Dynamic Cheatsheet epistemic report

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

Source-and-claim boundary

The question is whether and how Dynamic Cheatsheet acquires or produces truth-apt content, checks it, grants reliance, retains it for later use, and changes later behavior. The frozen boundary is the implementation and design material in the supplied Source register: SRC-1 at commit 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9 and SRC-2 for README.md. The caller, model provider, search services, supplied embeddings, external persistence, host process configuration and any enclosing application are excluded. The analysis treats outputs and prior answers as candidates whose correctness is unestablished. It does not infer operation from implementation or documentation.

The README makes claims about persistent, evolving memory (EPI-CLM-1), experience-driven learning without labels or human feedback (EPI-CLM-2), and benchmark performance gains (EPI-CLM-3). The implementation evidence supports within-call candidate generation, tag-based admission of a returned sheet, and caller-controlled reuse; it does not show epistemic acceptance or a correctness check. No execution trace, answer oracle, benchmark evidence, or causal experiment is supplied. The runtime records establish wired implementation paths, not that a model call or any later effect occurred.

Entry point or operation	Source path	Coverage	Exclusion, uninspected reason, or conclusion prevented
LanguageModel.advanced_generate and its default, cumulative, retrieval-synthesis, cumulative-retrieval, full-history, and dynamic-retrieval branches	dynamic_cheatsheet/language_model.py; dynamic_cheatsheet/utils/extractor.py	RT-RTE-1, RT-RTE-2, RT-RTE-3, RT-RTE-4, RT-RTE-5; EPI-OBJ-1, EPI-OBJ-2, EPI-OBJ-3, EPI-OBJ-4; source-inspected deductions	Providers, templates, embedding production, caller state and execution traces are outside the registered evidence; content correctness, source warrant, activation and benefit remain unknown.
LanguageModel.generate, local code-execution branch and provider tool dispatch	dynamic_cheatsheet/language_model.py; dynamic_cheatsheet/utils/execute_code.py; text_generation/simple_unified_client.py	RT-RTE-7, RT-RTE-8, RT-BAP-1, RT-BAP-2; EPI-OBJ-5, EPI-OBJ-6	The host and provider execution environments are uninspected; no execution, side effect, approval or rollback outcome is observed.
Direct UnifiedLLMClient.generate, generate_from_messages, and chat	text_generation/simple_unified_client.py	RT-RTE-6, RT-OBJ-3; Tavily and provider-native search source-inspected at SRC-1	Provider and search-service behavior is outside the boundary; no returned search result, source citation, model response or later use is observed.
README usage and design claims	README.md	EPI-CLM-1, EPI-CLM-2, EPI-CLM-3, SRC-2	Documentation establishes stated design and claims only, not operation, evaluation evidence or causal attribution.
CLI dispatch and hooks	No CLI or hook implementation path appears in the frozen Source register; README.md contains usage examples	Unassessed	Whether a separate CLI or hook route changes admission, authority or consumers cannot be determined; this prevents a system-complete claim about those entry points.
Embedding generation, caller persistence and external application consumers	Outside registered source paths and the stated boundary	Excluded	Their provenance, checks, storage, activation and effects cannot be attributed to this library.

The runtime report is the supplied account for the listed RT-* records. This member adds source-inspected direct-client web search because the runtime account does not cover that exposed operation. Its source wiring is described below; no provider operation is inferred. No separate hook is identified in the registered source scope, which is a scope limit rather than evidence of repository-wide absence.

Epistemic-object inventory

object ID	candidate truth-apt content or none	claimed role	evidence source ID and local anchor	gap/limit
EPI-OBJ-1	Returned generator answer text; truth-apt when it states a proposition over the query’s domain	Answer for the current query	SRC-1 dynamic_cheatsheet/language_model.py; extraction in dynamic_cheatsheet/utils/extractor.py	No answer oracle, evidence-consuming acceptance, provenance validation or execution trace.
RT-OBJ-1	Caller-supplied cumulative cheatsheet string; may contain claims, instructions or non-truth-apt material	Context reused by cumulative routes	SRC-1 dynamic_cheatsheet/language_model.py; SRC-2 README.md	The caller’s text and curator-template guidance are not registered; no retained sheet instance or correctness evidence is supplied. The model’s proposed replacement is separately inventoried as EPI-OBJ-8.
EPI-OBJ-2	Caller-supplied prior input strings; may state facts but their content is not established	Retrieval keys/context paired with earlier answers	SRC-1 dynamic_cheatsheet/language_model.py	Part of RT-OBJ-2; the inputs’ source, domain and warrant are unavailable.
EPI-OBJ-3	Caller-supplied prior generator outputs; truth-apt candidate answers may be present	Retrieved examples and input to later curation	SRC-1 dynamic_cheatsheet/language_model.py	Part of RT-OBJ-2; output provenance, correctness, acceptance and source warrant are unknown.
EPI-OBJ-4	Numeric embedding rows; no truth-apt proposition is identified in the vectors themselves	Similarity ranking signal that selects which paired examples reach a generator or curator	SRC-1 dynamic_cheatsheet/language_model.py	Part of RT-OBJ-2; producer, representation semantics, version, alignment and any validation are outside the boundary.
RT-OBJ-3	In-memory user/assistant message dictionaries, including candidate answer text on the direct chat path	Later prompt context within the client object	SRC-1 text_generation/simple_unified_client.py	The role and provenance of each caller-supplied message are not validated; no client-object trace is supplied.
EPI-OBJ-5	Model-proposed Python source; executable content can encode claims and instructions but is not itself shown to be true	Candidate operation admitted by the local flag/fence trigger or provider tool request	SRC-1 dynamic_cheatsheet/language_model.py; dynamic_cheatsheet/utils/execute_code.py; text_generation/simple_unified_client.py	Part of RT-OBJ-4; no execution result or host/provider execution evidence is supplied.
EPI-OBJ-6	Captured stdout/stderr or provider-returned execution text; truth-apt only insofar as it reports a named computation or environment state	Input appended to a later model generation, or returned to caller	SRC-1 dynamic_cheatsheet/utils/execute_code.py; text_generation/simple_unified_client.py	Part of RT-OBJ-4; no independent check, output trace, or evidence that a later answer relied on it.
EPI-OBJ-7	Search-result title, URL and content strings from Tavily; truth-apt claims may occur in snippets	Context prepended to a direct client query	SRC-1 text_generation/simple_unified_client.py	No counterpart among supplied operative-object records after comparison; the service response, source reliability and page contents are uninspected.
EPI-OBJ-8	Model response segment bounded by <cheatsheet> tags; candidate truth-apt content may be present	Proposed replacement for the current cumulative sheet	SRC-1 dynamic_cheatsheet/language_model.py; dynamic_cheatsheet/utils/extractor.py	Part of RT-OBJ-1 as the proposal stage of that sheet route; no trace, content-level criterion or reviewer.

Authority-route ledger

The table separates content generation and selection from operational admission. implemented means source inspection supports the wiring. No observed or causal status is assigned because the boundary supplies no execution trace or experiment.

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-1	content transformation	implemented	EPI-OBJ-1; current sheet RT-OBJ-1	truth-apt transformation: indeterminate — the model may restate supplied content, derive from it, or add claims; inspected output is unavailable to distinguish these	query plus current sheet to model response	Configured model proposes; no reference answer, evidence test or acceptance criterion is provided	Each selected generation call; operation unobserved	Candidate answer text	Returned as text to caller; no epistemic acceptance	None established beyond candidate status for this query	Caller may present or use the answer	Caller, returned text, advisory by interface, current call; later horizon depends on caller	SRC-1 dynamic_cheatsheet/language_model.py; see RT-RTE-1	EPI-CLM-1, EPI-CLM-2, EPI-CLM-3	none	No provider response, answer check, or consumer trace.
RT-RTE-1	disposition/acceptance	implemented	EPI-OBJ-8	truth-apt transformation: non-ampliative reshaping — code trims and extracts the tagged segment; it does not establish that the model’s content preserves or improves the sheet’s meaning	Presence of <cheatsheet> in response; old string is fallback	Program checks tag syntax only, not truth, entailment, evidential fit or intended use	Immediately after the curator response in each cumulative round	Tagged segment replaces the in-call value, otherwise the previous value is returned	Automatic admission to returned string and later rounds	No epistemic authority licensed; syntax is the entire check domain	Allows the extracted string into the next round and return value; caller separately decides cross-call reuse	Wrapper, prompt context, permissive string admission, later round in same call; later calls only if caller resubmits	SRC-1 dynamic_cheatsheet/utils/extractor.py, dynamic_cheatsheet/language_model.py; see RT-RTE-1	EPI-CLM-1, EPI-CLM-2	none	This is operational admission, not evidence-consuming acceptance for a stated use or scope.
RT-RTE-2	content transformation	implemented	EPI-OBJ-2, EPI-OBJ-3, EPI-OBJ-4; temporary sheet	truth-apt transformation: indeterminate — the curator may reorder, summarize, derive or add claims; no returned content is available to distinguish these	Selected input/output pairs and prior sheet to a curator response; tags bound returned segment	Similarity code selects examples; configured model proposes synthesis; neither is an answer oracle or content evaluator	Per retrieval-synthesis call before generation; operation unobserved	Query-specific candidate sheet	Delivered to current generator; not retained internally	No epistemic authority established for retrieved examples or synthesized claims	Changes the current generator prompt only; no durable write by the wrapper	Wrapper to generator, prompt context, permissive, current call	SRC-1 dynamic_cheatsheet/language_model.py; see RT-RTE-2	EPI-CLM-1, EPI-CLM-2	none	Embedding producer, pair alignment, prompt content, provenance and activation are uninspected.
RT-RTE-2, RT-RTE-3, RT-RTE-5	operational admission/selection/consumption	implemented	EPI-OBJ-2, EPI-OBJ-3, EPI-OBJ-4	no content change to selected strings; rank/order selects supplied pairs, without checking their claims	Current embedding against earlier supplied vectors; top-k indices select paired input/output rows	Program computes cosine similarity and sorts; no human veto, expected-answer oracle, or correctness evaluator	At each retrieval call; cumulative retrieval only ranks when at least two rows exist	A selected, ordered subset of caller-supplied pairs	Prompt inclusion in generator or curator input	Similarity licenses relevance ranking under the supplied vector space only; it does not license truth or correctness	Determines which examples the current call consumes	Wrapper to generator/curator, prompt context, ranking and permissive inclusion, current call	SRC-1 dynamic_cheatsheet/language_model.py; see RT-RTE-2, RT-RTE-3, RT-RTE-5	EPI-CLM-1	none	Vector semantics, alignment and retrieval effect are not established.
RT-RTE-4	operational admission/selection/consumption	implemented	Prior strings in RT-OBJ-2	truth-apt transformation: non-ampliative reshaping — pairs are serialized into a prompt; their truth is not changed or checked by serialization	All caller-supplied prior inputs and outputs	Program appends supplied pairs; no selection check or correctness criterion	Each chosen full-history call; operation unobserved	Formatted prior pairs and query are sent to the model	Prompt context only	No epistemic authority is added; supplied answers remain unchecked candidates	Changes current prompt context	Wrapper to generator, advisory context, current call	SRC-1 dynamic_cheatsheet/language_model.py; see RT-RTE-4	EPI-CLM-1, EPI-CLM-2	none	Supplied history provenance, alignment and effect are unobserved.
RT-RTE-6	content transformation	implemented	RT-OBJ-3; answer candidate EPI-OBJ-1	truth-apt transformation: indeterminate — provider response may derive, import or conjecture; no output evidence distinguishes the possibilities	Raw caller prompt and optional prior messages to provider response	Provider model is opaque; no answer oracle or acceptance criterion is wired by the client	Each direct generation/chat request; operation unobserved	Text or stream returned; chat appends user and assistant turns	Returned to caller and available as later in-object chat context	No epistemic authority beyond unverified candidate text	Later chat requests may consume stored messages until caller clears, replaces or discards the object	Client object to provider, message channel, advisory context, object lifetime	SRC-1 text_generation/simple_unified_client.py; see RT-RTE-6	none	none	Provider behavior, response correctness and later effects are unobserved.
EPI-RTE-1	content transformation	implemented	EPI-OBJ-7	truth-apt transformation: acquisition/import (source warrant unknown): Tavily result strings enter the prompt; provider-native search is also wired for supported client paths	Query to returned titles, URLs and snippets; then supplied as search context to model	Tavily service supplies results; code formats them. Native tools are provider-managed. Neither path includes an inspected source-quality check or answer oracle	Direct client generate only when caller enables configured Tavily search, or enables supported provider-native search; operation unobserved	Search context is prepended to prompt, or provider tool request is made; model response is returned	Search text can inform the generated answer; no separate acceptance gate	At most imported candidate material with unknown source warrant; no endorsement is established	Can change current prompt and model response; not automatically retained by this library’s chat list unless included in the prompt history by another caller action	UnifiedLLMClient to provider model, prompt/tool channel, permissive context, current request	SRC-1 text_generation/simple_unified_client.py	none	none	Tavily/provider responses, cited page contents, provider tool execution and output provenance are unavailable. In OpenAI chat-completions code, native web search proceeds only for search-preview models; regular models log a warning and continue without a tool.
RT-RTE-7	behavior/policy adaptation	implemented	EPI-OBJ-5, EPI-OBJ-6, RT-OBJ-4	non-truth-apt policy/content update: model text is admitted as executable code after the configured flag and fence checks	Proposed Python source and its host-process effects	Wrapper checks a textual trigger only; host Python executes; no content check, approval, sandbox or rollback is established	When caller enables execution and response has the configured flag after a code fence	Captured output is appended to generation text; external effects are possible but unobserved	Automatic subprocess launch with host process authority	No epistemic license; output is an unchecked candidate result	Can execute operations allowed to the host and change the next generation context; side effects may outlast the request	LanguageModel.generate to host Python, execution channel, enforcing launch, recursive call and possible host-side horizon	SRC-1 dynamic_cheatsheet/language_model.py, dynamic_cheatsheet/utils/execute_code.py; see RT-RTE-7, RT-BAP-1	none	none	Runtime permissions, actual execution, output correctness and effects are unknown.
RT-RTE-8	behavior/policy adaptation	implemented	EPI-OBJ-5, EPI-OBJ-6, RT-OBJ-4	non-truth-apt policy/content update: caller-selected tool request delegates code execution to a provider	Provider tool request and returned output	Wrapper constructs OpenAI/Claude requests; provider owns tool execution; no source-inspected result check or rollback	When caller selects supported provider option and required container/sentinel state; operation unobserved	Provider may return execution-informed text	Request wiring only; provider-side effects and control are outside inspected layer	No epistemic license established	Provider tool use may change returned text and provider environment; local wrapper does not admit a checked result	Client to provider tool, request channel, provider-controlled force and horizon	SRC-1 text_generation/simple_unified_client.py; see RT-RTE-8, RT-BAP-2	none	none	Provider execution, visibility, persistence, isolation and outcome are uninspected.

The deterministic extraction classifications above are source-inspected deductions from the inspected code. The tag parser can preserve the selected substring’s literal text, but that does not establish semantic preservation relative to a previous sheet because a model supplied the candidate text. Similarity ranking likewise establishes only computational ordering, not the quality of the paired answer.

new_cheatsheet = extract_cheatsheet(response=cheatsheet_output, old_cheatsheet=current_cheatsheet) cheatsheet = new_cheatsheet --- dynamic_cheatsheet/language_model.py:434-435 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

if "" in response: try: txt = response.split("")[1].strip() txt = txt.split("")[0].strip() return txt except: return old_cheatsheet else: return old_cheatsheet --- dynamic_cheatsheet/utils/extractor.py:78-86 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

similarities = cosine_similarity([current_original_input_embedding], prev_original_input_embeddings) top_k_indices = np.argsort(similarities[0])[::-1][:retrieve_top_k] --- dynamic_cheatsheet/language_model.py:510-511 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

if use_web_search and self.web_search_enabled: query = search_query or prompt search_context = self._web_search(query) if search_context and not search_context.startswith("Web search failed"): prompt = f"Web Search Results:\n{search_context}\n\nUser Query: {prompt}" --- text_generation/simple_unified_client.py:470-474 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

results.append( f"{idx}. {result.get('title', 'No title')}\n" f" URL: {result.get('url', 'N/A')}\n" f" {result.get('content', 'No content')}\n" ) --- text_generation/simple_unified_client.py:383-387 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

web_search_enabled: bool = False, tavily_api_key: Optional[str] = None, default_web_search: bool = False, --- text_generation/simple_unified_client.py:105-107 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

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-1	Persistent, evolving memory across queries	SRC-2 README.md (doctrine/design)	The README describes a growing knowledge base and demonstrates passing the returned sheet to a later call.	RT-RTE-1, RT-RTE-3; caller-owned handoff in RT-RTE-1	None supplied	None supplied; no comparison or persistence trace	The implementation returns an extracted candidate sheet and later calls can consume caller-resubmitted text. This supports a caller-mediated memory affordance, not library-owned durable persistence, truth, or successful evolution.	The README’s “persistent” wording is broader than the inspected library’s storage responsibility; no operation is observed.
EPI-CLM-2	Experience-driven learning without ground-truth labels or human feedback	SRC-2 README.md (doctrine/design)	The README states the experience-driven learning and zero-shot claims.	RT-RTE-1, RT-RTE-2, RT-RTE-3, RT-RTE-4, RT-RTE-5	None supplied	None supplied; there is no capability measure, control comparison or attribution evidence	The routes can place caller-resubmitted sheets or prior examples into later prompts. No evidence shows improved future capacity or that a candidate update is correct.	The implementation has no evidence-consuming acceptance gate for learned content; whether the resulting behavior improves is unknown.
EPI-CLM-3	Benchmark performance gains, including Game of 24 and knowledge-intensive tasks	SRC-2 README.md (doctrine/design)	Performance figures are stated in README; the boundary excludes benchmark notebooks, datasets and paper evidence.	RT-RTE-1, RT-RTE-7	None supplied	No inspected benchmark run, comparison design or causal evidence	Only that the README reports those outcomes. This run cannot validate the numbers or attribute them to a route or component.	Reported outcomes remain unsupported by the frozen evidence set.

A lightweight framework that gives language models (LMs) a persistent, evolving memory during inference time. --- README.md:9-9 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

• Experience-Driven Learning: Bridges the gap between isolated inference events and cumulative learning --- README.md:21-21 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

• Puzzles: GPT-4o's success rate on Game of 24 increased from approximately 10% to 99% after discovering and reusing Python-based solutions --- README.md:26-26 @ 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9

Bounded conclusion

The inspected library produces model-written answer candidates and, in cumulative modes, model-proposed cheatsheet text. Its only inspected check on a sheet proposal is whether expected tags appear; a successful parse admits the enclosed string for later rounds and return. That check does not assess truth, evidence, intended use or scope. The wrapper does not implement an answer oracle or an evidence-consuming acceptance decision. Returned text becomes later context only through another within-call round or caller resubmission.

Retrieval imports caller-supplied input/output strings with unknown provenance and warrant. Similarity ranks their paired vectors and chooses which examples enter a prompt. That rank licenses relevance only within the supplied vector space. Search-enabled direct clients can also insert external result snippets or request provider-native search; source warrant and the returned content were not observed. Neither path verifies imported claims. The client’s direct chat history can retain returned candidate text in the client object, but that does not endorse it.

Local Python execution is a consequential operational route. Caller configuration and a response trigger can launch model-proposed code with the host process’s authority, and captured output can enter another model request. This creates operational authority without an epistemic check; the actual execution environment and effects are unobserved. The provider-native tool route is wired but its execution, authority and persistence remain provider-owned and uninspected.

The README’s persistent-memory claim is supported only as caller-mediated reuse of returned text. The evidence does not establish durable library storage, accurate accumulated knowledge, experience-driven improvement, or the reported performance gains. CLI dispatch and hooks are outside the registered implementation paths, so their routes remain unassessed. No single epistemic verdict extends beyond these inspected paths.

Shared records

Operative objects

EPI-OBJ-1 — Returned answer candidate

• Source-native identity: answer text returned by advanced_generate or a direct client generation call.
• Representational form: natural-language model response, sometimes parsed from answer tags or a legacy answer format.
• Storage substrate: return value and, on chat, optionally the client’s in-memory message list.
• Distinct identity comparison: RT-OBJ-3 covers the client message history, which can contain an assistant response, but not every advanced_generate return; this record denotes the returned answer candidate across the wrapper path. No distinct correctness or lineage record was supplied.
• Evidence: SRC-1 at dynamic_cheatsheet/language_model.py, dynamic_cheatsheet/utils/extractor.py, and text_generation/simple_unified_client.py; see RT-RTE-1 and RT-RTE-6.

EPI-OBJ-2 — Prior input strings

Part of: RT-OBJ-2

• Source-native identity: original_input_corpus entries aligned by index with supplied outputs and embeddings.
• Representational form: caller-provided text; propositions, questions and non-truth-apt task instructions may coexist.
• Storage substrate: caller-owned in-memory argument arrays/lists.
• Evidence: SRC-1 at dynamic_cheatsheet/language_model.py; parent RT-OBJ-2 is a broader heterogeneous grouping.

EPI-OBJ-3 — Prior generated output strings

Part of: RT-OBJ-2

• Source-native identity: entries in generator_outputs_so_far paired by index with prior inputs.
• Representational form: natural-language model output; may contain truth-apt answer candidates.
• Storage substrate: caller-owned in-memory list supplied to a call.
• Evidence: SRC-1 at dynamic_cheatsheet/language_model.py; parent RT-OBJ-2 groups these with inputs and embeddings despite their distinct content and epistemic role.

EPI-OBJ-4 — Supplied embedding rows

Part of: RT-OBJ-2

• Source-native identity: rows in original_input_embeddings, including the final current-query row.
• Representational form: numeric vectors; no proposition or truth condition can be identified from the registered implementation.
• Storage substrate: caller-provided numeric array consumed in process memory.
• Evidence: SRC-1 at dynamic_cheatsheet/language_model.py; parent RT-OBJ-2 also groups the paired text and output records. The producer remains unidentified (RT-CMP-2).

EPI-OBJ-5 — Proposed executable code

Part of: RT-OBJ-4

• Source-native identity: Python code extracted from model response or instructions submitted to a provider-native execution tool.
• Representational form: executable source text; its instructions may have effects, but no truth-apt claim is established by its syntax alone.
• Storage substrate: model response, temporary local Python file, or provider tool request.
• Evidence: SRC-1 at dynamic_cheatsheet/language_model.py, dynamic_cheatsheet/utils/execute_code.py, and text_generation/simple_unified_client.py; parent RT-OBJ-4 combines proposal text with execution environments and results.

EPI-OBJ-6 — Execution output text

Part of: RT-OBJ-4

• Source-native identity: captured stdout/stderr or provider-returned execution-related response.
• Representational form: plain text that may state a computation result or environment observation.
• Storage substrate: return string and, in recursive local execution, the next generation prompt; provider output otherwise returns to caller.
• Evidence: SRC-1 at dynamic_cheatsheet/utils/execute_code.py and text_generation/simple_unified_client.py; parent RT-OBJ-4 is heterogeneous across local and provider-controlled substrates.

EPI-OBJ-7 — Search result strings

• Source-native identity: title, URL and content values formatted from Tavily results for direct client generation; native provider search results are not exposed as inspectable strings in the registered code.
• Representational form: external natural-language snippets, potentially containing truth-apt claims.
• Storage substrate: temporary prompt string constructed by UnifiedLLMClient.generate.
• Distinct identity comparison: no supplied operative-object record covers this direct search-result prompt material; it is separate from the cumulative-sheet and example-array records.
• Evidence: SRC-1 at text_generation/simple_unified_client.py; the result-formatting and prompt-insertion passages above support the implementation path.

EPI-OBJ-8 — Proposed cumulative sheet segment

Part of: RT-OBJ-1

• Source-native identity: the substring selected from a curator response between <cheatsheet> and </cheatsheet>.
• Representational form: model-authored plain text, potentially mixing factual propositions, strategies and instructions.
• Storage substrate: curator response, then local variable; on a successful tag parse it replaces the in-call sheet and is returned to the caller.
• Evidence: SRC-1 at dynamic_cheatsheet/language_model.py and dynamic_cheatsheet/utils/extractor.py; the transient proposal is a part of the broader cumulative sheet route, not evidence of an accepted fact.

Routes

EPI-RTE-1 — Direct-client web search

• Endpoints and progression: caller invokes UnifiedLLMClient.generate with Tavily search enabled or supported provider-native search selected; a query is sent to Tavily or a provider tool, result context is placed into the current prompt or provider request, and a model response returns to the caller.
• Owner: caller configures search and query; client formats Tavily snippets and assembles provider requests; external service and provider control retrieval and generation.
• Context, state and action effects: Tavily title, URL and content strings are prepended to the prompt; native provider search is requested only through supported provider paths. The library does not itself retain a validated source record.
• Decision roles: Tavily/provider retrieval chooses returned material; code performs formatting and insertion. No human reviewer, source-quality evaluator, expected-answer oracle or claim-level acceptance gate is wired.
• Answer oracle: none wired; search results serve as context, not a known-answer criterion.
• Operating mode and improvement trigger: caller-triggered open query; no bounded curriculum or correctness-feedback update is implemented. Search config can include provider-specific limits/filter settings, but no source validation method is registered.
• Admission: result strings are included when returned without the failure prefix; native provider tool use is requested according to client options and provider/model route. No epistemic check precedes prompt inclusion.
• implementation conclusion status: wired
• operation conclusion status: uninspected
• Guarantee strength: no claimed guarantee
• Immediate return: provider response text or stream; Tavily snippets are intermediate prompt context.
• Later read-back: inapplicable — this route does not retain search results as library memory; caller may retain the returned answer outside the inspected boundary.
• Delegated visibility: uninspected — provider-native search internals and visibility are provider-owned.
• Selection predicate: Tavily query is search_query or prompt; provider-native search depends on caller option/configuration and provider-specific branch. OpenAI chat-completions only uses automatic search for search-preview models; ordinary models continue without a search tool.
• Invalidation or expiry: inapplicable — no local result cache or expiry is implemented in this route.
• Activation or effect: source code inserts Tavily result strings into the model prompt or requests native search; actual retrieval and effect on response are unobserved.
• Evidence limits: SRC-1 at text_generation/simple_unified_client.py. Source inspection establishes request and prompt wiring only; external results, source reliability, provider behavior, operation, answer correctness and later effect are unknown.

Claims

EPI-CLM-1 — Persistent, evolving memory

• Claimed operation: language models receive persistent and evolving memory during inference, including reuse of a returned sheet on a later query.
• Source: SRC-2 at README.md; doctrine/design.
• Evidence: the README claim passage above and its two-call example; implementation paths RT-RTE-1 and RT-RTE-3.

EPI-CLM-2 — Experience-driven learning without labels or feedback

• Claimed operation: repeated query experience yields cumulative learning without ground-truth labels or human feedback.
• Source: SRC-2 at README.md; doctrine/design.
• Evidence: README.md feature claims; implementation candidate routes RT-RTE-1, RT-RTE-2, RT-RTE-3, RT-RTE-4, RT-RTE-5. The source contains no checked capability update or learning result in this boundary.

EPI-CLM-3 — Benchmark performance gains

• Claimed operation: performance improved on named mathematics, puzzle, arithmetic and knowledge-intensive benchmarks.
• Source: SRC-2 at README.md; doctrine/design.
• Evidence: README performance claims, including the retained Game of 24 passage above. The boundary explicitly excludes the paper, benchmark notebooks, datasets and execution results.

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member