Autogenesis epistemic report

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

Source-and-claim boundary

The question is which routes acquire or produce truth-apt content, evaluate it, license reliance and change later behavior. Scope and exclusions are those of boundary.md; evidence identities and layers are in its Source register. All new reads below are SRC-1 at the frozen commit. Executable code supplies implementation evidence; prompts, docstrings describing intended guarantees and README claims supply doctrine/design. No target execution, observed-run source or causal comparison was supplied or performed.

The material boundary includes user tasks and action results, child returns, web acquisition/synthesis, tiered execution-history memory, corpora and ranking, evaluator reports, benchmark scoring, workflow verification/evaluation records, extension admission and withdrawal, and administrative changes. Remote provider internals, other repositories and live services remain excluded. Their exclusion prevents identifying provider reasoning, source reliability and deployed behavior. Available but uninspected leaves are not absent routes. In particular, this is not a system-complete negative about knowledge production.

Claims assessed are RT-CLM-self-evolution, RT-CLM-user-approval, EPI-CLM-continuous-improvement and EPI-CLM-smoke-safety. Their declared meanings are not copied here.

Entry point or operation	Source path in SRC-1	Coverage and records	Limit / conclusion prevented
Console command, TUI, serve	autogenesis/cli.py	Direct read of entry/dispatch classes; supplied Gateway account: RT-RTE-task-execution	No alternate console answer oracle found in this read; complete control-command internals not independently traced.
Model decisions, children, skills, native roster, playground	autogenesis/agent/native_tools.py, autogenesis/agent/types.py, autogenesis/protocol/server.py, autogenesis/gateway/service.py	Direct roster read; other mechanisms supplied: RT-RTE-child-delegation, RT-RTE-skill-instructions, RT-RTE-model-playground	No provider weights, exact deployed model or decisions; content classification of unseen completions remains indeterminate.
Web fetch, search, summary and merge	autogenesis/tool/default/web_fetcher.py, autogenesis/tool/default/web_searcher.py	Direct fetch handler and search output/summary/merge methods: EPI-RTE-web-reports	Search-engine adapters and URL extraction internals uninspected; cannot establish faithful original-page extraction or exhaustive retrieval.
History ingestion, compaction, context delivery	autogenesis/memory/default/tiered.py, autogenesis/hook/default/compact.py	Direct emit/get, todo summary, compaction, persistence, session creation/eviction and compact hook: EPI-RTE-history-summary	Full trace producer, renderer and all alternative memory implementations uninspected; cannot establish complete capture or universal restart recall.
Corpus ingestion and retrieval	autogenesis/knowledge/server.py, autogenesis/knowledge/default/rankers.py	Both files read: RT-RTE-knowledge-admission, EPI-RTE-corpus-ranking	External ranker-library implementations uninspected; ranking relevance is not truth verification.
Tool and agent evaluators	autogenesis/agent/evaluator/tool_evaluate_agent.py, autogenesis/agent/evaluator/agent_evaluate_agent.py, autogenesis/prompt/default/tool_evaluate_agent.html	Direct classes and tool rubric: EPI-RTE-evaluator-reports	Other evaluator prompts/classes except workflow evaluator remain uninspected; no universal evaluator-quality conclusion.
Workflow verification, benchmark, result delivery	autogenesis/workflow/runtime.py, autogenesis/benchmark/server.py, autogenesis/benchmark/default/exact_match.py	Direct VERIFY, benchmark dispatch and normalization, batch scoring, exact match: EPI-RTE-workflow-checks	Dataset-bound benchmarks, canvas compiler and downstream user programs uninspected; no general answer correctness or causal component attribution.
Process record transformations	autogenesis/process/default/records.py	Entire file read: EPI-RTE-record-transforms	Text/table processors, dataset adapters and scientific computations uninspected; prevents a complete derivation inventory.
Workflow evaluation write, summary, inspection and reload	autogenesis/tool/default/evolution.py, autogenesis/agent/evaluator/workflow_evaluate_agent.py, autogenesis/workflow/context.py, autogenesis/tool/default/inspect_workflow.py	Direct tool/class, context initialization and evaluation methods, inspect consumer: EPI-RTE-workflow-evidence	File save/load bodies located but not fully read; atomicity not assessed. No actual evaluation artifact or subsequent decision supplied.
Capability registration, smoke, rollback, unload	autogenesis/extension/smoke_gate.py, autogenesis/tool/default/evolution.py, autogenesis/extension/server.py, autogenesis/hook/default/tool_registration.py	Direct smoke and evolution tool; supplied registration/promotion: RT-RTE-capability-admission	General capability effects and rollback internals supplied, not re-audited. No evidence that rollback reverses arbitrary effects.
Runtime configuration and shell	autogenesis/gateway/service.py, autogenesis/tool/default/bash.py	Supplied RT-RTE-runtime-configuration, RT-RTE-shell-execution	Administrative content can change decisions outside evolution checks; command success is not a content oracle.
Browser/reviewer, connectors, environments, science/kernel, canvas, other data/process/benchmark leaves	autogenesis/, extension/, examples/, frontend/	Included functional families; only supplied endpoint coverage in RT-RTE-workflow-execution and boundary	Detailed evaluators, action effects and report consumers unassessed. Prevents system-wide absence of acceptance, derivation or theory building.

Missing execution evidence prevents saying any candidate was actually checked, accepted, integrated or beneficial. Missing concrete model input/output pairs prevent distinguishing semantic compression, entailment and ampliative conjecture for particular reports. Missing independent source/reference warrant prevents upgrading imported text or numeric comparisons into established external facts. These limits do not erase implemented checking and behavioral-authority paths.

Epistemic-object inventory

Fields owned by existing records are reached through see, not redeclared.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-meta-agent, RT-OBJ-self-evolving	Diagnostic/method claims about when capability evolution is appropriate	see records	see records	Instructional imperatives are not themselves truth-apt; method applicability claims may be. Delivery is not compliance.
RT-CMP-model-manager	Not individually recoverable from weights	see record	see record	Provider-parametric state inaccessible; local calls do not establish parameter training.
EPI-OBJ-web-report	Page assertions, query summaries and cited synthesis	see record	see record	Original-source warrant unknown; citations do not check entailment.
EPI-OBJ-tiered-history	Claims that actions ran, failed, returned paths/results; summaries of those claims	see record	see record	Success labels are operational. Summary fidelity unverified.
EPI-OBJ-corpus-records	Caller-supplied text and metadata	see record	see record	Content lineage depends on caller; score may overwrite caller field.
EPI-OBJ-evaluation-report	Tool/agent quality and failure assertions, scores, suggestions	see record	see record	Rubric requests evidence but does not establish it was gathered.
EPI-OBJ-workflow-evaluation	Version-scoped run success, quality and aggregate health assertions	see record	see record	Run identity checks support lineage, not task correctness or independent quality judgment.
EPI-OBJ-record-values	Source records, derived returns and reference-relative correctness scores	see record	see record	Formal/numeric relation has conditional warrant; input fact truth remains unestablished.
RT-OBJ-extension-source	Executable policy: none necessarily; embedded descriptions may assert capability behavior	see record	see record	Code behavior change and accepted explanatory knowledge are different outcomes.
RT-OBJ-extension-manifest, RT-OBJ-workflow-checkpoint	Active/version/run-status assertions	see records	see records	Recovery identity licenses execution/reuse, not correctness of output.

Authority-route ledger

A repeated route citation identifies a different function of the same source-native route. Shared-record fields, evidence and activation conditions are owned by that record (see). All implemented rows describe inspected capacity, not observed operation. No row licenses an unqualified truth verdict.

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
RT-RTE-task-execution	content transformation	implemented	RT-CMP-model-manager	truth-apt transformation: indeterminate	Task/context to model answer or action	see record	see record	Candidate answer or instructions	see record	No generic answer oracle	see record	see record and RT-BAP-prompt-guidance	see record	none	lifecycle-success-not-truth	No concrete completion to classify; remote inference opaque.
EPI-RTE-web-reports	content transformation	implemented	EPI-OBJ-web-report	truth-apt transformation: acquisition/import	Page/search material enters fetched results	Fetch success, not source verification	see record	Text with source URLs	Advisory delivery	Source warrant unknown	Calling model can use returned message	see record	see record	none	none	Transport/extraction fidelity unassessed.
EPI-RTE-web-reports	content transformation	implemented	EPI-OBJ-web-report	truth-apt transformation: indeterminate	Truncated page to summary to merged research report	Model; no semantic-preservation checker	see record	Cited report or mechanical fallback	Advisory	Candidate synthesis, no established acceptance	Later foreground model receives report, not full original pages	see record	see record	none	citations-not-warrant	Could reshape, derive or conjecture; concrete pairs needed.
EPI-RTE-history-summary	content transformation	implemented	EPI-OBJ-tiered-history	truth-apt transformation: indeterminate	Event records to short summary	CompactHook model; success/nonempty test only	Overflow, after removal of old chunk	Summary or restored chunk on empty result	Context replacement	No semantic truth check	Old verbatim records can leave prompt-visible recent history	see record	see record	none	none	Exact values are requested, not verified.
EPI-RTE-history-summary	retention	implemented	EPI-OBJ-tiered-history	no content change	Summary enters working deque and later get	Bounded queue/window selectors	see record	Summary delivered in later context	Advisory context	Retained candidate, not accepted claim	Calling model can rely on summary	see record	see record	none	retention-before-acceptance	No direct report reload in inspected get/session creation.
RT-RTE-knowledge-admission	retention	implemented	EPI-OBJ-corpus-records	truth-apt transformation: acquisition/import	Supplied documents appended to corpus	Backend-name and I/O contract	see record	Persistent normalized documents	Storage admission	Source warrant unknown; no truth criterion	Later retrieval can select text	see record	see record	none	ingestion-not-acceptance	No named proposition accepted for a defined use.
EPI-RTE-corpus-ranking	operational admission/selection/consumption	implemented	EPI-OBJ-corpus-records	truth-apt transformation: non-ampliative reshaping	Query ranks retained documents	BM25 lexical score or fitted TF-IDF/cosine	On retrieval	Top-k records with scores	Ranking, not enforcing truth	Similarity ordering only	Workflow/data consumer gets selected subset	see record	see record	none	none	Model injection depends on subsequent caller.
EPI-RTE-evaluator-reports	check/evidence production	implemented	EPI-OBJ-evaluation-report	truth-apt transformation: indeterminate	Named tool/agent behavior and source quality	Model agent, requested source/direct cases and rubric	Caller-requested evaluation	Scored report, partial report or failure	Advisory	Potential evidence about tested cases only	Parent can request revision/rollback	see record	see record	RT-CLM-self-evolution	rubric-not-independent-oracle	Expected answers chosen by evaluating model; no observed case/check.
EPI-RTE-workflow-checks	check/evidence production	implemented	EPI-OBJ-record-values	truth-apt transformation: entailed derivation	Prediction versus caller ground truth	ExactMatchBenchmark numeric tolerance/string equality	Benchmark step/batch	Per-task score and mean	Data result, not adoption gate	Equality within encoded comparison domain	Scope outputs can drive user-authored next steps	see record	see record	none	reference-truth-unchecked	Does not prove reference, encoding or producing mechanism.
EPI-RTE-workflow-checks	check/evidence production	implemented	EPI-OBJ-evaluation-report	truth-apt transformation: indeterminate	VERIFY item	min_votes model-agent calls	VERIFY step	One verdict or verdict list	Returned evidence only	No consensus aggregation or truth license from call count	Workflow receives results for separately authored logic	see record	see record	none	verify-name-not-acceptance	Repeated calls need not be independent.
EPI-RTE-record-transforms	content transformation	implemented	EPI-OBJ-record-values	truth-apt transformation: non-ampliative reshaping	Record fields/order/subset/evaluation envelope	Fixed Python transformations	Process step	Projected, renamed, filtered or shaped records	Data output	Preservation limited to retained fields/records	Downstream pipeline consumes transformed records	see record	see record	none	lineage-can-be-dropped	Projection can discard provenance.
EPI-RTE-record-transforms	content transformation	implemented	EPI-OBJ-record-values	truth-apt transformation: entailed derivation	Numeric consecutive price values to return	Python arithmetic under field/order premises	derive_return call	New return column	Data output	Conditional arithmetic relation, not market explanation	Pipeline can use return values	see record	see record	none	none	Missing/non-numeric values and zero predecessor yield null; input truth not warranted.
EPI-RTE-workflow-evidence	check/evidence production	implemented	EPI-OBJ-workflow-evaluation	no content change	Evaluation identity, terminal state, success consistency	Code against registered version/hash and retained run	Before recording	Reject or normalized evaluation	Enforcing evidence admissibility	Recorded-run linkage and metric provenance only	Invalid evidence write blocked	see record	see record	none	none	Quality remains caller judgment; static negative needs only case ID.
EPI-RTE-workflow-evidence	disposition/acceptance	implemented	EPI-OBJ-workflow-evaluation	no content change	Evaluation record admission, not workflow quality acceptance	Same validation results	Before append	Evaluation retained or exception	Enforcing write gate	Acceptance as tracking evidence under lineage criterion	Admits records, not permission to execute workflow	see record	see record	none	evidence-admission-not-product-acceptance	No evidence-consuming adoption decision established.
EPI-RTE-workflow-evidence	retention	implemented	EPI-OBJ-workflow-evaluation	no content change	Version-linked evaluation append/save/reload	Context manager	see record	Persisted evaluations; current-version readback	Tracking	Evidence ledger, not accepted theory	inspect_workflow can expose current evidence	see record	see record	none	none	Old-version records filtered from current view.
EPI-RTE-workflow-evidence	check/evidence production	implemented	EPI-OBJ-workflow-evaluation	truth-apt transformation: entailed derivation	Current-version aggregates and healthy flag	At least 3 distinct cases, success rate ≥0.8, average quality ≥0.7	Summary/inspect call	Boolean and aggregate values	Advisory summary	Threshold satisfaction over submitted evidence	Foreground agent may keep/optimize/rollback	see record	see record	RT-CLM-self-evolution	threshold-not-mandatory-adoption	No automatic keep/rollback consumer found in inspected path.
RT-RTE-capability-admission	check/evidence production	implemented	RT-OBJ-extension-source	no content change	Generic agent loop health after registration	Response.success and no constraint stop; not candidate exercise	Optional default-on post-registration smoke	Probe success/failure	Feeds rejecting consumer	Synthetic loop completion only	Failed check triggers best-effort revert/unload	see record	see record	EPI-CLM-smoke-safety	candidate-not-required	Does not compare exact OK or prove benefit/safety.
RT-RTE-capability-admission	disposition/acceptance	implemented	RT-OBJ-extension-source	non-truth-apt policy/content update: executable capability admission	Proposed file to live registry	Path/syntax/overwrite/registration, optional smoke	Automatic generator hook or Gateway promotion	Install or reject/recover	Permissive then rejecting	Operational admissibility, not epistemic endorsement	Code may execute before probe; future callers use active version	see record and RT-BAP-extension-execution	see record	RT-CLM-user-approval	approval-not-invariant	Arbitrary import/external effects not undone by archive restore.
RT-RTE-capability-admission	retention	implemented	RT-OBJ-extension-manifest	no content change	Active files/manifests/archive	see record	see record	Reloadable installed version	Executable persistence	No accepted ampliative claim established	Startup and later dispatch load retained capability	see record	see record	EPI-CLM-continuous-improvement	installation-not-benefit	Later use and improvement unobserved.
RT-RTE-capability-admission	behavior/policy adaptation	doctrine only	RT-OBJ-meta-agent	non-truth-apt policy/content update: keep, optimize or withdraw using evaluation	Evaluator evidence to choice of capability version	MetaAgent prose policy; human/model compliance unobserved	After evaluator/optional comparison	Requested keep/rollback/optimize	Advisory instruction to model	No independent transfer/mechanism license	Model can use evolution_tool rollback/unload	see record and EPI-RTE-workflow-evidence	see record	RT-CLM-self-evolution	requested-not-exercised	Standing machinery exists; actual evaluation-shaped choice absent from evidence.
RT-RTE-runtime-configuration	operational admission/selection/consumption	implemented	RT-CMP-model-manager	non-truth-apt policy/content update: endpoint and instruction configuration	Caller edit to active configuration	Field/type/provider validation	see record	Updated configuration	Enforcing schema, permissive update	No quality comparison	Later requests may use changed endpoint/skill	see record	see record	none	alternate-admission	Human authorship unresolved; restart persistence uninspected.
RT-RTE-workflow-execution	lineage/freshness/recovery	implemented	RT-OBJ-workflow-checkpoint	no content change	Resume identity and completed invocation reuse	Version/hash/runtime checks	On resume	Reject, reuse completed record, or requeue incomplete work	Enforcing execution identity	Identity/cache eligibility only	Controls reexecution, not truth of cached output	see record	see record	none	recovery-not-revalidation	Partial external effects may repeat.

No lifecycle-integration row is claimed: inspected retention/operational use does not by itself establish post-acceptance integration of a truth-apt ampliative claim. That is a limit of these routes, not an absence conclusion for all shipped operations.

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
RT-CLM-self-evolution	see record	see record	see record	RT-RTE-capability-admission, EPI-RTE-evaluator-reports, EPI-RTE-workflow-evidence	None registered	No before/after or component intervention	Improvement-directed change, evaluators, evidence ledger and withdrawal affordances	Does not establish exercised evidence-responsive improvement.
RT-CLM-user-approval	see record	see record	see record	RT-RTE-capability-admission	None registered	Automatic tool-hook route is counterexample to universal approval enforcement	Prompt policy, not invariant	Per-content user answer not consumed on that path.
EPI-CLM-continuous-improvement	see record	see record	README claim; evolution/evaluation guidance	RT-RTE-capability-admission, EPI-RTE-workflow-evidence	None registered	Quality scores and installed changes do not identify benefit	Implemented capability ecosystem change path	Continuous beneficial operation unestablished.
EPI-CLM-smoke-safety	see record	see record	Smoke module's safety-net description	RT-RTE-capability-admission	None registered	General agent can finish without invoking candidate; probe runs after registration	Bounded generic-loop health check with rejecting consumer	Not candidate correctness, import sandbox or comprehensive first-use failure detection.

Bounded conclusion

Autogenesis has several independent routes, not one epistemic evaluator. Web pages and corpus documents are acquired with unknown source warrant. Search reports and history compaction request factual preservation but do not check semantic fidelity. Their later model-facing messages can affect decisions without a truth-apt acceptance step. Corpus ranking selects lexical/vector similarity; it neither endorses truth nor by itself demonstrates model reliance. These findings are bounded by EPI-RTE-web-reports, EPI-RTE-history-summary, EPI-RTE-corpus-ranking and RT-RTE-knowledge-admission.

The deterministic record and exact-match routes support conditional deductions. A numeric return follows from numeric predecessors under the inspected arithmetic and ordering convention; a benchmark score follows from a prediction/reference pair under the comparator. Neither warrants source prices, reference answers, explanatory mechanism or transfer. EPI-RTE-record-transforms and EPI-RTE-workflow-checks retain concrete source-level examples. No target run is needed for these implementation deductions.

Model evaluators request tests and scored criticism, while VERIFY collects agent verdicts without deciding a consensus. Those are implemented evidence-production paths, not evidence of observed correct reports. The workflow evidence ledger is stronger on lineage than a free-form score: registered identity/hash, terminal state, recorded success and duplicate-run checks constrain successful submissions. Its healthy flag is a reproducible aggregate over submitted records, but quality and case identity still partly originate with the caller. Inspection returns the evidence to agents; keep/rollback remains requested decision policy. This newly inspected route supplements, rather than supersedes, the valid runtime container RT-RTE-workflow-execution.

Operational change admission is clear even where epistemic acceptance is not. Registration grants immediate executable authority and durable future dispatch; smoke may subsequently withdraw it. Smoke acceptance means generic-loop completion under its flags, not verified use, safety or effectiveness of the proposed capability. Rollback/unload are implemented intervention affordances; choosing them because a particular criticism was correct remains unobserved. There is no evidenced acceptance of an ampliative truth claim against a named criterion for an intended use, followed by lifecycle integration, on the assessed routes. Uninspected scientific, browser, reviewer and other evaluator leaves prevent extending that statement to the whole system.

For the capability ecosystem across later dispatches/restarts, the task-completion objective is independent of the update. RT-RTE-capability-admission establishes a standing improvement-directed self-change path and live consumers. No supplied episode establishes that objective-bearing evidence selected this particular update or that subsequent operation causally depended on it; successful self-improvement and improved future capacity remain unestablished. Reflection, autonomous theory building and each theory-builder condition retain the runtime's separate evidence limits; a scored report or retained healthy flag does not fill missing content-directed criticism and iterative use. Provider weights and internal updates remain inaccessible; the local model configuration can resolve another model without demonstrating local parameter learning.

Shared records

Operative objects

EPI-OBJ-web-report

Label: Web source text and research report

Source-native identity: WebFetcherTool formatted page text, WebSearcherTool per-result summary and merged_summary. Natural-language assertions in transient dictionaries and Response.message; URLs/titles are symbolic lineage fields. Distinct from RT-OBJ-meta-agent, which is governing instruction, not acquired evidence. No counterpart among supplied objects. SRC-1: autogenesis/tool/default/web_fetcher.py, autogenesis/tool/default/web_searcher.py (implementation).

formatted = f"Title: {title}\nContent: {markdown}" --- autogenesis/tool/default/web_fetcher.py

Content: {raw_content[:self.max_length]} --- autogenesis/tool/default/web_searcher.py

EPI-OBJ-tiered-history

Label: Tiered history records and working summaries

Source-native identity: TieredMemory state.recent, state.working and get result. Natural-language event assertions and model summaries in RAM deques; timestamps/status/session keys are symbolic selectors. Persistence serializes through subclass renderers to per-session files. Distinct from RT-OBJ-workflow-checkpoint, which controls workflow replay rather than model history. No supplied object counterpart. SRC-1: autogenesis/memory/default/tiered.py (implementation).

lines.append("## Working Memory") lines += [f"- {s}" for s in working] --- autogenesis/memory/default/tiered.py

EPI-OBJ-corpus-records

Label: Knowledge corpus records

Source-native identity: corpus.jsonl documents and retrieval records; natural-language text plus symbolic metadata/score/backend fields, on disk and in RAM. This is the material retained by RT-RTE-knowledge-admission, which is a route, not a duplicate object. Source warrant is not established by coercion; dictionaries retain fields but retrieval overwrites any existing score. SRC-1: autogenesis/knowledge/server.py (implementation).

results = [{**docs[index], "score": score} for index, score in hits if 0 <= index < len(docs)] --- autogenesis/knowledge/server.py

EPI-OBJ-evaluation-report

Label: Model evaluator reports

Source-native identity: AgentEvaluateAgent/ToolEvaluateAgent Response.message and workflow VERIFY verdicts. Natural-language judgments and localized score labels interpreted by a later model; carried in transient Responses and any retained action history. Distinct from governing RT-OBJ-self-evolving; no supplied report object counterpart. SRC-1: autogenesis/agent/evaluator/agent_evaluate_agent.py, autogenesis/agent/evaluator/tool_evaluate_agent.py, autogenesis/prompt/default/tool_evaluate_agent.html, autogenesis/workflow/runtime.py (implementation; rubric doctrine/design).

• Read the structured results and assign per-dimension scores using the rubrics in <evaluation-rules>.
• Call done_tool with the complete report. --- autogenesis/prompt/default/tool_evaluate_agent.html

EPI-OBJ-workflow-evaluation

Label: Version-scoped workflow evaluations

Source-native identity: WorkflowEvaluation, _evaluations and evaluation_summary. Symbolic version/run/hash/success/quality/metric records and aggregates in RAM and evaluations.json; notes are natural-language explanations, with no inspected content-directed criticism instance. Distinct from RT-OBJ-workflow-checkpoint: it records evaluative evidence about runs, not execution resumption state. SRC-1: autogenesis/workflow/context.py, autogenesis/tool/default/evolution.py (implementation).

self.evaluation_path = str(Path(evaluation_path or os.path.join(self.base_dir, "evaluations.json"))) --- autogenesis/workflow/context.py

EPI-OBJ-record-values

Label: Pipeline records and comparator scores

Source-native identity: process record dictionaries, derive_return output column, ExactMatchBenchmark task.score and batch mean. Symbolic typed fields/numbers and fixed consumer rules, in transient Python lists/Task objects and workflow scope; not distributed model parameters. Different from RT-OBJ-workflow-checkpoint, though pipeline results can enter a checkpoint. No supplied record-value counterpart. SRC-1: autogenesis/process/default/records.py, autogenesis/benchmark/default/exact_match.py, autogenesis/benchmark/server.py (implementation).

task.score = score self._tasks.append(task) --- autogenesis/benchmark/default/exact_match.py

Routes

EPI-RTE-web-reports

Label: Web acquisition and report synthesis

Distinct from RT-RTE-task-execution: this is a callable tool's internal content pipeline, not agent scheduling. Owner: WebFetcherTool/WebSearcherTool. URL/query → fetch/search results → optional page summary → optional merged report → Response.message. Public flags govern fetch/summary/merge. Source text is truncated before model summarization; output includes a reference list but no entailment check. Models propose text; nonempty output admits the summary/report; exceptions use placeholders or mechanical fallback. No per-content human decision is consumed; caller identity unresolved. Open query mode, not a controlled experiment. SRC-1: autogenesis/tool/default/web_fetcher.py, autogenesis/tool/default/web_searcher.py (implementation).

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Formatted fetched page or search/report text in message and query/count/merged-summary metadata in data.
• Later read-back: uninspected — this tool does not itself establish a durable store or later model call; action-memory retention can retain its message.
• Delegated visibility: Foreground caller gets synthesized report or fallback source listing; original raw_content is not returned as a complete records array in the inspected search Response.data.
• Selection predicate: Query/URL, search-engine results and enabled fetch/summary/merge flags; each page summary receives only max_length characters.
• Invalidation or expiry: Transient result dictionaries leave scope after return; no per-source withdrawal or durable revision handler here.
• Activation or effect: Message supplies candidate evidence to caller; compliance, reliance and benefit unobserved.
• Evidence limits: Extraction/search adapters and source reliability uninspected; requested factuality/citation is not checked preservation or acceptance.

if response and response.message.strip(): result["summary"] = response.message.strip() --- autogenesis/tool/default/web_searcher.py

if merged_summary: output_message = merged_summary --- autogenesis/tool/default/web_searcher.py

EPI-RTE-history-summary

Label: Trace history compaction and later context

Distinct from RT-RTE-task-execution: internal retained-context transformation rather than task lifecycle. TieredMemory ingests action/agent events, stores event.message or error and status, removes old recent chunks on overflow, calls CompactHook, and appends returned summaries to bounded working memory. get selects recent tail and working summaries. Session creation constructs fresh state, not a file loader. Long todo descriptions also use model summarization with length truncation; todo content's downstream authority is not established by this get path. SRC-1: autogenesis/memory/default/tiered.py, autogenesis/hook/default/compact.py (implementation).

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: emit mutates memory; CompactHook returns text; get returns selected Markdown or None.
• Later read-back: get returns last working_fetch summaries and recent tail by session ID; the agent consumer is supplied in RT-RTE-task-execution. Summaries are retained between model invocations; disk persistence is separate from restart recall.
• Delegated visibility: Context ID selects separate session history; no automatic parent/child history union in this path.
• Selection predicate: recent_max overflow triggers code-selected oldest compact_chunk records; latest working summary accompanies them; get applies working_fetch and short_term_n.
• Invalidation or expiry: Compaction removes chunk; empty summary restores it; bounded working deque and session eviction can remove prompt-reachable material. No semantic-validation rollback is present in inspected compaction.
• Activation or effect: Summary replaces verbatim history available to later context; model behavior change unobserved. Old retained file content is not automatically dereferenced by a path in a prompt.
• Evidence limits: Summary can reshape or add claims; absence of input/output examples prevents assigning semantic preservation or ampliative conjecture. This does not establish universal memory behavior.

chunk = [state.recent.popleft() for _ in range(k)] --- autogenesis/memory/default/tiered.py

if text: state.working.append(text) else: async with state._lock: # lossless: restore and stop state.recent.extendleft(reversed(chunk)) break --- autogenesis/memory/default/tiered.py

text = response.message.strip() if response.success else "" --- autogenesis/hook/default/compact.py

EPI-RTE-corpus-ranking

Label: Corpus backend ranking

Part of: RT-RTE-knowledge-admission

This part owns the backend ranking and returned-record transformation within the supplied ingest/retrieve route. BM25 lowercases/splits texts and query; TF-IDF fits current texts and cosine-ranks query. Numeric weights are freshly computed representations, not evidence of a trained persistent epistemic policy. Owning code calls external ranking libraries, whose internals are uninspected. SRC-1: autogenesis/knowledge/default/rankers.py, autogenesis/knowledge/server.py (implementation).

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Ranked index/score pairs become top-k corpus records in Response.data.
• Later read-back: Corpus is reread on retrieval; freshly ranked retained documents reach caller, not automatically the model.
• Delegated visibility: Workflow normalization preserves data records; Agent paths returning only message would need a separate bridge to convey records.
• Selection predicate: Query lexical overlap or TF-IDF cosine, descending score, top_k at least one; no minimum truth/relevance threshold.
• Invalidation or expiry: Corpus/backend manifest changes affect next retrieval; no persistent index/update policy in these backend methods.
• Activation or effect: Ranking selects and orders what caller sees; does not establish changed model behavior or source endorsement.
• Evidence limits: Derived similarity is a retrieval signal; input provenance and external library correctness remain unestablished.

matrix = vectorizer.fit_transform([str(text) for text in texts]) query_vector = vectorizer.transform([str(query)]) sims = cosine_similarity(query_vector, matrix)[0] --- autogenesis/knowledge/default/rankers.py

EPI-RTE-evaluator-reports

Label: Tool and agent quality reports

Different from RT-RTE-capability-admission: produces advisory evidence rather than installs a capability. AgentEvaluateAgent and ToolEvaluateAgent delegate unchanged to base loop. Tool rubric requests known valid/invalid cases, interface inspection, source review and qualitative performance checks. Target is a named tool's behavior on the selected cases and source, not universal task success. Model chooses tests, expected answers and scores; code enforces ordinary routing/budgets, not rubric validity. Native roster respects allowlists, so prompt assertion that any registered target is callable is conditional on exposure. No mandatory human evaluator or independent oracle is supplied. Open task-directed evaluation; expected values are requested, not independently authenticated. SRC-1: autogenesis/agent/evaluator/tool_evaluate_agent.py, autogenesis/agent/evaluator/agent_evaluate_agent.py, autogenesis/prompt/default/tool_evaluate_agent.html, autogenesis/agent/native_tools.py (implementation and rubric doctrine/design).

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Ordinary evaluator Response, intended to carry scored report; no registrable artifact.
• Later read-back: uninspected — evaluator classes do not add a durable report ledger; caller/action memory may retain response text.
• Delegated visibility: Parent receives evaluator message under the supplied child-return route; not necessarily all structured Response.data or raw test evidence.
• Selection predicate: Caller names target and cases; registered evaluator, exposed tools, max steps and model decisions govern actual work.
• Invalidation or expiry: Model may revise report within task; no automatic score-triggered rejection/expiry of capability in these classes.
• Activation or effect: Parent can read report and request optimization/rollback; no observed evaluation-consumption episode.
• Evidence limits: Rubric is not a measured oracle. Self-selected expected answers, omitted cases and source-only performance assessment limit warrant.

return await super().call(task=task, files=files, ctx=ctx, **kwargs) --- autogenesis/agent/evaluator/tool_evaluate_agent.py

• Verify success=True, correct message, and correct data values (results live in response.data, a dict — NOT in extra).
• Score guide: 20 = all cases pass; −4 per failing case. --- autogenesis/prompt/default/tool_evaluate_agent.html

EPI-RTE-workflow-checks

Label: Workflow verdicts and benchmark checks

Part of: RT-RTE-workflow-execution

VERIFY invokes min_votes agent calls per item and returns verdicts; it does not compare or vote them into an acceptance decision. BENCHMARK invokes a named benchmark on normalized results; exact_match is the directly inspected comparator. Caller provides prediction/reference pairs; missing IDs and evaluation exceptions score zero in batch manager. A workflow benchmark Response succeeds even for score zero. _normalize blocks failed Responses but preserves data on successful ones, so data-dependent rejection needs separate authored control logic. No human per-item approval is read; caller origin unresolved. Bounded comparison mode, no verified improvement trigger. SRC-1: autogenesis/workflow/runtime.py, autogenesis/benchmark/server.py, autogenesis/benchmark/default/exact_match.py (implementation).

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Verdict(s), or benchmark mean/count in successful Response.data; workflow normalization preserves message/data/files.
• Later read-back: uninspected — checkpoint persistence belongs to the supplied outer route; comparator retains Task list only in process for stats, not an inspected cross-restart result ledger.
• Delegated visibility: VERIFY agents receive resolved item arguments; workflow gets verdict values, not automatically a causal explanation or complete trace.
• Selection predicate: Workflow step type/target/items/min_votes; benchmark task ID, prediction and ground_truth; exact match strips strings and tries floats before string equality.
• Invalidation or expiry: Batch errors map to zero; exact-match reset clears task list; failed Response normalization raises, but zero score alone does not reject workflow.
• Activation or effect: Result is available to downstream scope references; no automatic truth acceptance or policy update in these branches.
• Evidence limits: Reference warrant, case selection and model-verdict independence unknown. Full dataset-specific evaluators uninspected.

Concrete source deduction, not target execution: prediction " 2 " and ground_truth "2.0" strip to 2 and 2.0; float difference is zero, so exact_match assigns 1.0. This proves agreement under the comparator, not the factual truth of either string. Conversely two missing/empty values do not pass because pred must be nonempty.

if pred: try: score = 1.0 if abs(float(pred) - float(gt)) < 1e-9 else 0.0 except (TypeError, ValueError): score = 1.0 if pred == gt else 0.0 --- autogenesis/benchmark/default/exact_match.py

return verdicts[0] if step.min_votes == 1 else {"item": item, "verdicts": verdicts} --- autogenesis/workflow/runtime.py

return {"message": value.message, "data": value.data, "files": value.files} --- autogenesis/workflow/runtime.py

EPI-RTE-record-transforms

Label: Record reshaping and arithmetic derivation

Part of: RT-RTE-workflow-execution

Process manager calls can invoke registered record processors. The directly read file implements projection, head, rename, filtering, sort, evaluation-envelope mapping and return derivation. Fixed Python rules, not model judgment, select rows/fields and derive numbers. Caller specifies fields/mapping/order; no evidence/human acceptance decision is consumed. These functions are not truth checks or improvement selection. SRC-1: autogenesis/process/default/records.py; outer dispatch supplied by runtime (implementation).

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Response.data records/count or argument failure.
• Later read-back: inapplicable — processors themselves retain no updated artifact; workflow output/checkpoint consumers are separate.
• Delegated visibility: Downstream workflow sees normalized data; projections may remove provenance while preserving selected values.
• Selection predicate: Configured fields, filters, sort key, count, evaluation mapping or arithmetic source field.
• Invalidation or expiry: Per-call output only; invalid required fields/operators reject; no rollback or endorsement history.
• Activation or effect: Produces transformed values available to downstream pipeline; no observed use or benefit.
• Evidence limits: Return interpretation assumes chronological numeric prices. In missing-value cases prev tracks previous numeric value, not necessarily the immediately preceding row.

Concrete source deductions: select_fields on [{"text":"A","source":"S"}] with fields ["text"] yields [{"text":"A"}], retaining text but dropping its source. derive_return on [{"close":100},{"close":110}] produces returns null and 0.1. The second is a new arithmetic proposition conditional on supplied values, not ampliative market knowledge. These are implementation deductions, not observed executions.

out = [{k: row.get(k) for k in keep} for row in rows if isinstance(row, dict)] --- autogenesis/process/default/records.py

if isinstance(current, (int, float)) and isinstance(prev, (int, float)) and prev != 0: ret = (current - prev) / prev out.append({**row, output: ret}) if isinstance(current, (int, float)): prev = current --- autogenesis/process/default/records.py

EPI-RTE-workflow-evidence

Label: Workflow evaluation evidence ledger

Part of: RT-RTE-workflow-execution

This is an evaluation-maintenance part of the workflow lifecycle, not a duplicate of runtime's executable invocation/recovery account. evolution_tool record_workflow_evaluation creates a WorkflowEvaluation from caller inputs. WorkflowContextManager checks active version, retained run identity/hash, terminal state and success agreement; it substitutes measured run token/time fields and rejects duplicate run IDs. A static failure needs case_id; successful evidence needs a real run. Caller supplies quality and notes. A read-only WorkflowEvaluateAgent explicitly allows this recording action and can expose the target workflow. No human veto or independent quality oracle is consumed. Evaluation mode is caller/model-selected cases; guideline thresholds target keep/optimize/rollback, not automatic promotion. Writes persist records, not a new capability. SRC-1: autogenesis/tool/default/evolution.py, autogenesis/workflow/context.py, autogenesis/agent/evaluator/workflow_evaluate_agent.py, autogenesis/tool/default/inspect_workflow.py (implementation).

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Stored evaluation and current summary from evolution_tool, or failed Response on exception.
• Later read-back: Context initialization loads evaluation records; current-version filtering feeds summary and inspect_workflow. inspect message includes aggregate summary and count; data additionally contains individual evaluations.
• Delegated visibility: Evaluating agent can record and inspect evidence; a parent receiving only inspect/evaluator message need not receive individual notes present only in data.
• Selection predicate: Active workflow/version, run/hash/state/success checks or static failure case_id, nonduplicate run; current-version records aggregated. Healthy requires 3 distinct cases, success rate at least 0.8 and average quality at least 0.7.
• Invalidation or expiry: Different active version filters old records out of evaluations(name); duplicate run rejected. Individual evidence rollback/deletion not established by inspected methods; capability rollback/unload is a separate evolution_tool action.
• Activation or effect: Successful record admission appends/saves and changes computed summary; inspect returns summary to model as advisory evidence. No mandatory automatic keep/rollback branch consumes healthy in the inspected runtime dispatch/evidence path.
• Evidence limits: Success agrees with recorded workflow lifecycle, not necessarily user-goal satisfaction. Quality and case labels are caller supplied; no supplied trace establishes an actual evidence-shaped change.

if evaluation.success != run.successful: raise ValueError("Evaluation success must match the recorded Workflow run") --- autogenesis/workflow/context.py

if evaluation.run_id and any(item.run_id == evaluation.run_id for item in existing): raise ValueError("This Workflow run has already been evaluated") existing.append(evaluation) self._save_evaluations() --- autogenesis/workflow/context.py

"healthy": len(distinct_cases) >= 3 and success_rate >= 0.8 and quality >= 0.7, --- autogenesis/workflow/context.py

f"- Evaluation Summary: {summary}\n\n" --- autogenesis/tool/default/inspect_workflow.py

Claims

EPI-CLM-continuous-improvement

Label: Continuous capability improvement

Claimed operation: ongoing generator/optimizer/evaluator activity improves tool, skill and agent ecosystem. Broader outcome wording than RT-CLM-self-evolution's task-serving change claim; the overlap is explicit, not an independent observed result. SRC-1: README.md (doctrine/design).

A self-evolving multi-agent framework. A MetaAgent orchestrates sub-agents to complete user tasks, while optimizer / evaluator / generator agents continuously improve the tool, skill, and agent ecosystem. --- README.md

EPI-CLM-smoke-safety

Label: Smoke first-use safety net

Claimed operation: synthetic replay catches evolved components that import but break the loop before manifest commitment. Distinct from RT-CLM-user-approval; both concern admission but different warrants. SRC-1: autogenesis/extension/smoke_gate.py (module description doctrine/design; probe body implementation). Runtime's admission record supplies counterevidence about manifest timing. The code's general-agent choice and success predicate do not require candidate invocation or exact OK equality.

A newly generated tool/agent/prompt that imports cleanly but blows up the loop (bad schema, exception on first use, runaway) is caught here before it is committed to the manifest. --- autogenesis/extension/smoke_gate.py

probe_agent_name = "general_agent" --- autogenesis/extension/smoke_gate.py

ok = bool(getattr(resp, "success", False)) and not data.get("stopped_by_constraint", False) --- autogenesis/extension/smoke_gate.py