Arsumbris epistemic report

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

Source-and-claim boundary

The question is whether typed knowledge, diagnostics and agent governance establish truth-bearing knowledge or primarily organize content and control actions. The evidence boundary is the documented platform plus the implemented device seeder, not the sibling engine, host, agent kernel or knowledge packages. See the boundary and exclusions and Source register. All source findings below refer to SRC-1 at the frozen commit. No execution evidence or causal comparison is supplied. Runtime findings are provisional; direct reads below check alternate operations and claims against the same frozen source.

The source's consequential claims concern a typed graph, actionable diagnostics, governance and reusable knowledge/capabilities: EPI-CLM-typed-knowledge, RT-CLM-diagnostic-saves, RT-CLM-mediation-floors, and RT-CLM-user-privileges. The entry documentation does not itself supply substantive research claims to evaluate. Knowledge-package names are not evidence of their contents or evaluators.

Entry point or operation	Source path directly read	Covering records / limits
Install prompt, package acquisition, release claims and user launch handoff	README.md, INSTALL.md, VERSION.md, MIGRATION.md	RT-RTE-device-seeding, EPI-RTE-source-ingest; release testing is reported operation only.
Sole tracked executable CLI: option dispatch, discovery, parse/merge/render, dry run/write	scripts/seed-device-config.py, read completely	RT-RTE-device-seeding, EPI-OBJ-registry-paths; implementation supports producer semantics, not successful later launch.
Engine daemon start/status/stop, reads, mutations, subscriptions, dependency/workspace indexing	au-engine.md, MIGRATION.md	EPI-RTE-graph-index, EPI-RTE-type-diagnostics; engine implementation unavailable in this boundary.
Host launch, file read/write, config save, selection, intents, focus/view state, preview/theme and unmount cleanup	au-host.md	RT-RTE-host-edit, EPI-RTE-graph-index; layout/preview cleanup is not withdrawal of knowledge. Actual consumers remain uninspected.
MCP start, adapter launch, tools and hooks, profiles, skills/injects	au-mcp.md	RT-RTE-mediated-action, RT-BAP-profile-guidance, EPI-RTE-hook-lineage; direct reads confirm observer, stamper and session-start operations as well as mediator.
Executable plugin/projection discovery, rejection, rebuild/reload and arbitrary effects	au-mcp.md, au-host.md, MIGRATION.md, SAFETY.md	RT-RTE-extension-load; no content-warrant evaluator follows from SDK compatibility.
File-to-source-note ingest and tree research, rules/governance/writing knowledge packages	INSTALL.md	EPI-RTE-source-ingest; only package descriptions accessible within scope. Missing transformations and acceptance consumers prevent a whole-platform negative about knowledge production.
Harness inference and provider state	INSTALL.md	RT-CMP-agent-harness; no model/version/parameter-update evidence, preventing model-learning or internal-reasoning conclusions.

This is a material-operation coverage account, not a whole-system implementation audit. Direct source reads add no repository or source ID. The banner remains excluded; licensing is nonmaterial here. Actual knowledge bodies, provenance-preserving ingest, evaluator implementations, exact guidance delivery, acceptance records, rejection/withdrawal consumers and later reliance traces are missing. Respectively these gaps prevent semantic classification of knowledge updates, implemented warrant judgments, evidence of compliance, accepted-claim attribution and exercised lifecycle-integration conclusions. None is converted into evidence that the platform lacks those routes.

Epistemic-object inventory

Fields owned by supplied records say see; the rows identify material distinctions rather than redefine those records.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-knowledge-files	Potential assertions in Markdown/YAML bodies; no concrete research proposition individuated	Typed knowledge substrate	SRC-1 au-engine.md; see record	File typing and storage do not warrant assertions.
EPI-OBJ-type-definitions	Type constraints are symbolic policy; diagnostics can assert conformance/nonconformance in that formal domain	Define fields, references and body shape	SRC-1 au-engine.md; see record	Constraint validity is not source truth or encoding fidelity.
EPI-OBJ-session-ledger	Events may assert that actions occurred; exact fields not determinable	Observation	SRC-1 au-mcp.md; see record	No event schema or downstream replay consumer.
EPI-OBJ-write-metadata	Metadata assertions not individuated	Governed-write provenance	SRC-1 au-mcp.md; see record	No stamp schema or authenticity check.
EPI-OBJ-registry-paths	Name/location associations usable as propositions about local installation; also symbolic launch configuration	Bootstrap resolution	SRC-1 scripts/seed-device-config.py; see record	Stale entries preserved; parser supports a subset, not arbitrary YAML.
RT-OBJ-agent-profile	Configuration, not a knowledge claim; see record	Session selection	SRC-1 au-mcp.md; see record	Guidance bodies separate and unavailable.
RT-BAP-profile-guidance	Guidance/context may include claims, but actual content not determinable; see record	Natural-language instruction to harness	SRC-1 au-mcp.md; see record	Delivery neither establishes truth nor obedience.
RT-RTE-extension-load	Executable definitions are capability policy, not accepted propositions; see record	Admit tools/views	SRC-1 MIGRATION.md, SAFETY.md; see record	Compatibility and privileges differ from knowledge authority.
RT-CMP-agent-harness	Produced answer content and distributed-parametric state uninspected; see record	Optional inference	SRC-1 INSTALL.md; see record	No concrete generated proposition or provider-state evidence.

Authority-route ledger

Each function is separate even where it cites the same canonical container. All possible results below are source-supported capacities or declarations, not observed outcomes. See defers canonical route fields to the cited record; this ledger owns architectural status and the local epistemic assessment.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
EPI-RTE-graph-index	content transformation	doctrine only	EPI-OBJ-knowledge-files	truth-apt transformation: acquisition/import	Mounted files into graph	Parser/indexer, declared typed-file domain	Workspace start/file change; see record	Queryable graph; source warrant unknown	No inspected engine consumer	Access and representation, not endorsement	Supplies graph to UI/tools	Graph clients, versioned reads/subscriptions, informational, live workspace	SRC-1 au-engine.md	EPI-CLM-typed-knowledge	none	Semantic fidelity and source lineage unverified.
EPI-RTE-type-diagnostics	check/evidence production	doctrine only	EPI-OBJ-type-definitions	no content change	Field/ref/body conformity, not truth of prose	Engine type rules, formal constraint domain	Index/read changes; exact schedule uninspected	Diagnostics on bad data	Declared informational feedback	Conditional structural conformance only	Makes errors available for response	Human/agent, diagnostics, advisory, current graph	SRC-1 au-engine.md	RT-CLM-diagnostic-saves	none	No execution or evaluator implementation.
RT-RTE-host-edit, RT-RTE-mediated-action	retention	doctrine only	EPI-OBJ-knowledge-files	truth-apt transformation: indeterminate	Supplied content into persistent files	Writer/API; not an evidence-consuming truth decision	Save/tool invocation; see records	Retained edits despite diagnostics	Declared write path, not inspected	No semantic acceptance license established	Files may affect later graph consumers	Engine/UI/tools, file edits, operational, until later edit/delete	SRC-1 au-host.md, au-engine.md, au-mcp.md	RT-CLM-diagnostic-saves	none	Human initiation does not prove per-claim approval. Relation could be reshaping, derivation or conjecture; actual input/output needed.
RT-RTE-mediated-action	check/evidence production	doctrine only	EPI-OBJ-knowledge-files	no content change	Prior read/tool prerequisite witness	hasRead/tool-precondition, action-history domain	Before governed action; see record	Allow/deny/inject/ask	Declared mediator decision	No truth license: reading is not validating	Produces pending-action decision	Mediator to kernel, decision channel, advisory until consumed, pending action	SRC-1 au-mcp.md	RT-CLM-mediation-floors	none	Read history accuracy/lifetime uninspected.
RT-RTE-mediated-action	operational admission/selection/consumption	doctrine only	RT-OBJ-agent-profile	no content change	Permit or prevent action	Kernel/adapter consume mediator decision	Mediate before act; unreachable daemon fails open	Action withheld or allowed by declaration	Intended enforcing policy only on covered path	Operational permission, not accepted knowledge	Allows/blocks governed write; no OS-wide control	Tool executor, mediation/redirect, enforcing conditionally, one action	SRC-1 au-mcp.md, SAFETY.md	RT-CLM-mediation-floors	Fail-open countercase limits universal reading	Actual grants, human ask handling and bypass effects uninspected.
EPI-RTE-hook-lineage	lineage/freshness/recovery	doctrine only	EPI-OBJ-write-metadata	truth-apt transformation: indeterminate	Metadata folded into governed write	Stamper described as trustworthy; trust derivation/domain not specified	Write-time hook; see record	Metadata on governed content	Declared write augmentation	Cannot establish authenticity or source warrant from adjective	Changes provenance received by later readers	Governed writer/readers, stamped file metadata, informational, retained-file horizon unspecified	SRC-1 au-mcp.md	none	none	No metadata schema or evaluator; cannot choose import/derivation/conjecture.
EPI-RTE-hook-lineage	retention	doctrine only	EPI-OBJ-session-ledger	truth-apt transformation: acquisition/import	Pipeline events into ledger	Observer cannot deny; append stage	After act by declaration	Session events retained, source warrant unknown	Declared append, durability uninspected	Activity record, not acceptance record	No future policy update consumer established	Session ledger, event append, recording only, session horizon unspecified	SRC-1 au-mcp.md	none	none	Ledger replay/revision/withdrawal not established.
RT-BAP-profile-guidance	operational admission/selection/consumption	doctrine only	RT-OBJ-agent-profile	no content change	Select guidance/context and visible tools	Profile/adapter, session configuration domain	Trigger/session start; see record	Materialized skills/injects/spawn command	Intended visibility control/instruction	No truth endorsement or compliance license	Changes available tools/context	Harness, materialized text/profile surfaces, advisory text and declared visibility control, session	SRC-1 au-mcp.md	EPI-CLM-typed-knowledge	none	Graph-walk omission/transformation and model consumption uninspected.
RT-RTE-extension-load	check/evidence production	doctrine only	RT-RTE-extension-load	no content change	Runtime metadata/contract compatibility	Loader, SDK integer and concrete-type metadata domain	Discovery/start/load	Mismatch or missing/ambiguous declaration rejected	Declared compatibility result	Interface eligibility only; not code safety or content truth	Supplies load/refusal decision	Loader, metadata comparison, enforcing by declaration, current load	SRC-1 MIGRATION.md	RT-CLM-user-privileges	none	No loader implementation.
RT-RTE-extension-load	operational admission/selection/consumption	doctrine only	RT-RTE-extension-load	non-truth-apt policy/content update: executable capability admission	Install/load code	Loader consumes compatibility result; installer chooses source	Install/rebuild/reload	Tools/views gain execution privileges	Declared load/refusal	Not epistemic acceptance	Admitted code can run commands/read/write/network	Host/kernel, executable module, permissive, until unload/reload	SRC-1 SAFETY.md, MIGRATION.md	RT-CLM-user-privileges	none	Audit recommendation is not a supplied audit result.
RT-RTE-device-seeding	content transformation	implemented	EPI-OBJ-registry-paths	truth-apt transformation: acquisition/import	Existing config/manifests into selected registry rows	Text parsers/discovery/merge, local path domain	CLI dispatch; see record	Associations selected/rendered; input warrant not upgraded	Producer code implemented	Local parser/discovery result only	Emits launch configuration for documented readers	Seeder then host/engine, YAML, permissive configuration, until manual change	SRC-1 scripts/seed-device-config.py	none	none	Concrete lossy parser counterexample in object record; downstream readers absent.
RT-RTE-device-seeding	retention	implemented	RT-OBJ-device-config	no content change	Rendered configuration to device files	Caller --write, deterministic file IO	After dry-run branch; see record	Sequential registry rewrite and paths create/append	Enforced write flag in inspected CLI	Persistence does not confirm locations still valid	Enables later configured resolution, not verified launch	Device config readers, filesystem, permissive by declared contract, across launches	SRC-1 scripts/seed-device-config.py	none	none	No stale eviction, transaction or automatic rollback.
EPI-RTE-source-ingest	content transformation	doctrine only	EPI-OBJ-knowledge-files	truth-apt transformation: indeterminate	Files to typed source notes	au-ingest implementation unavailable	Package use, timing unknown	Declared source notes	No inspected transformation	Source warrant preservation unknown	Potential later graph/guidance use uninspected	Knowledge-package consumers, typed files, force/horizon unknown	SRC-1 INSTALL.md	EPI-CLM-typed-knowledge	none	Could import/reshape/derive/conjecture; source notes and transform needed.

No separate acceptance or lifecycle-integration route is established by this inspected entry artifact. This is not a system-complete absence claim: opaque knowledge/governance packages could supply such routes. No invented evaluator is assigned to that gap. Unmount subscription disposal, source-file graph refresh, contract refusal and manual config correction have distinct targets; none by itself withdraws a previously accepted claim or its downstream uses.

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-typed-knowledge	Reuse knowledge/capabilities; typed graph checks fields/references and supplies diagnostics	SRC-1 README.md, doctrine/design; see record	EPI-RTE-graph-index, EPI-RTE-type-diagnostics, RT-BAP-profile-guidance	None inspected for knowledge loop	None supplied	No comparison; typedness cannot establish empirical truth	A coherent declared organization/access architecture	Actual semantic production, warranted reliance and benefit unknown; do not inflate “knowledge” into a claim of proven truth.
RT-CLM-diagnostic-saves	Bad data diagnosed without blocking save/read; see record	SRC-1 au-engine.md, doctrine/design	EPI-RTE-type-diagnostics, RT-RTE-host-edit	None inspected	None supplied	Structural scope only	Diagnostic feedback and retention are deliberately separate	Not evidence of failed truth acceptance; no such gate promised here.
RT-CLM-mediation-floors	Always-on floors and fixed pipeline; see record	SRC-1 au-mcp.md, doctrine/design	RT-RTE-mediated-action	None inspected	None supplied	Unreachable daemon, host writes and arbitrary code limit coverage	Conditional operational governance, not warrant	Fail-closed load and fail-open mediation coexist; universal enforcement unsupported.
RT-CLM-user-privileges	Installed code has user privileges, no sandbox; see record	SRC-1 SAFETY.md, doctrine/design	RT-RTE-extension-load	None inspected	None supplied	Warning, not independent security test	Compatibility admission is not safe-code or true-content certification	Actual deployment envelope unobserved.

Bounded conclusion

Arsumbris declares retention of authored files, live graph retrieval and structural diagnostics, plus profile-selected context and operational action gates. Those are useful but different functions. Checks target field/reference/body conformity, prior-action prerequisites and SDK eligibility. No inspected check licenses the truth of natural-language knowledge. Diagnostics explicitly leave save/read possible; read-before-write licenses an action, not a proposition. The admission force and horizons are the separate paths in the ledger, not a shared truth oracle.

The only inspected executable route imports and persists local installation associations. It preserves existing selections even when discovery conflicts, warns rather than repairs them, and implements only a text subset of YAML. Thus the implementation licenses claims about its own selection/rendering, not freshness, source authenticity or successful downstream launch. The concrete parser deduction below supplies a source-supported countercase to general semantic-preservation claims.

Files entering the graph are acquired content with unknown source warrant. Graph indexing declares organization and access, not novel accepted propositions. File edits, stamped metadata and external source-note generation have indeterminate semantic relations at this boundary. No concrete warranted premises and checked derivation of a domain claim, non-entailed conjecture, evidence-consuming acceptance criterion for an intended knowledge use, or post-acceptance integration is established. The missing knowledge-package implementations and claim instances prevent stronger positives and whole-platform negatives alike. Operational use before such acceptance remains operational use.

The observer/ledger route records events; the stamper claims trustworthy metadata without exposing its warrant mechanism. Neither a trace-fed policy update nor later consumer reliance on a criticism is supplied. Guidance delivery and extension loading could change behavior, but do not establish reliable compliance, reflection, learning, autonomous theory construction or exercised self-improvement. These properties remain independently unestablished, not jointly disproved. The optional model dependency's mutability, exact version and endpoint behavior remain uninspected as recorded in RT-CMP-agent-harness; nothing about provider learning follows from static file or SDK inspection.

No runtime execution, accepted candidate trace or intervention comparison grounds exercised operation, benefit or causal attribution. This partial documentary loop supports the separate route conclusions above, not an unqualified epistemic verdict on the full platform.

Shared records

Operative objects

EPI-OBJ-knowledge-files

Label: Markdown and YAML knowledge files

Source-native Markdown/YAML repositories store authored content on disk; the engine declares an in-memory graph over mounted repositories. Natural-language bodies carry possible assertions; YAML/frontmatter and reference syntax are symbolic at graph consumers. Generic content does not establish a particular truth-apt proposition. The closest supplied RT-RTE-host-edit and RT-RTE-mediated-action describe producers/consumers, not a separate declaration of the knowledge substrate. RT-OBJ-device-config is different: pre-graph launch configuration, not knowledge files. SRC-1 au-engine.md, doctrine/design:

It indexes every mounted repo (deps) and holds them as one graph in memory. Change a file, and the graph updates. --- au-engine.md

The engine will parse markdown and YAML, and the type system only applies on those filetypes for now. But media files or other assets are also be indexed as untyped members. --- au-engine.md

EPI-OBJ-type-definitions

Label: YAML type definitions and diagnostics

Source-native type definitions are symbolic YAML files with fixed intended constraints for fields, references, inheritance and required body sections. Diagnostic messages are claims about those constraints, not endorsements of file assertions. Their storage/output encoding is uninspected beyond the declarative examples. This separates the constraint/checking part from EPI-OBJ-knowledge-files. The closest supplied RT-CLM-diagnostic-saves is the save contract, not this object. SRC-1 au-engine.md, doctrine/design:

myInteger: Number{integer} # must be whole number myPositive: Number{>=0} # must be positive --- au-engine.md

that file must have the type "someOtherType" required by the field

myReference: [[instance-of-someOtherType]] --- au-engine.md

The nonnegative check admits zero despite the comment “positive”; the explicit expression, not that informal adjective, bounds the documented domain. Even a correct structural result would need faithful encoding of an external claim before it could warrant anything about that claim.

EPI-OBJ-session-ledger

Label: Session ledger events

Part of: RT-RTE-mediated-action

This is the observation-output part of that route, not another mediated-action container. Ledger events are declared symbolic records, but schema, backing store, durability and attribution validation are not supplied. Event storage is session-specific. This differs from natural-language guidance in RT-BAP-profile-guidance and from write metadata below. SRC-1 au-mcp.md, doctrine/design:

• observer — watch the event stream, never deny.
• mediator — intercept a pending action: allow / deny / inject / ask.
• stamper — fold trustworthy metadata into a governed write.
• session-start — compute context to inject in front of the agent at open. --- au-mcp.md

The pipeline passage retained in RT-RTE-mediated-action supplies event append evidence, not evidence of acceptance or replay.

EPI-OBJ-write-metadata

Label: Governed write metadata

Part of: RT-RTE-mediated-action

This is the stamped-output part of the supplied action route, distinct from its session event log. The declared stamper folds symbolic metadata into a governed write, so later file readers rather than an identified ledger replay consumer could receive it. Actual keys, assertion content, storage encoding, authentication and later interpretation remain uninspected. SRC-1 au-mcp.md, doctrine/design; the hook-shape passage retained in EPI-OBJ-session-ledger supplies the evidence. The adjective “trustworthy” is a declared property without a supplied verification mechanism. No inference of source authenticity or warranted premises follows.

EPI-OBJ-registry-paths

Label: Registry name and path associations

Part of: RT-OBJ-device-config

The name/location and template-path part of device configuration is symbolic YAML on disk. Its consumer uses paths to locate code/repositories; an association can also assert a local installation fact. It differs from the parent's tool-path map, while sharing its bootstrap producer RT-RTE-device-seeding. SRC-1 scripts/seed-device-config.py, implementation, supplies the parse/serialization limits:

def yaml_unquote(s): """Inverse of yaml_scalar: unquote+unescape a quoted scalar, else strip an inline comment.""" s = s.strip() if len(s) >= 2 and s[0] == '"' and s[-1] == '"': return _yaml_unescape(s[1:-1]) --- scripts/seed-device-config.py

# a bare scalar never carries a space or '#' (we quote those), so a '#' here is a comment.
return s.split("#", 1)[0].strip()

--- scripts/seed-device-config.py

    if body[i] == "\\" and i + 1 < len(body):
        out.append(body[i + 1]); i += 2

--- scripts/seed-device-config.py

if unsafe:
    return '"' + v.replace("\\", "\\\\").replace('"', '\\"') + '"'
return v

--- scripts/seed-device-config.py

Concrete deduction without target execution: an existing registry entry with name: demo and path: /tmp/a#b is parsed as association demo → /tmp/a, since the bare scalar is split at #. Regeneration therefore writes the shortened path, not the original text. Within the script's own quoted subset, value /tmp/a#b is instead rendered as "/tmp/a#b" and unquoted back intact. This countercase rules out treating arbitrary imported config text as lossless reshaping; it does not claim that an actual installed config had either value. Similarly _yaml_unescape copies the character following each backslash, rather than implementing all YAML escapes. Existing selections survive merge without validating continued location, as the parent's retained setdefault passage shows. No endpoint beyond the producer was inspected.

Routes

EPI-RTE-graph-index

Label: Mounted graph indexing

The engine is the declared owner: entry repository plus dependencies/workspace members → parser/indexer → live graph → versioned reads/subscriptions to UI and agent tools. This is acquisition and organization of file content, not an evidence-consuming acceptance decision. Closest supplied RT-RTE-host-edit and RT-RTE-mediated-action are clients/writers, not the graph-index route; it is not wholly contained in either. The author supplies content, workspace declarations select mounting, and filesystem changes trigger updates; no per-claim approval is described. Type diagnostics do not veto this route. No answer oracle or improvement curriculum is established. SRC-1 au-engine.md, doctrine/design; supporting indexing passage is retained in EPI-OBJ-knowledge-files.

• Immediate return: Declared graph reads observe a version and subscriptions stream changes.
• Later read-back: Changed files are declared to reach later graph reads; exact cache/reload implementation uninspected.
• Delegated visibility: Consumers see mounted graph content; session/sub-agent isolation uninspected.
• Selection predicate: Entry repo, dependencies and workspace edit/discover members; queries select graph material.
• Invalidation or expiry: File change updates graph by declaration; actual invalidation and stale snapshot semantics uninspected.
• Activation or effect: Declared graph availability to clients, with no supplied run.
• Evidence limits: Entry documentation only; engine and clients' actual transformations outside frozen source.
• design conclusion status: claimed
• implementation conclusion status: uninspected
• operation conclusion status: uninspected

• The daemon serves consumers over a Unix socket.
• Reads observe a version.
• Mutations go through one guarded write path.
• Subscriptions stream changes. --- au-engine.md

EPI-RTE-type-diagnostics

Label: Typed content diagnostics

The engine evaluates file fields, references and body structure against EPI-OBJ-type-definitions, generating diagnostic feedback. This route is distinct from the supplied RT-CLM-diagnostic-saves: that claim establishes permissive saving, not a route declaration for evaluation. Content authors propose; type authors define criteria; engine judges formal conformity. A user or model may respond, but no mandatory revision or acceptance follows. There is no external truth oracle. Operating mode is open editing; reported diagnostics are a possible correction trigger, not an implemented improvement loop. The evaluator's license is limited to faithfully encoded constraints and references, not empirical truth. Supporting passages: EPI-OBJ-type-definitions and RT-CLM-diagnostic-saves, SRC-1 au-engine.md, doctrine/design.

• Immediate return: Diagnostic feedback by declaration; exact message interface uninspected.
• Later read-back: uninspected — diagnostic persistence and later consumption are not specified; changed file graph reads belong to the separate indexing route.
• Delegated visibility: Agents and UI are intended consumers; actual delivery and omissions uninspected.
• Selection predicate: Mounted Markdown/YAML content and its declared type, fields and references.
• Invalidation or expiry: uninspected — no diagnostic clearing/withdrawal implementation supplied.
• Activation or effect: Feedback can inform editing; no enforced rejection or observed correction.
• Evidence limits: Examples and contract only, not an implementation or test result.
• design conclusion status: claimed
• implementation conclusion status: uninspected
• operation conclusion status: uninspected

EPI-RTE-hook-lineage

Label: Hook metadata and event recording

Part of: RT-RTE-mediated-action

This is the lineage/observation part of that route; checking prior reads and consuming mediation decisions remain in the supplied container. Hooks watch events or fold metadata into a governed write; the pipeline appends events to a session ledger. Kernel and hook code are declared owners, while metadata proposals, stamp vetoes and field interpretation are unavailable. Observer has explicitly no denial authority. It is an open-work recording route, without an answer oracle, acceptance criterion or improvement trigger. Provenance and replay safety cannot be inferred from recording. Supporting minimum passages are retained in EPI-OBJ-session-ledger and RT-RTE-mediated-action, SRC-1 au-mcp.md, doctrine/design.

• Immediate return: Governed write may receive metadata; events are declared appended after act.
• Later read-back: uninspected — no ledger replay consumer or later metadata-validation route supplied.
• Delegated visibility: uninspected — multi-session daemon does not specify cross-session ledger access.
• Selection predicate: Hook shapes and pipeline events; exact event/stamp filter uninspected.
• Invalidation or expiry: uninspected — no ledger retention, stamp expiry or provenance-withdrawal rule supplied.
• Activation or effect: Declared event retention and metadata augmentation; no observed activation or demonstrated reliance.
• Evidence limits: Kernel/hook implementations and event examples absent from this source boundary.
• design conclusion status: claimed
• implementation conclusion status: uninspected
• operation conclusion status: uninspected

EPI-RTE-source-ingest

Label: File source-note ingestion

The entry installation describes the sibling au-ingest package as converting files to typed source notes. No counterpart route is declared in the supplied runtime; this is distinct from RT-RTE-device-seeding, which produces installation config rather than source notes. Input files → opaque package transform → typed source notes is the declared progression. The package owns the transformation; human/model proposal and veto roles, truth checks, recovery and guidance bodies are uninspected. No oracle or operating mode beyond a knowledge-work capability is established. No semantic-preservation, derivation, conjecture or acceptance classification can be resolved from the name. SRC-1 INSTALL.md, doctrine/design:

• au-ingest (files → typed source notes), and the au-tree-research feature package. --- INSTALL.md

• Immediate return: uninspected — source-note output is described but no invocation/result contract supplied.
• Later read-back: uninspected — knowledge-package writers and consumers are outside this entry source.
• Delegated visibility: uninspected — no worker or delivery boundary supplied.
• Selection predicate: uninspected — source file selection and filtering not specified.
• Invalidation or expiry: uninspected — source changes, note replacement and withdrawn-source handling unspecified.
• Activation or effect: Declared conversion capability only; no exercised candidate or checked result.
• Evidence limits: Package description does not expose evaluator, acceptance criterion, intended reliance scope or source lineage.
• design conclusion status: claimed
• implementation conclusion status: uninspected
• operation conclusion status: uninspected

Claims

EPI-CLM-typed-knowledge

Label: Reusable typed knowledge

The source claims composable knowledge/capability repositories, a typed graph, field/reference checks and diagnostics for agents. This is a knowledge-organization and capability claim, not an explicit claim that every saved assertion is true. It is distinct from RT-CLM-diagnostic-saves, which describes the consequence of bad data, and RT-CLM-mediation-floors, which describes action governance. SRC-1 README.md, doctrine/design:

Repos can depend on each other, letting you reuse knowledge and capabilities across projects.

The engine reads them as one typed graph, checks fields and references and gives agents diagnostics they can act on. --- README.md