semantic audit.

        research_state["evidence_ids"] = sorted(retained_evidence)

--- src/biofoundry/policies/llm.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

    transient_provider_failure = bool(scheduled) and any(
        trace.get("error_kind") == "provider_transient"
        for trace in traces.values()
    )
    self.provider_outage = bool(transient_provider_failure and freeze_on_outage)

--- src/biofoundry/policies/llm.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

async def actions(
    self,
    simulation: BioFoundrySimulation,
    show_progress: bool = True,
) -> dict[str, AgentAction]:
    await self.refresh_plans(simulation, show_progress=show_progress)
    return self.next_actions(simulation)

--- src/biofoundry/policies/llm.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-3 — Action consequences to feedback. Implementation conclusion status: wired. SRC-1 src/biofoundry/simulation.py:1022-1063,2109-2210. Simulator outcome updates access statistics and feedback history, then emits a readable action-result memory including intent where supplied. Later semantic observations include recent action results and count/streak summaries. Failure may trigger replanning. This route preserves result detail and optional intention, not a guaranteed causal explanation of failure.

            f"{detail[:400]}{intent_suffix}"

--- src/biofoundry/simulation.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-4 — Observation/experiment to reusable evidence. Implementation conclusion status: wired. SRC-1 src/biofoundry/simulation.py:1064-1125,1365-1399, src/biofoundry/science.py:431-491, src/biofoundry/knowledge.py:37-67,111-236. Passive sensing replaces cell evidence and acquisition records extraction outcomes. INSPECT writes a notebook observation and measurement-backed skill evidence. TEST writes an experiment result; enabled source-recipe retention preserves how the measurement was produced. These records can inform future plans; no independent learning benefit is established.

    if action.artifact == ArtifactType.NONE:
        raise RecipeValidationError("BUILD requires a nonzero artifact type")
    if self.research is not None and action.recipe is None:
        raise RecipeValidationError(
            "BUILD requires an explicit tested recipe in collective-science mode"
        )
    if self.research is not None and action.artifact_spec is None:
        raise RecipeValidationError(
            "BUILD requires an agent-authored artifact_spec in collective-science mode"
        )
    recipe = action.recipe or DEFAULT_RECIPES[action.artifact]
    agent_id = self.population.agent_ids[index]
    batch = self._execute_recipe_from_pool(index, recipe, action.causal_parents)
    program = (
        ArtifactProgram.from_dict(action.program, author=agent_id)
        if action.program is not None

--- src/biofoundry/simulation.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

    for resource, mass in required.items():
        personal = min(float(inventory[int(resource)]), mass)
        inventory[int(resource)] -= np.float32(personal)
        depot_amount = mass - personal
        if depot_amount > 1e-7 and self.config.science.shared_depot:
            contributors.extend(self.research.consume_depot(resource, depot_amount))
        elif depot_amount > 1e-7:
            raise RecipeValidationError("personal inventory is insufficient")

--- src/biofoundry/simulation.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-5 — Social record transfer and constrained combination. Implementation conclusion status: wired. SRC-1 src/biofoundry/simulation.py:770-885,1170-1274,1424-1500,1589-1642. Teach resolves explicitly cited private/archive records and known programs and copies them to nearby/addressed recipients. Publish adds authored content with structured grounding provenance; combined proposals require published records, structured provenance, enough distinct grounded authors and grounded recipe inputs. Provenance gates establish allowed evidence/material paths; they do not test truth of the publication's prose or whether a model derived its design from those words.

        record = self.memories[index].get(parent) or self.archive.by_id.get(parent)

--- src/biofoundry/simulation.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-6 — Program observation, verification, transfer and optional inheritance. Implementation conclusion status: wired. SRC-1 src/biofoundry/program_library.py:110-295, src/biofoundry/simulation.py:132-157,963-1021,1064-1125,1643-1716. Local visibility records unverified program familiarity; measured nearby artifacts mark verified evidence; teach transfers the latest measurement and status. With turnover enabled, replacement clears private/empirical memory and feedback; enabled cultural inheritance selects strongest verified prior skills by measurement count, recency and ID, within a cap. This preserves code/evidence across replacement inside an episode, not weights or a complete personal notebook. Measurements record outcomes, while instruction diffs record edits; neither alone explains why the program works.

        (entry for entry in existing.values() if entry["verified"]),

--- src/biofoundry/program_library.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

        self.memories[index] = AgentMemory(self.config.population.memory_capacity)
        self.empirical_knowledge[index] = EmpiricalKnowledge(self.world.width)
        self.action_feedback[index].clear()

--- src/biofoundry/simulation.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

MAX_INSTRUCTIONS = 64 MAX_REGISTERS = 16 --- src/biofoundry/programs.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

        value = _operand(instruction.get("value", 0.0), registers, sensors)
        if op == "set_open":
            outputs[op] = float(np.clip(value, 0.0, 1.0))
        else:
            outputs[op] = outputs.get(op, 0.0) + float(
                np.clip(value, 0.0, MAX_ACTUATION_PER_TICK)
            )
    return outputs

--- src/biofoundry/programs.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-7 — Context selection and access feedback. Implementation conclusion status: wired. SRC-1 src/biofoundry/memory.py:45-209,245-279, src/biofoundry/policies/llm.py:193-233,363-478, src/biofoundry/simulation.py:1940-2008,2179-2221. Private retrieval scans retained episodic+notebook content, tokenizes it, computes document-frequency weighting and scores the full candidate set; no vector index is wired. Archive retrieval similarly scans publications. Automatic skill focus extracts exact program IDs from research-state text, then favors matching and verified records. Locality/recency limit spatial evidence. Record excerpts bound delivered content without rewriting authoritative content. Context trimming drops selected memories first, then specified observation-list tails and whole optional fields; it can stop above the nominal budget when nothing removable remains. No hard token limit follows from the character target.

            str(entry["program_id"]) not in focus,
            not bool(entry["verified"]),

--- src/biofoundry/program_library.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

            elif not _shrink_observation(observation_packet):
                break

--- src/biofoundry/policies/llm.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-8 — Observer files and deterministic reconstruction. Implementation conclusion status: wired. SRC-1 src/biofoundry/events.py:64-154, src/biofoundry/counterfactuals.py:58-136, src/biofoundry/cli.py:1490-1575. Named CLI replay reads recorded content, checks revision/configuration/scenario compatibility, reconstructs research state and replays actions before comparing state digests. Counterfactual replay removes agents by replacing their recorded actions with passive controls. Neither path replans after altered recall; this is not a recalled-content dependence test. Template deduplication is storage compression, not memory synthesis or learning. The report claims code wiring only.

if record.get("planning_committed") is False:
    return

--- src/biofoundry/counterfactuals.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

        apply_recorded_model_trace(simulation, model_trace)

--- src/biofoundry/counterfactuals.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

for source in sources:
    knockout = copy.deepcopy(assay_source)
    knockout.artifacts.retired[source] = True
    knockout.artifacts.health[source] = np.float32(0.0)
    knockout.artifacts.performance[source] = np.float32(0.0)
    knockout.artifacts.services[source].fill(np.float32(0.0))
    knockout.artifacts.programs[source] = None
    result = _run_agent_free_assay(
        knockout,
        assay_horizon,
        description=f"Knock out artifact {source}",
        show_progress=False,
    )
    target_auc = np.asarray(result["artifact_auc"], dtype=np.float64)
    system_effect = float(intact["resilience_auc"] - result["resilience_auc"])

--- src/biofoundry/technology_ecology.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

    "evaluation_unit": "held-out disturbance schedule",
    "discovery_state_frozen": True,
    "agent_actions_per_schedule": 0,
    "seeds": evaluator_seeds,

--- src/biofoundry/technology_ecology.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-9 — Proposed route for deterministic derived context over retained action/social/empirical state. Implementation conclusion status: wired. SRC-1 src/biofoundry/simulation.py:795-823,1837-1862,2109-2221,2240-2272, src/biofoundry/science.py:794-864, src/biofoundry/knowledge.py:111-229. The model receives aggregate action counts and streaks, closest remembered resource/station sites, public author profiles and pending requests selected by participant/fulfillment state. These are read-time derived views; persistent evidence/counters and receipt records support later reuse. The views themselves are not separately durable learning artifacts. Request fulfillment automatically retains a receipt/result subset for later caller-agent context. No additional model summarizes or judges these projections.

        "passes_material_target": (
            self.material_utility(batch.batch) >= self.config.target_material_utility
        ),

--- src/biofoundry/science.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

RTE-10 — Mission/evaluation measurement and disposition. Implementation conclusion status: wired. The operator invokes scorecard/assay consumers over the simulated state. TEST calculates utility and threshold status; scorecard declares functional outcome from a passing recipe, present authored-spec fields, non-system program history, peak performance and novelty. It can inspect provenance contributions separately. This evaluates concrete outcome criteria, not the truth of every hypothesis/claimed function. The scorecard reports to the operator; no automatic semantic claim acceptance or online promotion from this aggregate was established. Source: SRC-1 src/biofoundry/science.py:469-491,530-680; src/biofoundry/technology_ecology.py:24-98,273-370,435-503.

            and is_invented
            and is_programmed
            and has_performance
            and is_novel
        )
        built_from_test |= is_tested
        causally_composed_artifact |= is_causally_composed
        collaborative_build |= is_collaborative

--- src/biofoundry/science.py @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

The following extensions complete route ownership and admission details. They amend the existing records, not their referents. All route implementation conclusion statuses remain wired; observed activation and benefit remain uninspected because no candidate-linked execution was admitted.

Route	Trigger, next owner and decision policy	Context, executor/effect, return and terminal output	Persistence, later read-back and delegated visibility	Selection, invalidation, recovery and admission limits
See RTE-1	Operator selects config/seed/CLI policy; server loop or environment caller drives ticks. Reset creates stores.	Symbolic CLI policy factory chooses scripted/scalable/research-oracle/LLM. Server can overlay manually supplied actions; environment caller receives observations/rewards/termination/truncation.	Episode state lives in the simulation; recorders optionally write files. Provider sees only assembled request; external caller sees its configured observation.	Max ticks/end of episode/reset bound lifetime. Human configuration controls grants. Manual or external actions bypass LLM output schema but still reach simulator parsing/preconditions. Reset is fresh state, not cross-task memory import.
See RTE-2	Scheduled macroturn or enabled event replan; LLMPolicy limits scheduled IDs by remaining call budget, dispatches parallel jobs and deterministically orders traces.	Prior state, current local evidence and selected memory guide external model; schema/coercion admits action/state fields; commit replaces queues and records state. Immediate return is success flag and traces, not completed physical work.	New state enters notebook and next query/prompt; plans persist until pop/replacement. Provider requests store: false; opaque server persistence is not established.	Any transient failure with freeze enabled restores scheduled agents' memory snapshots and preserves queues; CLI/server retry without world advance. Terminal/invalid output becomes WAIT. Direct actions() ignores the false flag and pops a queue, so it does not itself guarantee caller-level freeze. Call accounting distinguishes attempts from committed calls; no finite outage-retry cap was established in the CLI loop.
See RTE-3	Each tick submits actions; simulator parses, checks available actions/resources/energy and selects within action budget.	Symbolic transitions resolve movement/actions, step world/artifacts, update feedback, then return events/rewards/artifact score. Failed actions can trigger early planning under enabled event scheduling.	Outcome/intent records, access counters, empirical state and artifacts affect later observations/plans. Peers see only communication or local/public observations.	Missing/invalid actions become WAIT; action exceptions become rejection events, not proof of rollback. BUILD consumes materials before later program/artifact checks, so earlier effects can survive a downstream rejection. Global state rollback is not supplied by the inspected catch path.
See RTE-4	Agent proposes/combines recipe, OPERATEs, TESTs, BUILDs or INSPECTs; simulator/material lab owns admissibility and measurement.	Grounding/provenance checks gate proposals; inventory/process checks gate execution. TEST exposes latest untested personal batch; INSPECT measures nearby artifacts. Agent obtains numeric record and optional recipe association, observer obtains pass data.	Recipe IDs, batches, measurements and notebooks survive episode; later model views and skill observations consume them. Unmeasured latent properties stay outside private pending view.	Test criterion is simulator utility; no expected prose answer. BUILD's error text mentions tested recipe, but its inspected call path checks supplied recipe/spec and material availability without a prior TEST lookup. Invalid program after material consumption is not atomic rollback. New experiments or reset replace/extend state; old measurements remain historical.
See RTE-5	Model chooses PUBLISH/DEPOSIT/COMMUNICATE/TEACH/COMBINE; simulator checks local/addressed recipients and identity/provenance constraints.	Source record resolution copies content; free messages/publications can author new claims. Combined proposals require multiple published records, structured grounding and configured distinct contributors/resources. Return is delivery/proposal event, not semantic truth judgment.	Private recipients, archive and spatial deposits enable later reads. Receipts connect successful actions to requests; publication IDs can guide plans and eligibility.	Communication/config/locality gates exposure. Publication grounding is resource-name mention tied to evidence, not a test of prose. Append-by-ID and FIFO are not semantic deduplication or withdrawal; no semantic rollback was established in these functions.
See RTE-6	Model writes/forks/installs a program; observation/teaching and optional turnover update skill records.	Simulator validates known parent, changed content and ancestry; VM validates bounded DSL. Measurements mark verified, teaching preserves status, inheritance selects verified records by counts/recency. Returned events/history name installed program and lineage.	Installed code affects future artifact ticks; skills influence later context and fork eligibility; optional replacement inherits bounded skills while resetting notebook/empirical/feedback state.	No pre-install performance improvement gate. Replacement can install worse behavior; lineage preserves history without automatic rollback selection. Instruction-level bounds are not a universal aggregate actuation bound. Optional inheritance inactive when named profile economy is disabled.
See RTE-7	Scheduled model assembly automatically selects retained experience; scoring consumes state/observation/query plus counts.	Lexical overlap/salience/feedback/exploration and program-ID focus select memory/skills; coarse locality/recency/limits apply. Model receives excerpts; diagnostics expose selected records.	Selected/cited/outcome counts persist and change future ranking when enabled. A cited record may be trimmed before delivery; actual content dependence is uninspected.	Named profile targets six private records, six archive records,1800 chars per memory and60000 total characters. Trimming can stop when no removable field remains; target is best effort, not hard token bound. FIFO/reset/turnover forget content; metadata does not necessarily evict with content.
See RTE-8	Recorder writes events/model/actions/snapshots; operator requests replay/counterfactual/held-out assay.	Replay checks engine/scenario/config conditions, restores committed model states, executes recorded actions and compares serialized state digests. Agent removal replaces recorded actions; artifact knockout uses deep-copied discovery state and common future conditions.	Files preserve inspectable outputs. Read-back goes to replay/operator; no live model replanning in these counterfactual routes. State equality omits adaptive access metadata and cannot certify future prompts.	Engine mismatch can skip action replay while checking snapshot integrity; scenario hash mismatch rejects. A missing field is not restored by a digest. Held-out schedule evaluation freezes discovery state. No empirical effect is claimed without executions.
See RTE-9	Observation assembly projects action/social/empirical stores; successful chosen reply generates receipt.	Deterministic summaries expose counts/streaks, nearest sites, author profiles and pending requests. Receipt stores result IDs; projection returns context immediately.	Durable source/counter/receipt state enables later view; rendered aggregate is usually ephemeral. Recipient identity selects the request subset.	No separate summarizing model or semantic fulfillment judge. Limits/locality/participant state choose records; underlying reset/eviction governs expiry. Receipt persistence qualifies as trace-fed context; pure rendering does not.
See RTE-10	TEST/scorecard/operator assay runs configured numeric criteria.	Material threshold creates pass flag; functional scorecard joins measured input/output and field-presence/program/novelty criteria. Returns operator-facing outcome records.	Measurements/scorecards persist in research state or output files; online private measurement excludes target label. Assay outputs inform human analysis; automated online semantic promotion is uninspected.	Configured thresholds are evaluator definitions, not ground-truth explanations. Optional artifact removal/held-out schedules isolate physical artifact behavior under simulator assumptions, not recalled-content dependence or truth of prose.

Material revision admission is concentrated in RTE-2 (state/plan replacement), RTE-4 (experiment/product construction), RTE-5 (public/social knowledge admission), RTE-6 (executable behavior/skill inheritance) and RTE-7/RTE-9 (access metadata and receipts). Proposals come from the model or deterministic consequence processing; the simulator owns formal/material rejection and the operator chooses static configuration. There is no universal human semantic acceptance stage on these paths. Guidance is the shipped action manual, current local evidence, prior state and retained measurements; each remains either doctrine, input/outcome record, or authored theory candidate. Program/recipe syntax is individually inspectable and editable; code history can preserve changed instructions without preserving an argument for the change. Natural-language rationale may persist in hypothesis, progress, intent or predicted effects, and later context reads it. No route-level claim that criticism occurred follows from its possible storage.

For theory-bearing guidance in OBJ-26, OBJ-12, OBJ-13 and OBJ-24: formulation conclusion status: afforded; delivery/retention conclusion status: wired; operative application of its stated content conclusion status: uninspected; content-directed criticism conclusion status: uninspected; criticism-linked revision or changed reliance conclusion status: uninspected; improved capacity attributable to that process conclusion status: uninspected. The source provides model calls and measurements through which these can happen; opaque model processing and missing candidate traces prevent asserting the links. A program or recipe update is not by itself a formulated explanatory theory or criticism of one. This distinction does not imply that unobserved reasoning is absent.

Claims

CLM-1 — README characterizes SwarmWorld as a deterministic research environment where agent societies discover, test, exchange, inherit and embody technologies. Claim conclusion status: claimed. Source: SRC-1 README.md:1-30. RTE-2, RTE-4, RTE-5 and RTE-6 implement operations fitting that purpose; the commissioned evidence supplies no measured discovery improvement.

SwarmWorld is a deterministic research environment for studying how societies of language-model agents discover, test, exchange, inherit, and physically embody technologies in a shared world. Agents perceive locally, act through a bounded action contract, and leave persistent artifacts whose material effects and executable programs continue on later simulation ticks. --- README.md @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

CLM-2 — Architecture claims a simulator authority boundary, action replay and separation of evidence, authored claims and physical outcomes. Claim conclusion status: claimed. Source: SRC-1 docs/ARCHITECTURE.md:6-52,66-81,86-109. The inspected interfaces support bounded authority; tick atomicity, per-tick actuation wording and future-policy equivalence require the limits stated on RTE-3, RTE-6 and RTE-8.

BioFoundrySimulation is the only component allowed to mutate scientific state. PettingZoo, policies, model servers, FastAPI, and Godot submit actions or consume state; none defines physical outcomes. --- docs/ARCHITECTURE.md @ 6af7ae9fa36d98b07b0492cf139658e8af1f6eab

Evidenced absences

None established as source-wide ABS records. Missing candidate executions, hidden provider reasoning and untested faithfulness are limitations. Specific call-path limits are positive descriptions of inspected code, not claims that no other branch or external study exists.

Behavioral-authority paths

BAP-1 — Consumer: external policy model. Channel: system/manual plus observation, selected memory and prior research-state user input. Force: instructional for shipped doctrine, advisory/evidential for memories and self-guidance. Horizon: one request, renewed by episode retention. Delivery is wired; content activation is uninspected. Source: SRC-1 src/biofoundry/policies/llm.py:342-598.

BAP-2 — Consumer: simulator and ArtifactVM. Channel: AgentAction, recipe, geometry, installed DSL and configuration. Force: schema/precondition validation and deterministic execution/enforcement. Horizon: an action, queued plan, or installed artifact lifetime. This operational authority does not accept an accompanying explanation. Source: SRC-1 src/biofoundry/simulation.py:298-770,1275-1399,1526-1716; src/biofoundry/programs.py:65-222.

BAP-3 — Consumer: agents/selector/fork resolver. Channel: taught/published records, skill evidence, grounding/parent IDs and access metadata. Force: advisory for model evidence, ranking for context, permissive/enforcing for combined-proposal/fork/inheritance eligibility. Horizon: retained episode state until reset/eviction/turnover. Source: SRC-1 src/biofoundry/simulation.py:1170-1274,1424-1500,1643-1716; src/biofoundry/program_library.py:110-295; src/biofoundry/memory.py:79-209.

BAP-4 — Consumer: operator/replay checker. Channel: traces, digests, scorecards and artifact evaluation results. Force: replay consistency rejection and scoped numeric outcome reporting, not semantic acceptance of model prose. Horizon: recorded episode/evaluation projection. Source: SRC-1 src/biofoundry/cli.py:1490-1575; src/biofoundry/technology_ecology.py:273-370,435-503.

Runtime account

The ordinary model episode starts when an operator chooses a seed and configuration, requests LLM policy and supplies provider credentials. The CLI creates simulation/policy and records configuration/engine identity. Agent identities are population slots, not independent service principals. A macroturn scheduler selects active agents by interval/phase and, when configured, event-driven replanning. The policy reads each agent's own observation and retained state, automatically retrieves a bounded subset, calls the external provider with schema and action manual, and validates returned state/actions. Parallel calls finish in arbitrary order but traces are ordered by agent ID before commit. One queued action per agent executes each world tick. The simulator returns consequences; notebook/feedback/measurements then shape the next explicit context. Max ticks or action/call budgets bound the episode's operation; recorder outputs and scorecards are the operator's terminal artifacts. Sources: SRC-1 src/biofoundry/cli.py:355-479; src/biofoundry/policies/llm.py:245-598; src/biofoundry/simulation.py:274-410,624-770.

Alternate interfaces matter to guarantees. CLI scripted/scalable/research-oracle modes bypass model generation. The inspected factory establishes selectable policy routes, but internal oracle-policy privileged knowledge was not inspected; its name is not evidence of an expected-answer oracle. PettingZoo supplies decoded external actions and returns observations, rewards and termination/truncation; its caller's learning loop is external. Server manual actions overwrite planned actions before step. Thus LLM output schema is a model-interface gate, while simulator preconditions are the shared action boundary. No broad host sandbox is inferred: the source-installed Python runtime and operator control plane are trusted. Artifact programs use a restricted DSL, not unrestricted Python/shell. Sources: SRC-1 src/biofoundry/cli.py:127-148; src/biofoundry/env.py:94-180; src/biofoundry/server.py:247-341; src/biofoundry/programs.py:14-222.

The capability surface includes experiment, communication, teaching and program modification verbs; the current grant set depends on configuration and scenario capabilities. Named profile settings enable many science/memory features but disable economy, so mortality/inheritance is only a source-supported alternative. The deployment isolation envelope beyond simulator/VM checks is uninspected. Model responses are proposals, physical calculations are simulator-owned, and public prose is not interpreted as an executable physics amendment.

Four static forcing cases distinguish the boundaries:

1. Transient model failure: refresh_plans freezes on any scheduled transient provider failure with the option enabled, despite its docstring saying total outage. It restores private memory and preserves queues. CLI/server honor the false return and retry without stepping; the convenience actions() wrapper still consumes an action. Freeze is an entry-loop protocol, not a universal invariant across callers. Required external contract: the provider's retryability classification and eventual service recovery. Sources: SRC-1 src/biofoundry/policies/llm.py:245-340; src/biofoundry/cli.py:393-453; src/biofoundry/server.py:283-317.
2. Untested or invalid construction: _build requires a supplied recipe/spec in science mode, executes the recipe, then parses the program/adds the artifact. The inspected helper checks material availability, not prior TEST membership. The mission later computes tested-design success separately. Materials can be consumed before a subsequent validation failure; rejection does not prove atomic rollback. This is a call-path account, not a claim that every build succeeds or lacks all validation. Sources: SRC-1 src/biofoundry/simulation.py:1275-1320,1526-1587; src/biofoundry/science.py:530-620.
3. Bad but measured skill: INSPECT supplies a measurement and sets verified independently of threshold. Ordinary skill view ranks verified first but does not exclude unverified; inheritance does exclude unverified, then orders by measurement count/recency. Therefore verification is evidence presence and selection authority, not successful operation. Fork installation checks parent identity/content difference, not improvement. Sources: SRC-1 src/biofoundry/simulation.py:1064-1125,1643-1716; src/biofoundry/program_library.py:110-295.
4. Replay or stronger actuation claim: replay can recover committed research state/actions and serialized physical state, but not adaptive selector statistics omitted by the serializer/restorer. Matching digests do not entail equal later prompts. Separately, VM bounds instructions/registers and clips each actuator contribution, while repeated contributions accumulate; its per-instruction cap does not imply the same cap for an aggregated output. Further physical constraints remain in artifact/world calculations. Sources: SRC-1 src/biofoundry/counterfactuals.py:58-136; src/biofoundry/simulation.py:2548-2570; src/biofoundry/programs.py:14-16,141-222; src/biofoundry/artifacts.py:173-260.

No dynamic check planned. Provider outage, construction precondition, inheritance filtering and replay-statistics checks were considered; the selected claims are bounded call-path and dataflow conclusions visible in code, so no provider expense or scientific reproduction was warranted. No test was attempted or counted as an observed execution. Scientific usefulness, causal recall and simulator accuracy require different evidence.

Diagnosis is model-proposed from feedback when it occurs; formal rejection is simulator-owned. Candidate comparison is optional model deliberation plus configured outcome calculations, not a separately guaranteed critic. The operator chooses task, thresholds and treatment; a model can replace its plan/state or installed behavior within allowed actions, and the simulator can veto inadmissible syntax/material/identity conditions. No candidate-linked content-directed criticism was observed. In the ordinary LLM route, measurements supply outcomes rather than expected prose answers. The material pass threshold and outcome scorecard are code/configuration-defined references supplied by the experiment designer; they govern stated measurement criteria. They are not answers validating an explanation. Internal privileged inputs of the separately selectable research-oracle policy remain uninspected and are excluded from the ordinary model-episode conclusion.

Lens scoping

Memory/context scope

Depth: full. Trigger evidence: OBJ-1, OBJ-2, OBJ-3, OBJ-4, OBJ-5, OBJ-7 and RTE-2, RTE-3, RTE-4, RTE-5, RTE-6, RTE-7, RTE-8 and RTE-9. Inspected scope: private memory/research state/feedback, empirical knowledge, public/local/social records, experimental input/outcome records, program lineage and skills, adaptive access metadata and reconstructive trace read-back. Canonical parts OBJ-8, OBJ-9, OBJ-10, OBJ-11, OBJ-12, OBJ-13, OBJ-14, OBJ-15, OBJ-16, OBJ-17, OBJ-18, OBJ-19, OBJ-21, OBJ-22 and OBJ-23 preserve differences. Static OBJ-20, provider weights, presentation-only outputs and external studies are excluded. Explicit memory is central to repeated macroturns, so a brief lens would miss material behavior. Both raw/default and named/enabled branches are covered.

Epistemic scope

Depth: full. Trigger evidence: CLM-1, CLM-2; authored OBJ-26, OBJ-12, OBJ-13 and OBJ-24; measured OBJ-2, OBJ-15, OBJ-23; checking/admission on RTE-2, RTE-4, RTE-5, RTE-6 and RTE-10. The invoked standalone epistemic procedure is applied locally as a sparse overlay on these records. The question is which claims become measured, accepted or merely usable, and what each check permits. Assessed: model hypothesis/spec/publication generation, empirical acquisition, experiment measurements, formal/provenance admission, skill selection, traces/replay and scoped outcome evaluation. Unassessed: model-internal reasoning, study executions, every surrogate equation/scenario and full research-oracle policy. Those exclusions prevent a general theory-builder or real-world knowledge-production conclusion.

Lens outputs

Memory/context lens

The fresh specialist's complete report is integrated, with source evidence retained above. Explicit memory is supplied anew to the stateless request interface. Episode-lived stores include bounded personal records, accumulating public/program records and access metadata. Research-state goals/hypotheses contribute to the next query and prompt. That is wired continuity and self-guidance, not evidence that a useful theory changed a choice. The named profile's experience attention differs from unadapted defaults; optional turnover crosses agent generations inside an episode, not independent tasks.

Automatic trace-fed writes comprise RTE-2 research continuation, RTE-3 outcome/intent and access feedback, RTE-4 measurement/input association, RTE-6 measured skill records, RTE-7 adaptive counters and RTE-9 receipts. These routes feed later context, ranking or admissibility, supporting trace_learning: yes at wired strength. Source is event-streams, timing online, scope per-task (the configured episode), and outputs natural-language plus symbolic. Weight changes, raw JSONL recording and pure read-time rendering are not additional learning routes. Stored hypotheses can synthesize claims; curation also evolves entries, promotes measured skills and forgets via FIFO/reset. Exact-ID collision checks, lossless compression and substring grounding do not establish semantic deduplication, consolidation or invalidation.

At a scheduled macroturn, private/archive selection pushes lexical and coarse-selected content; exact program IDs in prior state focus skill views, and participant identities choose requests. TEACH and FORK_PROGRAM explicitly request retained IDs; operator replay requests recorded content. Those are pull operations distinct from later automatic context push. Read-time aggregates expose retained evidence but do not themselves become separate durable theories. Context trimming can exceed its nominal character target; selection diagnostics can include records omitted from delivered context. Citation and successful cited outcomes change counters without proving content activation. Faithfulness remains not-determinable: no recalled-content intervention with replanning and retained outcomes was admitted. Fixed-action contributor replay and artifact knockout ask different causal questions.

Epistemic lens

1. Source-and-claim boundary

System/revision and evidence boundary: see SRC-1 and Boundary and evidence. Question and assessed/unassessed route families: see Epistemic scope. Consequential knowledge/authority claims: CLM-1 and CLM-2, doctrine/design only as claims. Implementation supports the following routes; no candidate-linked trace or execution supplies an observed phase. Provider opacity prevents judging internal inference; missing experiment artifacts prevent measured benefit; excluded surrogate validation prevents external scientific warrant.

2. Epistemic-object inventory

Generic identity, producer/consumer, form, storage and anchors remain on the canonical records. This overlay names only truth-apt parts and their limits.

Object/part	Candidate content and transformation	Warrant boundary
OBJ-26	Hypothesis/progress can state ampliative explanations and alleged findings; exact generated content unobserved	A schema and delivery route do not certify implication or criticism
OBJ-27	Goal/checkpoint/collaboration intention; non-truth-apt policy content	Can steer query/context without a discovery claim
OBJ-12	Publications may copy evidence or conjecture mechanisms; transformation indeterminate without instances	Grounding supplies resource/source provenance, not prose entailment
OBJ-13	Taught source content is copied; free messages may originate claims; receipts assert action/request links	Copied lineage preserved within retained text; semantic satisfaction unknown
OBJ-24	Claimed function/predicted effects can be ampliative propositions about artifact behavior	Field presence is not measurement of these words
OBJ-25	Clipped geometry settings; non-truth-apt design parameters	Operational influence does not establish an explanation
OBJ-14, OBJ-17, OBJ-19	Experiment recipe, executable program and action plan; no necessary truth proposition	Formal validity or selection licenses execution only
OBJ-2, OBJ-15	Sensor/extraction/test evidence; acquisition and deterministic measurement projection	Domain is the configured simulator and executed input; not real materials
OBJ-9, OBJ-11	Readable outcome summaries and retained intent; non-ampliative rendering for simulator outcomes, authored intention otherwise	Abbreviation preserves a selected outcome subset, not all causal explanation
OBJ-18	Program familiarity/measurement status and history	Verified means measured evidence, not evaluator pass
OBJ-8	Numeric access/citation/success statistics	Direct selector adaptation; not truth-apt criticism
OBJ-21, OBJ-22	Recorded model/action content and state identity checks	Records establish intended replay data, not observed truthful claims or full policy state
OBJ-23	Calculated threshold/scorecard/assay outcomes	Scoped numeric outcome derivation, not acceptance of generating theory
OBJ-20	Shipped doctrine/configuration	Defines task/check domains; no runtime conjecture instance

Superseded aggregate OBJ-1, OBJ-3, OBJ-4, OBJ-5 and OBJ-6 are disposed through their parts. OBJ-7, OBJ-10 and OBJ-16 remain explicit containers with unlike part dispositions rather than one epistemic label.

3. Authority-route ledger

Every row's architectural status is separate from candidate observation. implemented below means inspected wiring only. All effects are possible implemented results, not reported executions. Behavioral-authority paths supply consumer/channel/force/horizon; canonical records carry anchors and complete progression.

Route/function	Architectural status	Target and content/update relation	Evaluator/condition, activation and result	Epistemic license	Operational/behavioral force	Claim and gap
RTE-2 / content transformation	implemented	OBJ-26, OBJ-27; possible ampliative conjecture and policy update from explicit context	Scheduled provider output creates state/plan; model reasoning uninspected	Candidate generation only	BAP-1 self-guidance; BAP-2 queue proposal	CLM-1; no instance or known entailment
RTE-2 / check/evidence production	implemented	OBJ-10, OBJ-19; no content change apart from coercion/reference filtering	Local schema and retained-ID checks yield valid output/WAIT	Structural conformity and existing references	BAP-2 blocks invalid output	CLM-2; no semantic support check
RTE-2 / retention	implemented	OBJ-10, OBJ-19; no content change after admission	Committed planning stores state/replaces queue; outage can rollback memory	No additional warrant	BAP-1 later context; BAP-2 actions	CLM-1; storage is not acceptance
RTE-3 / content transformation	implemented	OBJ-11 and OBJ-9; non-ampliative reshaping of outcome plus retained authored intent	Deterministic action-result formatter after each action	Outcome subset under simulator semantics	BAP-1 later evidence, replanning and BAP-3 counts	CLM-2; no guaranteed explanation
RTE-4 / check/evidence production	implemented	OBJ-2, OBJ-15; truth-apt transformation: acquisition/import	Local sensing, TEST and INSPECT produce measurements	Configured simulator-domain outcome	BAP-1 evidence; BAP-3 measured skill status	CLM-1; external validity uninspected
RTE-4 / operational admission/selection/consumption	implemented	OBJ-14, OBJ-25; non-truth-apt design update	Grounding/inventory/process checks admit proposal/execution/build	Allowed input/provenance, not prior experimental success	BAP-2 constructs product; can reject	CLM-2; BUILD does not enforce prior TEST on inspected path
RTE-5 / content transformation	implemented	OBJ-12, OBJ-13; copied content non-ampliative, new prose indeterminate	Agent supplies text/IDs; teaching copies or creates content	Preserved source text or authored claim, not new truth	BAP-3 later peer context	CLM-1; semantic content unobserved
RTE-5 / operational admission/selection/consumption	implemented	OBJ-12, OBJ-13, OBJ-14; no content change	Locality/recipient IDs, published grounding and contributor constraints	Identity/material provenance only	BAP-3 transfer/proposal eligibility	CLM-2; no proof of derivation or request satisfaction
RTE-5 / retention	implemented	OBJ-12, OBJ-13; no content change	Archive append, personal storage, receipt persistence	No added epistemic license	BAP-3 later recall	CLM-1; not post-acceptance integration
RTE-6 / check/evidence production	implemented	OBJ-17; no content change	Syntax/register/instruction/parent/difference checks before install	DSL validity and lineage admissibility	BAP-2 permits/rejects executable update	CLM-2; not improvement
RTE-6 / disposition/acceptance	implemented	OBJ-18; no content change to program, evidence-presence status update	Measured observation marks verified	Reliance licensed only as a measured-known skill, not true explanatory claim	BAP-3 ranking/inheritance eligibility	CLM-1; no quality threshold
RTE-6 / operational admission/selection/consumption	implemented	OBJ-17, OBJ-18; non-truth-apt behavior update	Install code; optional turnover selects verified bounded records	Operationally valid known code/evidence	BAP-2 future VM; BAP-3 recipient context/parent knowledge	CLM-1; can inherit poor performers
RTE-7 / behavior/policy adaptation	implemented	OBJ-8; non-truth-apt numeric ranking update	Selection/citation/success counts enter lexical/feedback/exploration score	No truth judgment	BAP-3 ranking; BAP-1 changed exposure	None; actual model content dependence uninspected
RTE-7 / operational admission/selection/consumption	implemented	OBJ-9, OBJ-10, OBJ-12, OBJ-18; no content change except excerpts	Scheduled automatic selection and best-effort trimming	No new warrant from retrieval	BAP-1 delivered evidence	None; selected versus delivered distinction
RTE-9 / content transformation	implemented	OBJ-2, OBJ-11, OBJ-13; non-ampliative reshaping	Deterministic aggregate/participant views and receipts	Preserves selected record facts, not semantic fulfillment	BAP-1/BAP-3 context	None; rendered summary not separately durable
RTE-8 / lineage/freshness/recovery	implemented	OBJ-21, OBJ-22; acquisition/import and no content change	Compatible replay restores states/actions, checks projection digest	Consistency of represented state only	BAP-4 reject/report	CLM-2; adaptive statistics not reconstructed
RTE-10 / check/evidence production	implemented	OBJ-23; entailed derivation in configured numeric domain	Thresholds/field presence/physical outcome rules compute criteria	Defined outcome fits, not theory truth	BAP-4 observer result	CLM-1; measurements unexecuted here
RTE-10 / disposition/acceptance	implemented	OBJ-23; no content change	Functional outcome flag accepts composite artifact outcome against stated thresholds	Accepted scope would be this simulator outcome criterion	BAP-4 reporting, not model-prose promotion	CLM-1; no observed candidate acceptance
RTE-8 / check/evidence production	implemented	OBJ-23; entailed numeric comparison	Frozen-state artifact knockout/shared future, or held-out schedule evaluation	Artifact behavior under simulator intervention design	BAP-4 comparison output	CLM-1; no execution or recall intervention
RTE-2 / lifecycle integration	not determinable	OBJ-26; no established post-acceptance transition	No candidate-linked semantic acceptance/reliance chain inspected	No accepted-theory claim	Possible BAP-1 ordinary context use is separately wired	CLM-1; retention cannot fill missing acceptance
RTE-5 / lifecycle integration	not determinable	OBJ-12, OBJ-13; no established post-acceptance transition	Provenance-gated publication/transfer does not supply semantic acceptance	No accepted-publication claim	BAP-3 pre-acceptance use remains wired	CLM-1; candidate content absent
RTE-6 / lifecycle integration	not determinable	OBJ-24; no semantic acceptance-to-use transition	Measurements concern physical outputs, not explicit claim semantics	No accepted-explanation claim	BAP-2 code use is separately wired	CLM-1; successful artifact and true explanation differ

4. Per-object lifecycle disposition

For potential ampliative hypothesis OBJ-26: transformation is ampliative conjecture when the field states a non-entailed explanation; no instance observed. Observation/anomaly: RTE-3 and RTE-4, architectural status implemented, observed candidate state no instance observed. Conjecture: RTE-2, architectural status implemented for generation interface, observed candidate state no instance observed. Derived consequence: model-mediated RTE-2, architectural status not determinable, observed candidate state no instance observed. Test/evidence: RTE-4 is implemented for measurements, but candidate-specific test mapping is not determinable; observed candidate state no instance observed. Acceptance: RTE-2, evaluator/criterion for truth of the particular claim uninspected, intended use guidance, architectural status not determinable, observed candidate state no instance observed, accepted scope unestablished. Lifecycle integration: RTE-2 post-acceptance change unestablished, architectural status not determinable, observed candidate state no instance observed. The wired notebook route is retention, not evidence of an accepted theory.

For potential ampliative artifact-function claim OBJ-24: transformation is ampliative conjecture for an unentailed predicted function; no instance observed. Observation/anomaly: RTE-4, architectural status implemented, observed candidate state no instance observed. Conjecture: RTE-2, architectural status implemented for supplied spec, observed candidate state no instance observed. Derived consequence: predicted-effects field affords formulation but no explicit entailment check was inspected, architectural status not determinable, observed candidate state no instance observed. Test/evidence: RTE-4 and RTE-10 measurements implemented, claim-semantic mapping not determinable, observed candidate state no instance observed. Acceptance: RTE-10 accepts configured physical criteria rather than these words; claim evaluator/criterion/accepted scope unestablished, architectural status not determinable, observed candidate state no instance observed. Lifecycle integration: RTE-6 installs code before a semantic acceptance route is established; architectural status not determinable for post-acceptance integration, observed candidate state no instance observed. Retained code/claim coexistence does not close this lifecycle.

OBJ-12 and the free-message portion of OBJ-13: transformation indeterminate. Copying measured content, summarizing it, deriving consequences and making new conjectures remain possible. RTE-5 preserves record provenance and supplies retention/use, while grounding does not decide entailment. Exact candidate text with its inputs and later checks would distinguish these transformations; no instance observed. The copied-teaching portion of OBJ-13 is non-ampliative acquisition/copy through RTE-5; discovery lifecycle not applicable, source warrant no stronger than original content. Request receipts are non-ampliative records of selected action/request links through RTE-9; they do not certify satisfaction.

OBJ-2 and OBJ-15: acquisition/import from simulator sensors/measurements through RTE-4; discovery lifecycle not applicable. Their warrant is domain-limited measured state, with real-world validity excluded. OBJ-9 and OBJ-11 outcome portions: non-ampliative reshaping through RTE-3/RTE-9; discovery lifecycle not applicable; selected abbreviated details do not carry omitted causal explanation. Authored intentions remain policy content. OBJ-18 verification facts: non-ampliative measurement-status update through RTE-6, discovery lifecycle not applicable; no semantic or performance pass follows. OBJ-21 and OBJ-22: acquisition/copy and identity checks through RTE-8, discovery lifecycle not applicable; replay does not newly warrant recorded prose. OBJ-23: numeric derivation through RTE-10/RTE-8, discovery lifecycle not applicable for the arithmetic outputs; validity stays within supplied simulator state and configured criteria, with no observed outcome instance admitted.

No lifecycle record for OBJ-8: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-7.

No lifecycle record for OBJ-14: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-4.

No lifecycle record for OBJ-17: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-6.

No lifecycle record for OBJ-19: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-2, RTE-3.

No lifecycle record for OBJ-20: no runtime candidate truth-apt output for this static control object; relevant direct-adaptation or update routes: RTE-1.

No lifecycle record for OBJ-25: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-4.

No lifecycle record for OBJ-27: no candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-2, RTE-7.

5. System-claim versus route comparison

Claim	Doctrine/design support	Implementation	Observed support	Causal support	Supported conclusion and mismatch/unknown
CLM-1	README/current architecture describe discovery, evidence exchange and executable inheritance	RTE-2, RTE-4, RTE-5, RTE-6, RTE-8, RTE-10	None admitted	None admitted; intervention code is not an executed comparison	Operations to generate, measure, exchange and inherit are wired. Useful discovery, criticism-driven improvement and recalled-content dependence are unresolved.
CLM-2	Simulator authority, deterministic tick/replay and separate evidence/prose	RTE-2, RTE-3, RTE-4, RTE-6, RTE-8	None admitted	None admitted	Numeric/action authority is bounded and wired; entry-path freeze differs, sequential rejection need not rollback, verified is measurement presence, replay omits selector metadata and VM accumulation differs from a simple total actuation cap.

6. Bounded conclusion

The runtime acquires local simulated evidence, retains authored hypotheses and plans, measures recipes/artifacts, checks formal/material/provenance constraints and uses code/records operationally. Numeric checks warrant their configured simulator outcomes. They do not generally accept the truth of research-state prose, publication explanations or predicted-function wording. Skill status changes later selection with a specifically limited license: measured and known. Adaptive recall statistics and plan/program replacement can change behavior without a truth-apt acceptance transition. Experimental improvement and content-directed criticism remain uninspected at this evidence boundary; these limits are not a claim that model reasoning lacks them.

Reconciliation

The specialist report's run, source, full revision, complete status, frozen input hash and method hash (7e86ed242caadc095d7f20ccd7fcbcb837b9d62e94fca50656b782c65cb0d675) were checked before integration. The coordinator read the entire typed report and retained its source evidence. No prior review or ingest supplied characterization. The specialist supplies memory analysis, not independent semantic clearance of this synthesis.

Proposal mapping uses exact tokens: MEM-OBJ-1 → OBJ-8; MEM-RTE-1 → RTE-9; MEM-ABS-1 → limitation on RTE-8, not an ABS record. The last proposal already carried uninspected status; its local prefix did not establish absence. Known profile axes retain the specialist's values, basis and scope, with canonical references expanded when objects were split. Superseded OBJ-1, OBJ-3, OBJ-4, OBJ-5 and OBJ-6 keep their original combined referents and evidence, with new monotonic part IDs OBJ-9, OBJ-10, OBJ-11, OBJ-12, OBJ-13, OBJ-14, OBJ-15, OBJ-16, OBJ-17, OBJ-18, OBJ-19 and OBJ-20. The source's broader containers OBJ-7, OBJ-10 and OBJ-16 retain identity while their unlike operative parts receive OBJ-21, OBJ-22, OBJ-23, OBJ-24, OBJ-25, OBJ-26 and OBJ-27. Lens overlays do not redefine these records.

All material integration issues are disposed: (1) adaptive statistics are OBJ-8 rather than memory content; (2) derived context/receipts are RTE-9, with ephemeral rendering distinguished from durable source records; (3) faithfulness remains an evidence limitation; (4) defaults, enabled profile and optional turnover are separate; (5) verified means measured, view sorting differs from inheritance filtering; (6) citation, successful cited action and fixed-action intervention do not prove recall dependence, and selected diagnostics can exceed delivered content; (7) replay does not reconstruct adaptive statistics or promise future prompt equivalence; (8) all automatic trace-fed continuation/outcome/recipe/skill/counter/receipt routes enter the comparison dependencies, while raw logs/rendering do not; (9) final ownership, checks and publication remain coordinator-owned.

The coordinator sent the worker questions about BUILD's tested-recipe wording, measurement-only verification and VM instruction versus aggregate bounds. Findings on those prompted issues are coordinated checks, not independent convergence. The parent retained the source-supported caveats on RTE-3, RTE-4 and RTE-6. Coordinator normalized three specialist citation endpoints to the full pinned blob bounds: events.py ends at154 and the named profile ends at99. The specialist could not be resumed because the shared thread slots were occupied; parent confirmed this mechanical correction. Quotation bytes, findings/profile, frozen input and method remain unchanged. The final report hash above identifies the corrected report. No unresolved substantive conflict requires choosing a stronger status. Memory generic records belong to the canonical register; the local epistemic lens adds warrant/phase annotations only. RTE-10 and BAP-4 make observer outcome criteria distinct from model/agent authority.

Bounded synthesis

SwarmWorld wires an episode-level experimental society: models propose local plans and technological artifacts, a deterministic simulator decides physical consequences, and explicit private/shared memory carries experience into later calls. The artifact programs remain executable after the authoring call, and optional inheritance can carry measured skill records across agent replacement. These are concrete persistence and revision mechanisms; their value is not contingent on proving provider weight change.

The strongest supported learning contribution is automatic continuation-state, measurement/input association, action feedback and adaptive retrieval written during an episode and read by later policies or selectors. Trace learning is wired in the comparison's structural sense. Criticism of an operative formulated theory improving capacity for future action has conclusion status uninspected: candidate-linked theory application, content-directed criticism and attributable gain were not supplied by this source-only pass. Stored hypothesis text and success counters cannot fill those links.

Reflection has conclusion status wired in a bounded operational sense. The system represents selected internal agent history/progress/intent and access statistics; action/measurement changes update those representations; subsequent context queries and ranking consume them and can change later behavior. The two-way dataflow on RTE-2, RTE-3 and RTE-7 supports this limited self-representation. Reflection specifically on a theory of the system's theory-building organization has conclusion status uninspected. Useful improvement does not follow from the reflective loop.

Self-improving operation has conclusion status afforded: evidence can guide plan, research-state, recipe and program replacement through RTE-2, RTE-4 and RTE-6; admission does not require better performance, and no retained comparison establishes net improved capacity. Actual self-improvement has conclusion status uninspected. This separates a route for revising behavior from demonstrated benefit and from criticism of a theory.

For simulator-domain experiments, the source usefully separates measurements from prose and retains provenance/input associations. Its gates have deliberately different licenses: syntax permits execution; resource/provenance permits combination; measured status permits priority/inheritance; scorecard criteria describe physical outcomes. None is a general truth gate for explanations. BUILD's tested-recipe wording, the broader verified label and replay's state projection need the concrete limits stated above. A follow-up could change this assessment by retaining candidate-specific hypothesis/consequence/test/revision traces, measured matched learning comparisons, recalled-content interventions with replanning, a tested-construction gate, or replay restoration/checks that include access metadata.

Limitations

Limitation	Affected IDs	Inspected boundary	Conclusion prevented	Evidence that would resolve it
No candidate-linked operation admitted	SRC-1, RTE-2, RTE-4, RTE-6, RTE-10	Implementation/doctrine only	Observed learning, criticism, improvement or accepted theory	Frozen episode artifacts linking exact candidates, inputs, outcomes and later use
Provider internals and exact weights opaque	CMP-3, RTE-2	Request/response interface	Hidden reasoning process, weight fixity/update, reliable inference	Inspectable provider/model boundary and applicable probes
Recall dependence not tested in admitted evidence	OBJ-8, RTE-7, RTE-8	Retrieval/citation/replay code	Faithful citation or causal use of recalled content	Retained interventions on recalled content with fresh decisions and controls
Physics validity excluded	CMP-1, OBJ-15, OBJ-23, RTE-10	Configured surrogate outcomes	Real-world scientific validity or explanatory transfer	External validation of equations/measurements and domain transfer evidence
Scenario and oracle-policy internals not exhaustive	OBJ-20, RTE-1	Default biological route/named profile and entry factory	Every scenario equivalence or oracle-policy information symmetry	Frozen scenario-specific and oracle-policy implementation analysis
Replay omits access metadata	OBJ-8, OBJ-22, RTE-8	Current serializer/restorer	Full future model-context equivalence or supported resume	Restoration and equality coverage for selectors/policy state
Static checks only	All inspected runtime routes	No target execution	Deployment reliability, outage recovery rate, scaling or empirical effect size	Bounded execution evidence under declared tools/services/configuration

Verification and blockers

Semantic verification

Checked source boundary, full revision and source-only independence; complete specialist identity and unchanged input/report/method hashes; canonical object splits and exact-token proposal mappings; all fourteen comparison axes; quote text occurrence and local anchors; route ownership and status separation; configuration alternatives; all material admissions and authority horizons. Every material route has immediate return, later read-back, visibility, selection/expiry, activation limit and recovery disposition. Dynamic behavior remains unobserved.

Profile checks cover RTE-2, RTE-3, RTE-4, RTE-5, RTE-6, RTE-7, RTE-8 and RTE-9. Trace-fed union includes continuation, outcomes/intent, recipe-linked measurements, skill observations, adaptive metadata and receipts with event-stream source, per-episode task horizon, online timing and prose/symbolic form. Pure rendering/raw recording and static configuration were excluded explicitly. Push consumer/trigger/input/selected part are recorded for lexical private/archive, coarse locality/recency/limits, exact program focus and participant requests. Requested record resolution and operator replay are pull. Inaccessible provider state is excluded rather than silently represented as a known memory form. Citation and delivery never establish activation; replay/knockout design never substitutes for an executed faithfulness comparison. Epistemic checking, disposition, retention and post-acceptance integration remain separate.

Deterministic validation

Validation target: kb/reports/state/agentic-system-analysis/AAS-2026-09-25-swarmworld-01/result.md. commonplace-validate --full passed cleanly. Full pinned-blob checks matched all40 quotes and verified citation endpoints; source-quote and anchor checks precede publication. No review-job or receipt is embedded in this result.

Blockers

None.