ByteRover CLI epistemic report

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

Source-and-claim boundary

This report asks what licenses reliance on the project knowledge ByteRover stores and supplies, rather than whether the files exist. Scope, dependencies, frozen source and exclusions: see boundary and its Source register, SRC-1. The boundary is a complete repository artifact but a partial knowledge loop: external authoring/consuming agents, providers, Cloud, shared UI and transport internals are excluded. Their absence prevents end-to-end judgments about authorship, factual checking, actual reliance and deployed behavior. No execution or causal-experiment source is registered.

Public claims assessed are RT-CLM-project-memory, EPI-CLM-agent-native, EPI-CLM-benchmark-results and EPI-CLM-query-answer. None is treated as evidence of current-route operation. Runtime findings are provisional: counterpaths already supplied there defeat a universal curate gateway, mandatory pre-write human approval and internal-model authorship of all knowledge. Direct inspection below further separates a recorded review decision from evidence-consuming factual acceptance, and a maturity label from factual validation.

Coverage

“Supplied” means reliance on a named runtime record, not a direct reread. “Direct” means additional SRC-1 source inspection in this member. All paths are commit-relative at the reviewed revision.

Entry point or operation	Source path	Coverage and records, or remaining limit
CLI dispatch and lifecycle hooks	package.json; bin/run.js; src/oclif/hooks/	Direct package registration; supplied launcher account in RT-RTE-curate-session. Init/update/config/error hooks are enumerated, not traced through all bodies. Product-update admission remains unassessed.
CLI/MCP curate	src/oclif/lib/curate-session.ts; src/server/infra/mcp/tools/brv-curate-tool.ts; src/server/infra/mcp/tools/index.ts	Supplied RT-RTE-curate-session, RT-RTE-topic-admission; direct MCP export inventory.
Writer and factual attribute validation	src/server/infra/render/writer/html-writer.ts; src/server/infra/render/elements/make-validator.ts; src/server/infra/render/elements/bv-fact/schema.ts; src/server/infra/render/elements/bv-fact/validator.ts	Direct; EPI-RTE-fact-validation.
Query, selection and cache	src/server/infra/executor/query-executor.ts; src/oclif/lib/query-retrieval.ts	Supplied RT-RTE-query-retrieval. Cache confidence/freshness is not endorsement.
Rendered versus raw read	src/oclif/commands/read.ts; src/oclif/lib/read-topic.ts; src/server/infra/render/reader/html-renderer.ts; src/server/infra/render/reader/html-parser.ts	Direct; EPI-RTE-render-content, EPI-RTE-read-topic. Raw read is a counterexample to universal renderer loss.
Search	src/oclif/commands/search.ts	Direct command description and request setup; EPI-CLM-query-answer. Search result formatting/excerpt completeness not independently traced; no equivalence to a complete topic asserted.
Internal curate, context compaction	src/server/infra/executor/curate-executor.ts; src/agent/infra/llm/agent-llm-service.ts	Supplied RT-RTE-agent-curation, RT-RTE-context-control, RT-CMP-content-generator. Semantic entailment of generated outputs remains unresolved.
Registered tools	src/agent/infra/tools/tool-registry.ts	Direct factory inventory includes map, code execution, curate, expand, file discovery/read/write, ingest, search, map and swarm operations. Supplied internal-control account covers dispatch controls, not every worker. Ingest extraction, map/swarm judgments and their semantic admissions remain uninspected; prevents a system-complete negative about inference or checking.
Review and withdrawal	src/server/infra/transport/handlers/review-handler.ts; src/server/infra/process/curate-html-log.ts; src/server/infra/transport/handlers/push-handler.ts	Direct review handler; supplied pending-selection and push consumers in RT-RTE-topic-admission, RT-RTE-cloud-sync; EPI-RTE-review-decision.
Direct edit, hub, Cloud	src/server/infra/transport/handlers/context-tree-handler.ts; src/server/infra/hub/hub-install-service.ts; src/server/infra/transport/handlers/pull-handler.ts	Supplied RT-RTE-direct-edit, RT-RTE-hub-install, RT-RTE-cloud-sync. Imported bytes have unknown source warrant; no semantic audit inferred.
Dream scan, finalize, undo	src/server/infra/dream/tool-mode/dream-session.ts; src/server/infra/dream/tool-mode/prune-candidates.ts; src/server/infra/dream/tool-mode/merge-candidates.ts; src/server/infra/dream/dream-undo.ts; src/oclif/commands/dream/undo.ts	Direct selected algorithms, archive and undo; EPI-RTE-prune-candidates, EPI-RTE-dream-undo, supplied RT-RTE-dream-maintenance. Link and synthesis generator internals not independently assessed; orchestration does not establish semantic synthesis.
Hook and maintenance instructions	src/oclif/commands/hook-prompt-submit.ts; src/server/templates/sections/brv-instructions.md; src/server/templates/skill/dream.md	Direct; EPI-RTE-hook-instructions. Shipped instructions request future behavior, not observed compliance.
Versioning, sources, worktrees, migration, reset, connector installation	src/oclif/commands/vc/; src/server/infra/transport/handlers/vc-handler.ts; src/server/infra/transport/handlers/source-handler.ts; src/server/infra/transport/handlers/worktree-handler.ts; src/server/infra/transport/handlers/migrate-handler.ts; src/server/infra/transport/handlers/reset-handler.ts; src/oclif/commands/connectors/	Available, not deeply inspected here; supplied runtime explicitly leaves these incomplete. They can alter retained content, scope or instruction delivery. Prevents complete lineage, rollback and admission judgments, not the positive route findings below.
Public evaluation	README.md; paper/README.md; paper/main.tex	Direct reported claims, metrics and ablation sections; EPI-CLM-benchmark-results. No question-level outputs, executed harness or intervention records supplied; prevents independent accuracy verification and causal attribution. Tests remain available but uninspected evaluators of implementation, not factual topic acceptance.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-topic-statements	Project/personal assertions, reasons and explanations, scoped by authored text	Typed facts and retained knowledge	SRC-1, src/server/infra/render/elements/bv-fact/schema.ts	A type label does not supply evidence or identify an actual true proposition in a supplied run.
EPI-OBJ-topic-rules	Prescriptive rules are not inherently truth-apt; any accompanying descriptive premises are assessed separately	Behavior guidance	see EPI-OBJ-topic-rules	Host interpretation determines compliance; renderer severity is not enforcement.
RT-OBJ-context-tree	Mixed container, not one proposition	see RT-OBJ-context-tree	Container identity remains valid; semantic parts above need no supersession.
EPI-OBJ-runtime-signals	Counters/timestamps support bounded arithmetic statements; importance/maturity are policy labels	Ranking and cleanup inputs	SRC-1, src/server/core/domain/knowledge/runtime-signals-schema.ts	validated is not a recorded factual acceptance criterion.
RT-RTE-topic-admission	HTML errors, write success, pending status and backup are symbolic operational objects, not factual verdicts	see RT-RTE-topic-admission	Selection and recovery refer to changes, not claim truth.
RT-RTE-query-retrieval	Matched topic text and cache payload carry candidate assertions; scores select rather than verify them	see RT-RTE-query-retrieval	Shared/unreadable hits may be omitted; rendered content may lose evidence links.
RT-RTE-context-control	Model summaries may contain truth-apt task claims; deterministic checkpoint metadata describes task state	see RT-RTE-context-control	No concrete supplied summary permits entailment or factual acceptance assessment.
RT-CMP-content-generator	Distributed parameters are not individually accessible propositions here	see RT-CMP-content-generator	Parameter mutation, exact weights and endpoint contents remain opaque; no parametric-learning inference.
RT-RTE-hub-install	Bundles may contain assertions; skills may contain prescriptions	see RT-RTE-hub-install	Imported content provenance is publisher-based, not locally warranted.

Authority-route ledger

Each row is a distinct consequential function. Generic route fields owned by the cited declaration are referenced with see; epistemic classifications here apply only to the named edge, not every possible output of its enclosing route. Architectural status denotes inspection support, not operation.

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-curate-session	content transformation	implemented	EPI-OBJ-topic-statements	truth-apt transformation: acquisition/import	Caller HTML enters target	Caller authors; semantic checking external	see RT-RTE-curate-session	see RT-RTE-curate-session	Authoring prompt advisory; envelope enforcing	Source warrant unknown	Caller can submit candidate for writing	Caller → continuation HTML → writer; permissive, session	SRC-1, src/oclif/lib/curate-session.ts	RT-CLM-project-memory	none	No attribution of caller reasoning to ByteRover
EPI-RTE-fact-validation	check/evidence production	implemented	EPI-OBJ-topic-statements	no content change	Topic structure and fact attributes	Registry/Zod; markup domain only	Before writer I/O	Schema success or errors	Enforcing input to writer	License only for checked syntactic predicates	Writer may proceed or return correction	Validator → writer result → caller, enforcing, one submission	SRC-1, src/server/infra/render/writer/html-writer.ts	none	none	Not a factual oracle
RT-RTE-topic-admission	disposition/acceptance	implemented	RT-OBJ-context-tree	no content change	Structural success and overwrite policy	see RT-RTE-topic-admission	see RT-RTE-topic-admission	see RT-RTE-topic-admission	Enforcing within this writer	Operational admission, not evidence-consuming acceptance of truth	Allows write or rejects candidate	Writer/caller, structured error/success, enforcing, submission	SRC-1, src/server/infra/render/writer/html-writer.ts	none	none	Direct edit and imports are counterpaths
RT-RTE-topic-admission	retention	implemented	RT-OBJ-context-tree	no content change	Persist admitted topic	see RT-RTE-topic-admission	see RT-RTE-topic-admission	see RT-RTE-topic-admission	Makes bytes available	No new semantic warrant from persistence	Later retrieval eligible before review	Tree → query/read, permissive, until changed	SRC-1, src/server/infra/render/writer/html-writer.ts	RT-CLM-project-memory	none	Not lifecycle integration after factual acceptance
RT-RTE-agent-curation	content transformation	implemented	EPI-OBJ-topic-statements	truth-apt transformation: indeterminate	Model extraction/grouping/final HTML	see RT-RTE-agent-curation	see RT-RTE-agent-curation	see RT-RTE-agent-curation	Prompt guidance then writer enforcement	No checked source-to-output entailment	Can author and replace retained guidance	Model output → writer → future readers, permissive, persistent	SRC-1, src/server/infra/executor/curate-executor.ts	EPI-CLM-agent-native	none	Reshaping, derivation and conjecture unresolved without concrete input/output
RT-RTE-query-retrieval	operational admission/selection/consumption	implemented	RT-OBJ-context-tree	no content change	Relevance and deliverability	see RT-RTE-query-retrieval	see RT-RTE-query-retrieval	see RT-RTE-query-retrieval	Ranking/filtering	Similarity licenses selection, not truth	Chooses what caller can use	see RT-BAP-topic-guidance	SRC-1, src/server/infra/executor/query-executor.ts	EPI-CLM-query-answer	Query answer label exceeds target synthesis on this route	No recall or actual reliance evidence
EPI-RTE-render-content	content transformation	implemented	EPI-OBJ-topic-statements	truth-apt transformation: non-ampliative reshaping	HTML assertions to text	Deterministic rendering; selected typed elements	During query/read	Selected text survives; href may not	Advisory delivery	Proposition text can survive while evidence lineage degrades	Caller receives incomplete source representation	Rendered markdown → caller, advisory, invocation	SRC-1, src/server/infra/render/reader/html-renderer.ts	RT-CLM-project-memory	none	Not lossless or a truth check
EPI-RTE-read-topic	operational admission/selection/consumption	implemented	RT-OBJ-context-tree	no content change	Explicit requested path	Path checks and readable file, not relevance scoring	Each read request	Raw or rendered content/error	Permissive delivery	Raw bytes preserve available source representation, not source truth	Supports inspection before rewrite	CLI content → caller, advisory, invocation	SRC-1, src/oclif/lib/read-topic.ts	RT-CLM-project-memory	none	No proof caller inspects evidence
EPI-RTE-review-decision	disposition/acceptance	implemented	RT-OBJ-context-tree	no content change	Pending operations by task/path	Requesting reviewer; no encoded factual criterion or oracle	Conditional post-write	Approved/rejected statuses	Enforcing status changes	Recorded choice, not established evidence-consuming factual acceptance	Approval keeps current state; rejection requests reversion	Request → handler/log → push filtering, enforcing, until later change	SRC-1, src/server/infra/transport/handlers/review-handler.ts	none	none	Human identity and evidence use not established
EPI-RTE-review-decision	lineage/freshness/recovery	implemented	RT-OBJ-context-tree	no content change	Rejected file change	Backup presence and I/O success	After reject request	Restore old bytes or unlink	Physical rollback, best effort extras	Withdrawal of current bytes is not refutation	Changes later available guidance	Backup → tree → later readers, enforcing file effect, persistent	SRC-1, src/server/infra/transport/handlers/review-handler.ts	none	none	Missing backup may delete; cache/index retraction not established
RT-RTE-cloud-sync	operational admission/selection/consumption	implemented	RT-OBJ-context-tree	no content change	Pending/rejected push paths	see RT-RTE-cloud-sync	see RT-RTE-cloud-sync	see RT-RTE-cloud-sync	Push exclusion policy	Approval status licenses export, not truth	Blocks selected legacy Cloud submissions	Handler → Cloud request, enforcing, push	SRC-1, src/server/infra/transport/handlers/push-handler.ts	RT-CLM-project-memory	none	Does not gate local query or establish Git VC policy
RT-RTE-direct-edit	retention	implemented	RT-OBJ-context-tree	truth-apt transformation: acquisition/import	Client-supplied replacement bytes	see RT-RTE-direct-edit	see RT-RTE-direct-edit	see RT-RTE-direct-edit	Direct persistent mutation	Source warrant unknown	Bypasses curate structural/review gateway	Client → file → readers, permissive, persistent	SRC-1, src/server/infra/transport/handlers/context-tree-handler.ts	none	none	Semantic content depends on request
RT-RTE-hub-install	retention	implemented	RT-OBJ-context-tree	truth-apt transformation: acquisition/import	Publisher bundle bytes	see RT-RTE-hub-install	see RT-RTE-hub-install	see RT-RTE-hub-install	Persistent file admission	Source warrant unknown	Bundle becomes eligible knowledge	Publisher → installer → tree, permissive, persistent	SRC-1, src/server/infra/hub/hub-install-service.ts	none	none	Skill branch is policy admission, separately below
RT-RTE-hub-install	behavior/policy adaptation	implemented	see RT-RTE-hub-install	non-truth-apt policy/content update: install agent skill	External host instruction surface	see RT-RTE-hub-install	see RT-RTE-hub-install	see RT-RTE-hub-install	Installation permissive; host force unresolved	No epistemic license from install	Makes skill available to host	Connector files → excluded host loader, permissive, installed lifetime	SRC-1, src/server/infra/hub/hub-install-service.ts	none	none	Not observed activation or self-improvement
RT-RTE-cloud-sync	retention	implemented	RT-OBJ-context-tree	truth-apt transformation: acquisition/import	Cloud snapshot	see RT-RTE-cloud-sync	see RT-RTE-cloud-sync	see RT-RTE-cloud-sync	Local snapshot admission	Source warrant unknown	Replaces eligible local state after divergence guard	Cloud → pull → tree, permissive, snapshot lifetime	SRC-1, src/server/infra/transport/handlers/pull-handler.ts	RT-CLM-project-memory	none	Remote correctness/access controls excluded
EPI-RTE-prune-candidates	check/evidence production	implemented	EPI-OBJ-runtime-signals	truth-apt transformation: entailed derivation	Low-importance/stale predicates	Arithmetic and symbolic tier rules	Scan	Predicate reason and age	Advisory candidate ranking	Entailed only relative to input fields and clock	Offers candidates, does not archive them	Scan envelope → caller, advisory, scan	SRC-1, src/server/infra/dream/tool-mode/prune-candidates.ts	none	none	Does not establish stale knowledge or semantic redundancy
RT-RTE-dream-maintenance	disposition/acceptance	implemented	RT-OBJ-context-tree	no content change	Caller archive list	see RT-RTE-dream-maintenance	see RT-RTE-dream-maintenance	see RT-RTE-dream-maintenance	Enforcing removal from active tree	Caller choice not factual refutation	Archives paths without proving survivor preserved facts	Caller → finalize → active tree, enforcing, until undo	SRC-1, src/server/infra/dream/tool-mode/dream-session.ts	none	none	Stateless session does not certify scan membership or semantic merge
EPI-RTE-dream-undo	lineage/freshness/recovery	implemented	RT-OBJ-context-tree	no content change	Last dream state	Log preconditions and saved data	Explicit undo	Restored files/signals or collected errors	Physical recovery	Restores prior state, not prior warrant	Can re-expose archived guidance and prune predicates	Log → files/signals → readers/scans, enforcing, persistent	SRC-1, src/server/infra/dream/dream-undo.ts	none	none	Not all historical operations reversible; errors can coexist with undone status
RT-RTE-context-control	content transformation	implemented	see RT-RTE-context-control	truth-apt transformation: indeterminate	Model summary replaces working history	see RT-RTE-context-control	see RT-RTE-context-control	see RT-RTE-context-control	Changes subsequent model input	Summary fidelity unchecked here	Continues with reduced input	Summary boundary → next model call, permissive, task	SRC-1, src/agent/infra/llm/agent-llm-service.ts	none	none	Deterministic checkpoint branch differs; no factual acceptance implied
EPI-RTE-hook-instructions	operational admission/selection/consumption	implemented	Static workflow text, not acquired experience	non-truth-apt policy/content update: supply workflow instructions	Host pre-prompt context	Template loader; host activation excluded	Hook invocation	Workflow text or silent production failure	Advisory request; host placement unresolved	No factual warrant	Requests query-before-work and curate-after-work	Stdout → host context → caller, advisory here, prompt	SRC-1, src/oclif/commands/hook-prompt-submit.ts	none	none	“MANDATORY” is not local enforcement

No accepted-claim lifecycle-integration route is established by the writer, review handler or dream paths inspected here. This is a bounded conclusion about those mechanisms, not a system-complete absence record: uninspected tool workers and excluded caller judgment prevent that wider negative. Topic links and synthesis can change organization before any factual acceptance is evidenced.

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-project-memory	see RT-CLM-project-memory	SRC-1, README.md, doctrine/design	see claim	RT-RTE-topic-admission, RT-RTE-query-retrieval, RT-RTE-cloud-sync	None supplied	No demonstrated improvement or teammate consumption	Persistent availability and local delivery are wired	Knowledge label does not establish acceptance or factual warrant
EPI-CLM-agent-native	see EPI-CLM-agent-native	SRC-1, paper/README.md, doctrine/design	see claim	RT-RTE-agent-curation, RT-RTE-curate-session, RT-RTE-query-retrieval	None supplied	No author/consumer traces	Internal model route exists; public tool mode delegates authorship and synthesis	LLM retrieval is not the current deterministic public query path; zero-infrastructure wording is not a provider-independence proof
EPI-CLM-query-answer	see EPI-CLM-query-answer	SRC-1, src/oclif/commands/search.ts, doctrine/design	see claim	RT-RTE-query-retrieval	None supplied	No host synthesis trace	Query supplies context from which caller may synthesize	Description overstates target-side synthesis on tool-mode route
EPI-CLM-benchmark-results	see EPI-CLM-benchmark-results	SRC-1, paper/README.md; paper/main.tex, reported operation/design	see claim	No implemented benchmark acceptance consumer traced; public retrieval is RT-RTE-query-retrieval	None supplied; scores are attributed reporting	Full-versus-no-tier contrast also removes separate justifier; curated tree fixed; write-path components explicitly not ablated; mixed published comparisons use different models/judges	Reported answer-scoring protocol names domain, reference and evaluator	Does not warrant all retained statements, present-route accuracy, mechanism transfer or isolated write-path effect

Bounded conclusion

ByteRover has implemented routes for acquiring candidate statements from callers and publishers, retaining them, selecting them for later use, and requesting behavior through retained rules and static host instructions. Source warrant of imports is unknown, not conferred by storage. The relevant paths are RT-RTE-curate-session, RT-RTE-hub-install, RT-RTE-cloud-sync and RT-BAP-topic-guidance. Direct edits independently admit content. The boundary does not establish actual later compliance or benefit.

Structural checking is consequential but narrowly licensed. EPI-RTE-fact-validation checks attributes and topic structure for writer admission; it does not test the statement against project evidence. EPI-RTE-review-decision records a reviewer-supplied choice and can revert bytes. Neither branch requires a named factual criterion, intended factual use and scope, or records the evidence consumed. Pending content is already eligible locally; legacy push filtering confers operational export authority only. Therefore these statuses do not establish accepted ampliative knowledge or post-acceptance lifecycle integration.

The deterministic renderer demonstrably reshapes rather than verifies content; EPI-RTE-render-content shows one preserved assertion with a lost source URL. EPI-RTE-read-topic supplies a raw alternative, so loss is route-specific. EPI-RTE-prune-candidates derives valid conditional arithmetic results from supplied state, not the truth or obsolescence of topic contents. No warranted source premises for broader project claims are supplied. Maturity, similarity and freshness must not be substituted for endorsement.

The internal model can author topics and compact task history, but accessible evidence does not individuate a produced claim and its premises well enough to decide reshaping, entailment or non-entailment; both generated-topic and summary relations remain indeterminate. The configured generator's parameter changes, exact version and endpoint behavior remain as bounded in RT-CMP-content-generator. This does not negate nonparametric retention. Dream maintenance offers structural candidates and caller-controlled removal, while EPI-RTE-dream-undo restores selected prior state without revalidating truth.

The benchmark evaluator has a named answer/reference domain, but its labels score answers, not each retained claim or its causal explanation. EPI-CLM-benchmark-results supports attributed outcomes and bounded comparison designs, not independently observed effectiveness of the frozen tool-mode route. No general learning, reflection, autonomous-theory-builder or self-improvement verdict follows: guidance delivery does not establish semantic criticism, iterative retention of criticism, or exercised evidence-responsive improvement. These are separate unestablished properties, not one inferred failure. Uninspected workers and excluded agents leave additional routes unresolved.

Shared records

Components

none declared in this member

Operative objects

EPI-OBJ-topic-statements

Label: Topic statements

Part of: RT-OBJ-context-tree

Source-native identity: textual statements in bv-fact and descriptive topic bodies; natural-language representational form, local HTML-file substrate. This is the semantic part of the supplied mixed container, not a replacement for it. Scope and truth depend on authored content; no particular persisted candidate instance was supplied. SRC-1 implementation/design, src/server/infra/render/elements/bv-fact/schema.ts:

The element's text content is the canonical statement; attributes are * the structured extraction. --- src/server/infra/render/elements/bv-fact/schema.ts

EPI-OBJ-topic-rules

Label: Topic rules

Part of: RT-OBJ-context-tree

Source-native identity: bv-rule bodies and their severity/ID signals, persisted in topic HTML. Natural-language normative consumption is distinct from the symbolic rendering of attributes and from descriptive statements. Storage substrate: local files; later reader receives advisory text. SRC-1 implementation, src/server/infra/render/reader/html-renderer.ts; minimum rendering passage and actual authority boundary: see RT-BAP-topic-guidance and RT-RTE-query-retrieval. The normative part differs from the statement part by its consumer function; neither a severity string nor a “Rule” label implements host enforcement.

EPI-OBJ-runtime-signals

Label: Runtime signals

Source-native identity: RuntimeSignals sidecar fields, symbolic numerical/tier state stored per machine rather than shared topic bodies. Distinct from RT-OBJ-context-tree; overlaps RT-RTE-dream-maintenance as its input, not its route identity. SRC-1 implementation, src/server/core/domain/knowledge/runtime-signals-schema.ts:

export const MaturityTierSchema = z.enum(['core', 'draft', 'validated']) --- src/server/core/domain/knowledge/runtime-signals-schema.ts

export const DEFAULT_IMPORTANCE = 50 export const DEFAULT_RECENCY = 1 export const DEFAULT_MATURITY = 'draft' as const export const DEFAULT_ACCESS_COUNT = 0 export const DEFAULT_UPDATE_COUNT = 0 --- src/server/core/domain/knowledge/runtime-signals-schema.ts

The checked schema constrains values, not their semantic warrant. The prune consumer below uses maturity as policy input; this source alone establishes no factual evaluation behind validated.

Routes

EPI-RTE-fact-validation

Label: Fact attribute validation

Part of: RT-RTE-topic-admission

Endpoints/progression: writer validateHtmlTopic walks parsed elements, calls the fact validator through the element registry, receives Zod attribute results, then returns errors or a valid topic to writer disposition. Owner: deterministic writer/validator. Context/state: submitted markup; no project evidence enters the fact attribute evaluator. This check is a material subroute, not a duplicate admission mechanism.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Attribute-validation errors or syntactic success.
• Later read-back: inapplicable — this pure check retains no result by itself; enclosing writer determines topic persistence.
• Delegated visibility: Caller can receive structured errors through the enclosing admission path.
• Selection predicate: Registered bv-* elements select corresponding attribute validators; exactly one topic and a safe nonempty path are also required.
• Invalidation or expiry: inapplicable — each call reevaluates current submitted markup.
• Activation or effect: Errors block this writer's write; success permits the remaining overwrite/write checks.
• Evidence limits: Supports structure/attribute correctness only; no semantic truth test or execution observed.

The proposer is the caller or internal model, the program can veto schema failure, and human contribution is unresolved. This is open-request admission, with schema as reference criterion, not an expected factual-answer oracle. Correction prompts can shape resubmission; enclosing route owns retries and recovery. SRC-1 implementation, src/server/infra/render/writer/html-writer.ts, src/server/infra/render/elements/bv-fact/validator.ts, src/server/infra/render/elements/bv-fact/schema.ts, src/server/infra/render/elements/make-validator.ts:

export const validateBvFact = makeAttributeValidator('bv-fact', BvFactAttributesSchema) --- src/server/infra/render/elements/bv-fact/validator.ts

const parsed = schema.safeParse(node.attributes) --- src/server/infra/render/elements/make-validator.ts

const result = ELEMENT_REGISTRY[el.tagName].validator(el) --- src/server/infra/render/writer/html-writer.ts

Because the checked fact schema uses optional strings/category enumeration and the generic validator parses attributes, changing a fact body's factual assertion alone does not cause this fact evaluator to test the new assertion's truth. This deduction is narrower than a claim about all tools or external reviewers.

EPI-RTE-render-content

Label: Render topic content

Part of: RT-RTE-query-retrieval

Endpoints/progression: raw topic → parseHtml normalized tree → selected topic children → semantic-prefix renderer → caller-facing markdown. Owner: renderer code. The same implementation is reused by the independent single-topic read route below. This is a transformation part of the supplied query route, not a new retrieval selector.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Markdown-like text, including selected header attributes, typed labels and collapsed inner text.
• Later read-back: Rendered envelopes can be cached by the enclosing query route and delivered later; renderer itself does not persist.
• Delegated visibility: Text descendants and selected metadata survive; ordinary link hrefs and unrecognized non-bv top-level elements need not survive.
• Selection predicate: First bv-topic; element children; type-specific rendering and nonempty inner text.
• Invalidation or expiry: inapplicable — renderer is a pure current-input transformation; cache invalidation belongs to enclosing query.
• Activation or effect: Alters content available to caller interpretation, with no new truth test or mandatory compliance.
• Evidence limits: Deduction below is implementation evidence, not a target execution; raw reads provide a lossless-byte alternative.

SRC-1 implementation, src/server/infra/render/reader/html-parser.ts, src/server/infra/render/reader/html-renderer.ts:

if (node.type === 'text') return node.text if (node.type === 'element') { if (node.tagName === 'img') return renderImgMarkdown(node.attributes) return node.children.map((c) => getInlineMarkdownRaw(c)).join(' ') } --- src/server/infra/render/reader/html-renderer.ts

Concrete input deduction, not execution: <bv-topic path="config/timeout" title="Timeout"><bv-fact>Timeout is 30 s. <a href="evidence.txt">Source</a></bv-fact></bv-topic> is parsed into the topic, fact, text and link nodes. Header extraction yields # Timeout; fact recursion joins its text descendants, yielding Timeout is 30 s. Source. The fact branch adds - **Fact**:, so the result is # Timeout\n\n- **Fact**: Timeout is 30 s. Source (escape sequences denote newlines). The proposition is reshaped non-ampliatively, but the evidence URL is discarded. Thus preserving a statement's words does not preserve its available source linkage or warrant. This deduction assumes the imported parser contract used by the inspected wrapper, not a verified dependency implementation.

EPI-RTE-read-topic

Label: Read topic path

Endpoints/progression: CLI read → readTopic → checked project-local path → UTF-8 file → raw or rendered result. Owner: helper and invoking caller. Distinct from RT-RTE-query-retrieval: explicit path selection, no query ranking/cache, optional raw bytes. It reuses EPI-RTE-render-content only in its default HTML branch.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Content/path/format or unsafe-path, not-found or read-failed error.
• Later read-back: Previously retained topic reaches a later read invocation; raw mode preserves available source bytes.
• Delegated visibility: Entire UTF-8 source string with raw flag; otherwise HTML renderer output or unchanged non-HTML text.
• Selection predicate: Caller-chosen relative path; lexical/resolved-root safety and file readability, not relevance or truth.
• Invalidation or expiry: Each read opens current file; no helper cache. File replacement/deletion changes subsequent results.
• Activation or effect: Supplies existing content for inspection or merge authorship; actual interpretation unobserved.
• Evidence limits: No target execution; lexical checks are not a proved physical symlink-containment guarantee.

Selection is an open caller request, code can veto unsafe/unreadable paths, human identity unresolved, no factual answer oracle. SRC-1 implementation, src/oclif/commands/read.ts, src/oclif/lib/read-topic.ts:

const format = relPath.toLowerCase().endsWith('.html') ? 'html' : 'markdown' const content = format === 'html' && !options.raw ? renderHtmlTopicForLlm(raw) : raw --- src/oclif/lib/read-topic.ts

EPI-RTE-review-decision

Label: Review task decision

Endpoints/progression: ReviewEvents.DECIDE_TASK request → pending operations grouped by task/file → optional file filter → reject restores/unlinks, approve leaves current bytes → successful operations' log status changes → backup deletion/notification. Owner: requesting reviewer for choice; handler for effects. Distinct endpoint and consumer chain spanning the post-write review support in RT-RTE-topic-admission and export selection in RT-RTE-cloud-sync; not a duplicate writer admission.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Processed paths, reverted flags and operation count.
• Later read-back: Updated log statuses can affect export selection; restored/deleted topic bytes affect later reads subject to other freshness mechanisms.
• Delegated visibility: Pending/list endpoints expose operation metadata; decision request carries choice, task ID and optional paths, not a factual proof.
• Selection predicate: Matching task ID, pending status, project-relative file path and optional caller file filter.
• Invalidation or expiry: Decision clears applicable backups; failed file processing remains pending. Later changes can replace topic bytes and statuses.
• Activation or effect: Reject changes filesystem content; approval/rejection changes processed operation statuses; notification best effort.
• Evidence limits: No encoded evidence-consuming factual criterion; reviewer human identity and evidence use unresolved; rejection does not demonstrate instantaneous cache/index retraction.

The requesting principal proposes the decision; software selects applicable pending operations and I/O failure can prevent completion. Operating mode is requested review, not an experiment. No expected-answer oracle is implemented. Optional review metadata shaped which changes were offered, not a mandatory semantic trial. Rollback depends on backups and is not guaranteed for every admission path. SRC-1 implementation, src/server/infra/transport/handlers/review-handler.ts; supplied backup failure and pending-selection passages: RT-RTE-topic-admission.

if (decision === 'rejected') { const absolutePath = join(contextTreeDir, relPath) const backupContent = await backupStore.read(relPath) --- src/server/infra/transport/handlers/review-handler.ts

batch.push({operationIndex, reviewStatus: decision}) --- src/server/infra/transport/handlers/review-handler.ts

Approval records a choice but cannot on these fields establish acceptance against a named factual criterion for an intended scope. Conversely, rejection has real physical force; it is not merely a cosmetic label. The export consumer's separate scope is retained in RT-RTE-cloud-sync.

EPI-RTE-prune-candidates

Label: Prune candidate predicates

Part of: RT-RTE-dream-maintenance

Endpoints/progression: loaded topic mtimes and sidecar signals → scope/core filtering → age and importance predicates → reason labels → oldest-first limit → scan envelope. Owner: deterministic candidate generator; caller owns archive/keep/merge judgment. This is a check subroute, not the encompassing semantic maintenance workflow. SRC-1 implementation, src/server/infra/dream/tool-mode/prune-candidates.ts.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Candidate HTML/path/signals plus computed age and low-importance/stale/both reason.
• Later read-back: inapplicable — generator retains no candidate decisions; later scans recompute from persisted files/signals.
• Delegated visibility: Caller receives original topic HTML and signals alongside reasons, not a factual obsolescence verdict.
• Selection predicate: Optional prefix; exclude core; importance below 35 or draft age at least 60 days/validated age at least 120; sort by age and limit by default 20.
• Invalidation or expiry: Changed signals, mtime, clock or scope change next result; no candidate lifetime certificate.
• Activation or effect: Advisory candidate selection only; enclosing finalize requires caller archive request.
• Evidence limits: Correctness is conditional on supplied clock/mtime/signals; no semantic freshness or actual maintenance benefit established.

const daysSinceModified = Math.max(0, (now - t.mtimeMs) / DAY_MS) const lowImportance = t.signals.importance < importanceThreshold const staleMtime = isStaleForMaturity(t.signals.maturity, daysSinceModified, draftDays, validatedDays) --- src/server/infra/dream/tool-mode/prune-candidates.ts

case 'draft': { return daysSinceModified >= draftDays } --- src/server/infra/dream/tool-mode/prune-candidates.ts

Concrete deduction: a non-core draft topic with importance 50 and mtime exactly 60 days before the supplied clock is selected as stale-mtime under defaults, even if its project assertion remains true. The implementation's boundary is inclusive (>=), whereas the dream skill abbreviates it as >60d; implementation governs this calculation. The entailed proposition is about fields and clock, not whether content is outdated. Code proposes, caller decides and can ignore the candidate; no human step or factual oracle is required. Scan is open requested maintenance; thresholds guide selection, not content acceptance. The merge generator separately delegates to BM25 pair discovery with threshold 0.85 (src/server/infra/dream/tool-mode/merge-candidates.ts); that parameter does not license semantic equivalence.

EPI-RTE-dream-undo

Label: Dream undo recovery

Endpoints/progression: explicit CLI undo → last dream ID and log preconditions → reverse logged operations → restore/delete and signal/mtime recovery → review cleanup → manifest rebuild → log undone and dream-state rewind. Owner: CLI helper. Distinct recovery route, not the archive endpoint in RT-RTE-dream-maintenance; restores the state that can feed its next scan.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Restored/deleted/archive-restored paths and errors, or failed precondition.
• Later read-back: Restored topic bytes can reach reads; restored mtimes/signals can make a topic qualify again for pruning.
• Delegated visibility: Caller receives errors alongside successes; no evidence of a factual reacceptance decision is returned.
• Selection predicate: Last logged dream, completed or partial, not already undone; reverse each logged operation.
• Invalidation or expiry: Only latest dream, not a history stack; operation data limits restoration; subsequent file changes are not compared against a content revision before restore.
• Activation or effect: Writes prior text, restores selected metadata, deletes created outputs; cleanup can alter review statuses. Marks undone even when collected per-operation errors remain.
• Evidence limits: Not transactional recovery. Legacy synthesis UPDATE explicitly lacks saved previous text; restoration of sidecar is best effort; no executed rollback supplied.

Caller requests rollback; log state and implementation can veto; there is no factual answer oracle or new semantic approval. Recovery guidance is fixed switch logic over saved operations, not model judgment. Phase-two independent curate writes are not part of this undo; CLI documentation states that boundary. SRC-1 implementation, src/oclif/commands/dream/undo.ts, src/server/infra/dream/dream-undo.ts:

if (op.action === 'UPDATE') { throw new Error(Cannot undo SYNTHESIZE/UPDATE: previousTexts not captured for ${op.outputFile}) } --- src/server/infra/dream/dream-undo.ts

await writeFile(fullPath, content, 'utf8')

      // Restore original mtime so stale-mtime prune candidates
      // re-qualify. `writeFile` stamps mtime=now; `utimes` overwrites.
      const savedMtimeMs = op.previousMtimes?.[filePath]

--- src/server/infra/dream/dream-undo.ts

EPI-RTE-hook-instructions

Label: Hook workflow instructions

Endpoints/progression: hidden pre-prompt command → template loader brv-instructions → stdout → excluded host's context placement. Owner: command for delivery, host for actual invocation/placement, calling agent for compliance. Distinct from retained-topic guidance RT-BAP-topic-guidance: static workflow instructions rather than retrieval of accumulated experience.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Template text on stdout; production load failures do not interrupt the agent.
• Later read-back: inapplicable — static shipped instructions are not accumulated-through-use read-back.
• Delegated visibility: Host receives workflow text asking query first, curate after code changes, and inspect curate logs.
• Selection predicate: Command invocation loads the named static section; external connector/host activation not established here.
• Invalidation or expiry: Instruction bytes follow installed package/template version; no content-derived policy update implemented by this command.
• Activation or effect: Supplies an advisory instruction channel; no local enforcement of the requested query/curate order.
• Evidence limits: No host execution trace, binding system-context placement or compliance observed; source comment describes host wrapping, not an inspected host implementation.

SRC-1 implementation, src/oclif/commands/hook-prompt-submit.ts:

const instructions = await templateLoader.loadSection('brv-instructions')

  // Output to stdout (agent wraps in system context)
  this.log(instructions)

--- src/oclif/commands/hook-prompt-submit.ts

SRC-1 doctrine/design, src/server/templates/sections/brv-instructions.md:

Code task received → brv query FIRST → Work → brv curate → Done --- src/server/templates/sections/brv-instructions.md

Claims

EPI-CLM-agent-native

Label: Agent native memory

Claimed operation: LLM curation, structuring and retrieval through a hierarchical context tree with zero external infrastructure. Distinct from the general persistent-memory purpose in RT-CLM-project-memory; this claim assigns the processing to the LLM. SRC-1 doctrine/design, paper/README.md:

Research paper for ByteRover - an agent-native memory architecture where the LLM itself curates, structures, and retrieves knowledge through a hierarchical Context Tree, requiring zero external infrastructure. --- paper/README.md

Current public tool mode supplies a countercase: caller authoring plus deterministic retrieval. Internal model curation exists independently. Neither source route proves every instance of the broader architectural account or zero model-service dependence.

EPI-CLM-query-answer

Label: Query synthesized answer

Claimed operation: query provides a synthesized answer, contrasted with ranked search. No distinct supplied claim record; closest RT-CLM-project-memory does not make this synthesis assignment. SRC-1 doctrine/design, src/oclif/commands/search.ts:

Use this for structured results with file paths. Use "brv query" when you need a synthesized answer. --- src/oclif/commands/search.ts

The implemented tool-mode retrieval in RT-RTE-query-retrieval leaves synthesis to the caller. The description may describe the overall caller workflow but does not establish target-side answer generation on that route.

EPI-CLM-benchmark-results

Label: Benchmark answer accuracy

Claimed operation: production-codebase memory evaluation with reported LoCoMo 96.1% and LongMemEval-S 92.8% LLM-judge answer accuracy. Distinct evaluation claim from RT-CLM-project-memory; no supplied record declares its protocol. SRC-1 reported operation, paper/README.md:

LoCoMo - 96.1% overall accuracy (1,982 questions, 272 docs).

LongMemEval-S - 92.8% overall accuracy (500 questions, 23,867 docs).

All metrics are LLM-as-Judge accuracy (%). --- paper/README.md

SRC-1 reported operation/design, paper/main.tex:

where a judge model evaluates each generated answer against the ground truth and returns a binary correctness label. For our own evaluations, we use Gemini~3 Flash as the judge with a separate justifier model (Gemini~3.1 Pro) that synthesizes an answer from retrieved context before scoring. --- paper/main.tex

Target/domain: answers on the named conversational benchmarks. Reference: benchmark ground truth. Evaluator: stated Gemini judge, distinct justifier. Result consumer evidenced here is reported comparative scoring, not an implemented admission gate for stored topics or a live policy update. No supplied question-level run artifact establishes those judgments actually occurred under the frozen tool-mode implementation.

The paper reports selective query-time ablations, a stronger design claim than an uncontrolled comparison but not registered causal-experiment evidence. Its tier-removal contrast changes both retrieval routing and the separate justifier, so attribution is no finer than that treatment bundle. Other LongMemEval published comparisons use differing backbone/judge configurations. The write-path limitation is explicit:

Since the curated \contexttree{} is held constant across all ablation conditions, their contribution cannot be isolated through query-time ablation on a static benchmark. --- paper/main.tex

All queries are routed to the full agentic loop (Tier~4), bypassing Tiers~0--3 and the separate justifier model --- paper/main.tex

Thus attributed answer accuracy does not independently establish factual reliability of each retained claim, replay safety, a causal curation/lifecycle improvement, or current-route transfer.

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member