SwarmWorld epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is how SwarmWorld's simulated science acquires evidence, evaluates candidates and licenses later reliance. Scope and exclusions are those of the boundary; source identity and layers are in its Source register. All direct reads below are SRC-1 at the frozen commit, supplying implementation evidence unless explicitly labelled doctrine/design. No target execution, paper dataset, observed run or causal experiment is supplied. Physical-world validity, actual model compliance, exercised learning and empirical scaling therefore remain unestablished. Provider internals and weights remain outside accessible evidence; see RT-CMP-configured-model.

The public research-environment claim is EPI-CLM-research-environment. The model's grounding instruction is RT-CLM-grounded-science. Neither is read as a claim of infallible scientific knowledge. The strongest inspected checks concern simulator measurements, recipe eligibility, typed plans, DSL validity and replay agreement. These are different targets.

Entry point or operation	Source path	Coverage and records, or remaining limit
CLI dispatch and ordinary model planning	src/biofoundry/cli.py; src/biofoundry/policies/llm.py	Direct reads of parser, main dispatch, prompt construction, evidence-ID filtering and output handling; supplied RT-RTE-model-planning covers scheduling/schema/freeze.
INSPECT, passive evidence and action feedback	src/biofoundry/simulation.py; src/biofoundry/knowledge.py	Direct reads; EPI-RTE-sensor-evidence. Skill-library verified=True is observed-measurement provenance here, not proof of general utility. Full program-library promotion/ranking lifecycle remains uninspected.
OPERATE and TEST	src/biofoundry/simulation.py; src/biofoundry/science.py	Direct reads; EPI-RTE-batch-measurement. Material-physics implementation is not fully audited; real-material accuracy cannot be concluded.
PROPOSE_RECIPE and COMBINE_DESIGN	src/biofoundry/simulation.py; src/biofoundry/science.py	Direct reads; EPI-RTE-recipe-admission, EPI-RTE-proposal-evaluation. These add evaluator and provenance routes not separated in runtime.
PUBLISH, DEPOSIT_INSIGHT and TEACH	src/biofoundry/simulation.py; src/biofoundry/memory.py	Direct reads; EPI-RTE-insight-circulation. Message transport outside these bodies is supplied runtime coverage, not a full communication audit.
Retained excerpts and feedback-weighted retrieval	src/biofoundry/memory.py; src/biofoundry/policies/llm.py; src/biofoundry/simulation.py	Direct reads of selection and later executor outcome feedback; EPI-RTE-memory-selection. Full public BM25 implementation and all semantic-observation omission paths remain uninspected; no complete-packet fidelity claim.
BUILD, WRITE_PROGRAM, FORK_PROGRAM	src/biofoundry/simulation.py; src/biofoundry/programs.py; src/biofoundry/artifacts.py	Supplied RT-RTE-program-install and RT-RTE-world-stepping; not independently audited here.
Scripted, human, live and environment actions	src/biofoundry/server.py; game/session.py; src/biofoundry/env.py; src/biofoundry/policies/scripted.py	Supplied RT-RTE-direct-control; model-schema conclusions do not extend to these routes.
replay, playback and counterfactual-replay	src/biofoundry/cli.py; src/biofoundry/counterfactuals.py; src/biofoundry/playback.py	Parser directly read; implementation account supplied by RT-RTE-recorded-replay. Playback presentation not audited; digest agreement is not semantic endorsement.
ecosystem-report and checkpoint/held-out evaluation	src/biofoundry/cli.py; src/biofoundry/technology_ecology.py	Report wrapper directly read; evaluator supplied by RT-RTE-ecology-assay. Reports consume results; no live plan-admission feedback established.
research-study, analyze-study, analyze-mobility, trace-report, benchmark	src/biofoundry/cli.py	Dispatch directly inspected; trace-report wrapper read. Study member selection, statistical inference, mobility and trace attribution internals not assessed. This prevents system-complete conclusions about evaluators or automatic successor selection.
technology-dossiers and technology-atlas	src/biofoundry/cli.py; src/biofoundry/technology_visualization.py; src/biofoundry/technology_atlas.py	Dispatch and wrappers directly read: optional image generation, embeddings and ranked shortlists are additional exposed routes. Downstream prompt construction, semantic fidelity, embeddings and image-model versions are uninspected; generated depictions cannot be certified as measured mechanisms.
world validate/describe, doctor and external world authoring	src/biofoundry/cli.py; src/biofoundry/scenarios.py; .agents/skills/swarmworld-world-builder/SKILL.md	Dispatch directly read; supplied RT-RTE-scenario-admission. Mechanical package acceptance differs from scientific validation or human review.

This is a bounded epistemic assessment inside the whole-system source boundary, not an exhaustive semantic audit. Available but uninspected operations above preclude a system-wide negative about acceptance, criticism, knowledge production or adaptation. No supplied runtime record needs supersession: most new routes separate material parts of world stepping; memory selection and outcome feedback span planning and world stepping rather than being contained in either.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-empirical-knowledge	See record	EmpiricalKnowledge	SRC-1, src/biofoundry/knowledge.py	Historical local evidence, not global current truth
EPI-OBJ-experimental-batch	See record	ExperimentalBatch	SRC-1, src/biofoundry/science.py	Simulated measurements; hidden target flag is not delivered privately
EPI-OBJ-memory-record	See record	MemoryRecord	SRC-1, src/biofoundry/memory.py	Text may be a claim, intention or policy; instances not supplied
RT-OBJ-action-queues	None necessarily; see record	Agent action queues	SRC-1, src/biofoundry/policies/llm.py	Executable action authority is not propositional warrant
RT-OBJ-artifact-program	Not determinable for an unspecified program; see record	Artifact behavior program	SRC-1, src/biofoundry/programs.py	A controller may encode a theory but no instantiated theory is supplied
RT-OBJ-scenario-prompt	See record	Scenario agent instructions	SRC-1, src/biofoundry/scenarios.py	External authored premises and guidance, not episode-produced knowledge
RT-CMP-configured-model	Not individuated in accessible weights; see record	Configured provider model	SRC-1, src/biofoundry/providers/openai_compatible.py	Parametric provenance, exact version and internal updates uninspected

Authority-route ledger

Fields owned by a declaring record are incorporated by citation, not duplicated. Content/update relations below classify the indicated edge only. Architectural status refers to implementation, not observed activation.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
EPI-RTE-sensor-evidence	content transformation	implemented	EPI-OBJ-empirical-knowledge	truth-apt transformation: acquisition/import	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/knowledge.py	none	none	Warrant restricted to simulator source and observation time
EPI-RTE-batch-measurement	check/evidence production	implemented	EPI-OBJ-experimental-batch	truth-apt transformation: entailed derivation	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/science.py	none	none	Arithmetic and state readout, conditional on engine premises
EPI-RTE-batch-measurement	retention	implemented	EPI-OBJ-experimental-batch	truth-apt transformation: non-ampliative reshaping	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/simulation.py	none	none	Measurement retention is not acceptance of an explanation
EPI-RTE-recipe-admission	disposition/acceptance	implemented	EPI-OBJ-memory-record	no content change	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/simulation.py	none	none	Operational admission only; not truth acceptance
EPI-RTE-proposal-evaluation	check/evidence production	implemented	EPI-OBJ-experimental-batch	truth-apt transformation: entailed derivation	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/science.py	none	none	Matched material ablation, not agent/component causal attribution
RT-RTE-model-planning	content transformation	implemented	EPI-OBJ-memory-record	truth-apt transformation: indeterminate	see record	see record	see record	see record	see record	see record	see record	see RT-BAP-model-guidance	SRC-1, src/biofoundry/policies/llm.py	RT-CLM-grounded-science	none	Hypothesis schema permits candidates, but no output proves entailment or non-entailment
RT-RTE-model-planning	disposition/acceptance	implemented	RT-OBJ-action-queues	no content change	see record	see record	see record	see record	see record	see record	see record	see RT-BAP-action-dispatch	SRC-1, src/biofoundry/structured_output.py	none	none	Shape/type admission is not empirical claim checking
EPI-RTE-insight-circulation	retention	implemented	EPI-OBJ-memory-record	truth-apt transformation: indeterminate	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/simulation.py	none	none	Prefixing/truncation can lose necessary scope; not guaranteed semantic preservation
EPI-RTE-insight-circulation	operational admission/selection/consumption	implemented	EPI-OBJ-memory-record	no content change	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/simulation.py	none	none	Public grounded-resource metadata can grant recipe eligibility, not truth of prose
EPI-RTE-memory-selection	content transformation	implemented	EPI-OBJ-memory-record	truth-apt transformation: indeterminate	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/memory.py	none	none	Unstructured truncation can discard a negation or condition
EPI-RTE-memory-selection	behavior/policy adaptation	implemented	EPI-OBJ-memory-record	non-truth-apt policy/content update: retrieval ranking statistics	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/memory.py	none	none	Planning citations and later executor success update future ranking; credit is not causal identification
RT-RTE-program-install	behavior/policy adaptation	implemented	RT-OBJ-artifact-program	non-truth-apt policy/content update: installed artifact instructions	see record	see record	see record	see record	see record	see record	see record	see RT-BAP-program-execution	SRC-1, src/biofoundry/programs.py	none	none	Useful effect or theory content requires further evidence
RT-RTE-recorded-replay	lineage/freshness/recovery	implemented	EPI-OBJ-memory-record	truth-apt transformation: acquisition/import	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/counterfactuals.py	none	none	Recorded-state warrant only; no new model reasoning
RT-RTE-ecology-assay	check/evidence production	implemented	RT-OBJ-artifact-program	truth-apt transformation: entailed derivation	see record	see record	see record	see record	see record	see record	see record	see record	SRC-1, src/biofoundry/technology_ecology.py	EPI-CLM-research-environment	none	Conditional simulated comparison, no supplied actual trajectory

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-research-environment	see record	SRC-1, README.md; doctrine/design	see record	EPI-RTE-sensor-evidence, EPI-RTE-batch-measurement, EPI-RTE-insight-circulation, RT-RTE-program-install, RT-RTE-ecology-assay	Uninspected; no supplied episodes	No actual intervention comparison; surrogate domain only	Implemented environment for measurement, exchange and persistent simulated control	Scientific discoveries, benefit, transfer and real-world mechanisms remain unestablished
RT-CLM-grounded-science	see record	SRC-1, src/biofoundry/policies/llm.py; doctrine/design	see record	RT-RTE-model-planning, EPI-RTE-recipe-admission, EPI-RTE-batch-measurement	Uninspected	Resource grounding mechanically restricts recipes; it does not check arbitrary hypotheses	Guidance and bounded evidence access are wired	Do not upgrade a request against invented observations into a guarantee of truthful prose

Bounded conclusion

SwarmWorld acquires local simulator evidence and retains it for later model input. Observation-time labels preserve a bounded lineage; current possession and historical observation grant resource eligibility, not knowledge of an unobserved world. TEST reads simulated batch properties and computes a utility threshold; its private consumer receives measurements and utility but not the threshold flag. These are evidence-producing and arithmetic routes, not evidence-consuming acceptance of a newly proposed scientific explanation (EPI-RTE-sensor-evidence, EPI-RTE-batch-measurement).

A counterexample to “only TEST evaluates candidates” is proposal-time latent material evaluation, including leave-one-inherited-material-out comparisons. A counterexample to “publication merely stores text” is the way named, personally grounded resources become citation-based recipe eligibility. These routes carry real computational and operational authority, but their license is restricted respectively to the simulator's utility contrast and to provenance/resource admission. Neither establishes acceptance of the publication's prose or the causal explanation named by a model (EPI-RTE-proposal-evaluation, EPI-RTE-recipe-admission, EPI-RTE-insight-circulation).

Model-authored hypotheses and progress can be stored and delivered again. Their content relation is indeterminate without an output instance; a field named hypothesis does not prove an ampliative claim. Evidence IDs are filtered for availability, not for support. Public and taught records also gain later availability without a truth-acceptance gate in the inspected writers. No lifecycle integration after epistemic acceptance is established by those writers; retention, citation eligibility and prompt consumption are separate functions. A complete system-wide absence cannot be concluded because study selection, program-library curation and analysis routes remain partly uninspected (RT-RTE-model-planning, EPI-RTE-insight-circulation).

Retrieval implements evidence-responsive ranking updates even without a formulated theory. Selection and citation processing occur during planning; action-success counters update during subsequent world stepping and feed later retrieval. The full feedback route spans both phases, not just pre-inference planning. Its credit counts cite-and-succeed associations, not independent variation of the record or causal truth. Excerpts can discard semantic conditions, and subsequent context trimming can omit selected records entirely. The stored original survives, so read-back is possible but not guaranteed delivery in a particular invocation. These changes establish an adaptive selection capacity, not improved future performance, exercised reflection or exercised self-improvement (EPI-RTE-memory-selection).

Programs have enforcing actuator authority after installation, while natural-language claims have interpretive force at model/human consumers. DSL validation licenses bounded execution, not usefulness; replay licenses agreement with compatible recorded state, not scientific correctness. Frozen-society assays can measure simulated service persistence and knockout contrasts, but the inspected episode path records their results rather than automatically selecting a successor controller. No inspected actual experiment supports a causal benefit claim (RT-RTE-program-install, RT-RTE-recorded-replay, RT-RTE-ecology-assay).

Theory-builder conditions, learning, reflection and autonomy remain independent. The model route affords localized hypothesis text and later persistence; content-dependent use, explicit blame-bearing criticism and criticism-led iteration remain uninspected, as recorded in RT-RTE-model-planning. Automatic scheduling is not evidence for those properties. Nothing in this account establishes weight updates: external parameter mutability remains uninspected independently of the implemented simulator-memory adaptation (RT-CMP-configured-model).

Shared records

Operative objects

EPI-OBJ-empirical-knowledge

Label: EmpiricalKnowledge

Identity: each agent's EmpiricalKnowledge, containing CellEvidence keyed by cell and extraction counters. Representational form: symbolic coordinates, resource/terrain/station enum values, times, masses and counts. Storage substrate: in-process dictionaries/counters, serializable through authoritative state. Candidate truth-apt content is that a cell had the recorded sensor values at the recorded tick, plus extraction-attempt outcomes. This is distinct from RT-OBJ-action-queues: evidence rather than selected actions. No supplied object is its counterpart. SRC-1, src/biofoundry/knowledge.py supplies implementation.

self.cells[y * self.width + x] = CellEvidence( --- src/biofoundry/knowledge.py

"scope_note": "Unlisted materials are unobserved, not proven absent.", --- src/biofoundry/knowledge.py

EPI-OBJ-experimental-batch

Label: ExperimentalBatch

Identity: research-ledger microbatch, source recipe and tests entries; distinguish pending latent properties from tested readout and hidden passes_material_target. Form: symbolic structured recipe/ID/provenance plus numerical material properties and utility. Substrate: ResearchLedger.batches, tests, proposals, private notebook experiment-result JSON and events. Truth-apt candidates concern simulated batch properties and matched utility differences, not real materials. Unlike RT-OBJ-artifact-program, this is measurement/evaluation state rather than installed behavior. No counterpart after comparison. SRC-1, src/biofoundry/science.py supplies implementation.

Pending samples expose their identity and authored recipe, but not latent material properties or utility. Those values enter the view only after TEST. --- src/biofoundry/science.py

EPI-OBJ-memory-record

Label: MemoryRecord

Identity: ID-addressed text records consumed by private retrieval, teaching and public archive, including self-authored research state and insight text. Natural-language form for model interpretation; IDs, timestamps, kind and causal-parent fields are symbolic metadata. Storage substrate: working/episodic/notebook collections and CulturalArchive.records/by_id, with optional trace serialization. Candidate propositions depend on content; goal and action-intent strings need not be truth-apt. Text is not made warranted by kind, salience or citation. Distinct from RT-OBJ-action-queues and RT-OBJ-scenario-prompt: use-generated retained content rather than executable queue or static instructions. SRC-1, src/biofoundry/memory.py, src/biofoundry/policies/llm.py, src/biofoundry/simulation.py supply implementation.

f"Prior self-authored research state: {prior_state}\n{tail}" --- src/biofoundry/policies/llm.py

Routes

EPI-RTE-sensor-evidence

Label: Sensor evidence acquisition Part of: RT-RTE-world-stepping

Endpoints: world/local artifact state → observation records and empirical cell evidence → knowledge view, retained notebook and optional measured program evidence. Owner: simulator and EmpiricalKnowledge. This is the sensing subroute, not a duplicate of the full step. SRC-1, src/biofoundry/simulation.py, src/biofoundry/knowledge.py; implementation.

• Immediate return: INSPECT emits sample_inspected; observe updates the most recent cell value; view returns counts, bases and nearest historically observed sites.
• Later read-back: notebook observations can enter subsequent retrieval; cell evidence enters later knowledge views. Historical sites retain last_seen_tick and last-seen mass.
• Delegated visibility: agent-local sensor evidence, not authoritative global arrays; human/analysis events may expose more than private model packets.
• Selection predicate: INSPECT current tile and nearby active programmed artifacts within distance one; empirical view groups observed cells and matter backed by observation or conserved possession.
• Invalidation or expiry: revisiting a cell replaces that cell's latest evidence; historical evidence is not a freshness guarantee about current world mass.
• Activation or effect: implementation writes notebook evidence and may mark observed program knowledge verified=True; model reliance unobserved. grounded_resources later gates recipe eligibility.
• Evidence limits: simulator source warrant preserved as time/locality-labelled measurements, not independently checked against physical reality; full program-library verification consumers uninspected.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Evaluator target/domain: local simulator cell and artifact values; program reads and numeric rounding, not model judgment. Force: measurement availability and enforcing resource eligibility through the recipe gate; program verification concerns measured observation. Consumer/channel/horizon: later knowledge.view, memory retrieval and the model prompt, advisory over subsequent macroturns; recipe checker consumes resource sets with enforcing force. No explanation accepted or integrated by this route. Recovery is new sensing or ordinary record retrieval; no universal withdrawal of stale claims is established.

self.program_library.observe( agent_id, program.program_id, tick=self.tick, source="measured_artifact", evidence_id=record_id, verified=True, measurements=measurement, ) --- src/biofoundry/simulation.py

EPI-RTE-batch-measurement

Label: Batch measurement and retention Part of: RT-RTE-world-stepping

Endpoints: controller-authored OPERATE recipe → fabricated/staged microbatch → TEST newest pending sample → numeric result and private notebook. Owner: simulator and ResearchLedger. This part differs from general action execution by its measured target and private information boundary. SRC-1, src/biofoundry/simulation.py, src/biofoundry/science.py; implementation.

• Immediate return: OPERATE emits sample identity/recipe/mass; TEST marks a batch tested and emits properties, utility and target-threshold result in the event.
• Later read-back: measured result retained in ledger and notebook, available to later private experiment view/retrieval. Source recipe is added to the test result when experience_reuse is enabled.
• Delegated visibility: private experiment view and notebook omit passes_material_target; events carry it. Pending samples do not reveal latent properties or utility through this view.
• Selection predicate: OPERATE requires recipe, suitable foundry/station/workspace, validated fabrication and feedstocks; TEST selects latest untested sample for that agent and rejects when none exists.
• Invalidation or expiry: testing consumes pending status, not the underlying result; old results remain historical tests rather than fresh observations.
• Activation or effect: fixed code computes numeric readout and threshold classification; later prompts can receive it, without demonstrated content dependence.
• Evidence limits: no executed sample; physics, encoding and utility adequacy not independently validated. A utility threshold is not acceptance of a causal theory.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Target/domain: the particular simulated batch's property dictionary and utility under ResearchLedger.material_utility. Evaluator: fixed code and configured target_material_utility; comparison result licenses only that utility criterion at the test tick. Model/human/script proposes actions, software vetoes invalid execution; no reference scientific answer or per-content human approval. Mode: bounded simulated experiment. TEST does not install a recipe, approve a publication or automatically revise a hypothesis. Separate retention follows measurement without a named explanatory acceptance decision. Consumer/channel/force/horizon: private notebook/result view → future model prompt, advisory; research event/ledger → researcher reporting, descriptive. Recovery: fabricate/test another sample; no automatic scientific rollback.

"passes_material_target": ( self.material_utility(batch.batch) >= self.config.target_material_utility ), --- src/biofoundry/science.py

observable_result = { key: value for key, value in result.items() if key != "passes_material_target" } --- src/biofoundry/simulation.py

EPI-RTE-recipe-admission

Label: Recipe provenance admission Part of: RT-RTE-world-stepping

Endpoints: PROPOSE_RECIPE/COMBINE_DESIGN → material validation, resource/provenance checks → recorded eligible proposal or rejection. Owner: simulator. This part handles eligibility, separately from EPI-RTE-proposal-evaluation's utility calculations. SRC-1, src/biofoundry/simulation.py; implementation.

• Immediate return: recipe_proposed/design_combined event and notebook proposal, or validation error.
• Later read-back: retained design record can be retrieved; proposal ledger preserves recipe and parents.
• Delegated visibility: record says recipe validated; that means these admission checks, not empirical confirmation of a mechanism.
• Selection predicate: valid recipe; own observed/possessed resources or allowed cited resources; combined designs require multiple published parents, structured provenance and configured distinct grounded contributors.
• Invalidation or expiry: historical proposal remains; this method implements no empirical-failure withdrawal of its validation label.
• Activation or effect: gate permits/rejects the proposal and records parent-resource inheritance. Later fabrication/build checks are separate.
• Evidence limits: a grounded material name or multiple authors do not prove the parent statement true, independent evidence sources, or a useful composition.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Target: recipe executability/provenance eligibility within collective science, not truth of prose. Proposal owner is selected model/human/script; fixed-rule software decides/vetoes, without a reference answer or required human approval. Guidance is the local material/manual constraint; what persists is recipe, IDs and validation result. Force is enforcing on this action. Consumer/channel/horizon: simulator proposal handler via resource sets/parent metadata, immediate action admission; retained labels later have interpretive force at model consumers. Recovery: revise recipe, obtain/observe inputs or cite eligible parents. No evidence-consuming acceptance of an ampliative explanatory claim occurs in this inspected gate.

"COMBINE_DESIGN requires empirically grounded publications from at " f"least {self.config.science.minimum_contributors} distinct agents" --- src/biofoundry/simulation.py

ungrounded = sorted(required - grounded, key=int) --- src/biofoundry/simulation.py

EPI-RTE-proposal-evaluation

Label: Proposal material evaluation Part of: RT-RTE-world-stepping

Endpoints: admitted normalized recipe → synthetic inventory execution → full utility and inherited-material ablations → proposal ledger. Owner: ResearchLedger.propose. Distinct check/evidence production within world stepping, not the eligibility gate. SRC-1, src/biofoundry/science.py; implementation.

• Immediate return: proposal includes evaluation_utility and optional composition_synergy.
• Later read-back: proposal ledger retains metrics for research-state/report consumers; the complete downstream scorecard/analysis consumer tree is uninspected here.
• Delegated visibility: authoritative evaluation is excluded from agent observation by design comment; private experiment view exposes measurements only after TEST. Proposal events are broader than that private view.
• Selection predicate: every admitted proposal; ablations only for combined proposals with inherited resource slots and a nonempty remaining recipe.
• Invalidation or expiry: historical proposal metrics persist; no automatic rejection or controller rollback based on negative synergy in this function.
• Activation or effect: computes full utility minus the strongest available leave-one-inherited-material-out utility under unchanged processing; writes result, not a live successor choice.
• Evidence limits: simulator derivation conditional on evaluator/recipe encoding. Ablation compares material inclusion, not the independent causal contribution of cited publications, communication or model intelligence. No observed comparison supplied.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Target/domain: utility contribution of inherited materials to the proposed recipe in the simulator. Fixed MaterialLab execution and material_utility form the evaluator; full recipe is the reference outcome, matched ablations are contrasts. Mode: bounded computational assay, not an empirical physical test. Consumer/channel/force/horizon: numeric proposal-ledger fields and events, descriptive to offline research; any further study selection remains uninspected. No criterion here licenses a new explanatory claim for reliance, and no performance-based acceptance is implemented in this function.

composition_synergy = ( round(full_utility - max(ablation_utilities), 6) if ablation_utilities else None ) --- src/biofoundry/science.py

Authoritative evaluation is never included in the agent observation.

"evaluation_utility": round(full_utility, 6), --- src/biofoundry/science.py

EPI-RTE-insight-circulation

Label: Insight retention and circulation Part of: RT-RTE-world-stepping

Endpoints: controller insight/message → private notebook/spatial deposit → optional public archive and grounded-resource publication metadata; or TEACH retained record/program IDs/message → nearby recipients. Owner: simulator/archive. This is a retained-content subroute; program transfer is noted but its full library curation is not assessed. SRC-1, src/biofoundry/simulation.py, src/biofoundry/memory.py; implementation.

• Immediate return: deposited insight or knowledge-taught event; record is retained, publication adds archive entry and structured provenance.
• Later read-back: archive retrieval, recipients' notebooks and own memory can supply later model input. Combined design can consume publication resource metadata.
• Delegated visibility: deposited content is author-prefixed and limited to 1000 characters; TEACH copies selected retained records or supplies a fresh message when no records resolve. IDs/provenance remain metadata, not semantic verification.
• Selection predicate: nonempty insight; TEACH requires nearby addressed recipients and retained records/known programs or nonempty message. Public resource metadata attaches only personally grounded resource names found in publication text.
• Invalidation or expiry: archive deduplicates by record ID; inspected writer has no truth-based retraction, semantic supersession or publication expiry. Other uninspected routes prevent a global absence conclusion.
• Activation or effect: archive/notebook availability, optional immediate-replan interrupts, and enforcing cited-resource eligibility in COMBINE_DESIGN. Actual model reliance remains unobserved.
• Evidence limits: lexical mention plus personal resource grounding establishes resource provenance, not support for every assertion in the publication. Teaching does not upgrade warrant.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Target/domain of admission: nonempty retained content and transfer/publication eligibility. Model/human/script authors; software decides these conditions, not propositional truth; no answer oracle or per-content human approval required. Guidance originates in action manual and controller intent. Consumer/channel/force/horizon: model via retained text, advisory across macroturns; simulator recipe checker via grounded_resources, enforcing for later combined proposals; replanning scheduler via boolean interrupts, permissive for later scheduling. Retention precedes epistemic acceptance; no post-acceptance lifecycle integration established. Recovery is subsequent content/action, not automatic scientific rollback.

content=f"{agent_id} | {content[:1000]}", --- src/biofoundry/simulation.py

normalized = content.casefold().replace("", " ") personal = self.grounded_resources(index, include_shared_depot=False) return { resource for resource in personal if self.resource_name(resource).casefold().replace("", " ") in normalized } --- src/biofoundry/simulation.py

for recipient in recipients: for record in records: self.memories[recipient].remember(record, notebook=True) --- src/biofoundry/simulation.py

EPI-RTE-memory-selection

Label: Memory excerpt and adaptive selection

Endpoints: prior research state/pending requests/local observation → query → scored retained records → bounded excerpts → optional packet trimming → model input; accepted citations update use statistics during planning, and subsequent world stepping updates action-outcome statistics that feed later rankings. Owner: AgentMemory, LLM policy and simulator feedback writer. This route spans RT-RTE-model-planning and RT-RTE-world-stepping, overlapping their selection/citation and executor-feedback stages respectively. It is not contained in either runtime route and does not duplicate either full route. It also overlaps MEM-RTE-retrieval-feedback in the selection/use/outcome feedback chain, while additionally tracing excerpt transformation and packet omission; the overlapping feedback mechanism has the same identity, not an independent evaluator. SRC-1, src/biofoundry/memory.py, src/biofoundry/policies/llm.py, src/biofoundry/simulation.py; implementation.

• Immediate return: retrieval returns ranked records and diagnostics and updates selection counts; packet construction renders bounded excerpts; accepted citation processing updates use counts during planning. Later _record_action_outcome calls in world stepping separately update outcome counts after planning has committed its queue.
• Later read-back: retained statistics change adaptive retrieval on later invocations; records omitted from a prompt remain in stores and can be retrieved again.
• Delegated visibility: model sees only remaining excerpts after trimming, with record ID/kind/tick; full retained source remains outside the message window. A selected diagnostic is not proof of delivery after trimming.
• Selection predicate: deduplicated notebook/episodic records, lexical score plus salience and tiny recency term; when experience_attention enables adaptation, citation/success feedback and exploration bonus modify ranking. Budget trimming drops memory entries from the end first.
• Invalidation or expiry: selection is per invocation; statistics persist in the in-process agent memory. Excerpting leaves source content unchanged; model conversation is rebuilt, not accumulated.
• Activation or effect: selection/use statistics from planning and action-success statistics from subsequent simulator execution feed fixed-rule ranking on later retrievals. Packet trimming then determines which selected excerpts reach the model. Actual dependence of subsequent model behavior and performance is unobserved.
• Evidence limits: cited-parent success is an association, not an intervention; ranking neither certifies truth nor preserves all conditions in truncated text. Full stored record provenance varies by producer.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Target/domain: retrieval relevance and heuristic past-use feedback, not propositional truth. Software chooses ranking and omissions; controller supplies citations; simulator supplies action success. No expected-answer oracle. Force: ranking at retriever, enforcing omission at packet builder, advisory at model; horizon later decisions within the agent's retained-state lifetime. Direct behavior/policy adaptation is the ranking-state update across planning and subsequent executor feedback, consumed by later retrieval. The inspected CLI commits planning and takes queued actions before simulation.step(actions); _record_action_outcome then calls mark_causal_outcome, so outcome credit is not a planning-stage result. SRC-1, src/biofoundry/cli.py, src/biofoundry/simulation.py, src/biofoundry/memory.py. No improved capacity or exercised self-improvement established: objective-aligned determination and causal dependence of subsequent live behavior are not demonstrated. Underlying model parameter fixity is separate.

Concrete deterministic transformation (implementation deduction, no execution): a non-JSON record Yield is high only after heating at max_characters=38 has length 32 and is returned intact. The same record at budget 25 takes the non-JSON truncation branch: the retained prefix has one character, so output is Y...[bounded excerpt]. The original necessary condition is no longer conveyed. This branch is therefore not guaranteed non-ampliative semantic reshaping; for arbitrary records the truth-apt relation is indeterminate. Structured artifact-measurement excerpts also project out service details, keeping artifact ID, program ID and performance before any final clipping. The remaining classifications require actual input/output propositions and their interpretation, not merely a text label.

return content[: max(0, max_characters - 24)] + "...[bounded excerpt]" --- src/biofoundry/memory.py

feedback = 0.5 * citation_rate + 0.5 * ( causal_success_rate if outcome_attempts else 0.0 ) --- src/biofoundry/memory.py

self.memories[index].mark_causal_outcome( set(map(str, action.causal_parents)), success=success ) --- src/biofoundry/simulation.py

The model-planning container also filters proposed research-state evidence IDs by retained availability before storage. The predicate checks resolvability, not whether a cited item supports the hypothesis; this bounds the lineage license of the retained text.

retained_evidence = { str(record_id) for record_id in research_state.get("evidence_ids", []) if simulation.memories[index].get(str(record_id)) is not None or str(record_id) in simulation.archive.by_id or simulation.program_library.knows(agent_id, str(record_id)) } --- src/biofoundry/policies/llm.py

Claims

EPI-CLM-research-environment

Label: Scientific research environment

Claim: SwarmWorld offers a deterministic environment for studying model societies discovering, testing, exchanging, inheriting and embodying technology. SRC-1, README.md; doctrine/design. Distinct from RT-CLM-grounded-science, which is a model instruction against invented observations; this claim concerns the environment's scope. It does not itself assert accepted explanatory truth or physical-world validity.

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. --- README.md

Behavioral-authority paths

none declared in this member