Archie epistemic report

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

Source-and-claim boundary

The question is whether Archie's agent-launch profiles acquire, produce, check or accept truth-apt content, and how their results affect later behavior. Scope and exclusions are as in boundary; evidence identities are in its Source register. This is a runtime-client subsystem, not the external agent loop or the whole desktop distribution. The assistant plan is design context, not an implemented evaluator.

Direct inspection used SRC-1 at the frozen revision: the complete profile module, shell command loader, Python CLI, assistant epic and agent-backed work item, plus the profile-reference section. The searches below covered operation names, hooks, tools, acceptance, rejection, rollback, provenance and confidence in those files. Deployment, shell initialization and override behavior are supplied runtime findings, not fresh source reads. No source was executed. There are no supplied traces, candidate answers, evaluation results or causal comparisons.

Entry point or operation	Source path	Coverage and records	Limit / conclusion prevented
Seven profiles, four aliases, direct helper call	deployment-packages/local/lib/zsh/commands-agents.sh	Direct complete read; RT-RTE-profile-launch, RT-RTE-default-update	External CLI semantics and model outputs uninspected
Shell load and deployment	deployment-packages/local/lib/zsh/commands.sh, deployment-packages/home/.zshrc, scripts/install.sh	Loader read directly; remaining chain supplied by RT-RTE-shell-load	No installed-environment evidence
Direct invocation and overrides	deployment-packages/home/.zshrc	Supplied RT-RTE-direct-handoff	Prevents universal profile-gate claims
Python CLI dispatch	src/archie/cli.py	Direct complete read; registers applet, downgrade, GUI and system groups, not an agent operation	Desktop implementations excluded; no whole-system negative follows
Documented profile guarantees	deployment-packages/local/lib/zsh/README.md	Direct profile-section read; EPI-CLM-profile-contract	Documentation differs from executable selectors
Documentation guidance, provenance, confidence and fallback	docs/epics/assistant.md, docs/work-items/assistant-03-agent-backed-documentation-guide.md	Direct complete reads; RT-CLM-assistant-guide	Planned; no candidate, check algorithm or consumer implementation
External tools, hooks, model evaluation and context management	External omp and providers	Opaque dependencies; RT-CMP-model-services	No external implementation or traces; cannot judge their epistemic routes

There is no knowledge-production guarantee in the inspected launcher. The consequential claims are the operational profile contract and the planned documentation guide cited above. Missing external runtime/provider sources prevent attribution of any internal acquisition, inference, evaluation, acceptance or tool restriction to Archie. Missing executions prevent claims of exercised authority or benefit. Missing candidate content prevents classification of any actual answer as imported, entailed or conjectural; answer content is not determinable. Excluded contributor workflows and unrelated desktop maintenance cannot support conclusions about this subsystem's knowledge production.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-model-roles	None established; symbolic role configuration, not a proposition accepted as true	see cited record	SRC-1, deployment-packages/local/lib/zsh/commands-agents.sh	Getter origin and durable backend opaque; existing fields are not evidence claims merely because they persist
RT-RTE-profile-launch	No truth-apt candidate in preset model/effort flags; forwarded caller input may contain one	see cited record	SRC-1, deployment-packages/local/lib/zsh/commands-agents.sh	Flags are symbolic controls. No actual caller payload or downstream answer is supplied
RT-CMP-model-services	Not determinable within accessible service state	see cited record	SRC-1, deployment-packages/local/lib/zsh/commands-agents.sh	Service computation may produce truth-apt content, but no model state or output is inspected; per-selector mutability and version limits remain in cited record
RT-CLM-assistant-guide	Intended documentation/architecture explanations; no candidate instance supplied	see cited record	SRC-1, docs/work-items/assistant-03-agent-backed-documentation-guide.md	Design-only context and provenance requirements; no operative answer object or evidence attachment

Authority-route ledger

Rows separate the functions performed by existing records; they do not allocate duplicate route identities. Fields owned by those records are incorporated by citation. Architectural statuses and content relations below are this report's 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
RT-RTE-default-update	content transformation	implemented	RT-OBJ-model-roles	non-truth-apt policy/content update: replace default selector	Current JSON to updated configuration	Symbolic jq assignment; ordinary JSON object domain	see route	see route	see route	No factual warrant conferred	see route	see RT-BAP-role-arguments	SRC-1, deployment-packages/local/lib/zsh/commands-agents.sh	EPI-CLM-profile-contract	Conditional preservation, not universal durability	Direct helper calls also afforded; no target execution
RT-RTE-default-update	retention	implemented	RT-OBJ-model-roles	no content change	Transformed value to config setter	Setter return status; backend opaque	see route	see route	see route	Retention is not acceptance of a proposition	see route	see RT-BAP-role-arguments	SRC-1, deployment-packages/local/lib/zsh/commands-agents.sh	EPI-CLM-profile-contract	Persistence requested, not observed	No read-back verification, lock, history or compensating restore in helper
RT-RTE-profile-launch	operational admission/selection/consumption	implemented	RT-OBJ-model-roles	no content change	Helper completion to process launch	Nonzero helper return stops profile; shell status domain, not answer truth or provider readiness	see route	see route	see route	None for generated claims	see route	see RT-BAP-role-arguments	SRC-1, deployment-packages/local/lib/zsh/commands-agents.sh	EPI-CLM-profile-contract	Gate is operational, not epistemic acceptance	Pipeline status need not expose every getter failure; no schema/read-back check
RT-RTE-profile-launch	behavior/policy adaptation	implemented	RT-OBJ-model-roles	non-truth-apt policy/content update: fixed role selection	Invoked function to runtime role request	Caller chooses preset; no evidence-responsive ranking or answer oracle	see route	see route	see route	Model names confer no answer warrant	see route	see RT-BAP-role-arguments	SRC-1, deployment-packages/local/lib/zsh/commands-agents.sh	none	none	Appended arguments and external parser bound effective selection; not evaluation-driven adaptation
RT-CLM-assistant-guide	other — intended documentation guidance and provenance presentation	doctrine only	see claim	truth-apt transformation: indeterminate	Intended documentation/context to explanation	No named answer evaluator or acceptance criterion implemented in inspected subsystem	Planned, optional mode	Intended constrained guidance; no observed answer	No implemented assistant force found in inspected profile/dispatch paths	No established license for factual reliance	None established for assistant restrictions	Intended assistant-to-user guidance, advisory; implementation and horizon undecided	SRC-1, docs/work-items/assistant-03-agent-backed-documentation-guide.md	RT-CLM-assistant-guide	Plan must not be transferred to launcher	No candidate to distinguish import, reshaping, derivation or conjecture

For the deterministic transformation, consider the hypothetical ordinary object {"default":"old","plan":"keep"} and helper argument openai-codex/gpt-6-luna:medium. The inspected .default = $default expression yields {"default":"openai-codex/gpt-6-luna:medium","plan":"keep"} modulo JSON serialization. The whole resulting value goes to the setter. This is an implementation-layer deduction, not a target execution or an entailed new factual proposition. It supports preservation of non-default keys in the transform's domain, not source truth, durable atomicity or protection from concurrent whole-object writes. The minimum implementation passage is supplied by RT-RTE-default-update.

The strongest countercase to a purely passive-launcher account is that it changes configuration before launch and later helper invocations request that value again. Hence operational retention and authority are real implemented requests, even though factual acceptance is not established. Conversely, setter success followed by launch failure does not restore the previous selector. Direct invocation bypasses the profile gate. These cases, supported by the same helper and RT-RTE-direct-handoff, prevent treating status gating as an answer validator or a universal runtime safety policy.

No wrapper answer check, rejection, withdrawal, cleanup of accepted claims or post-acceptance lifecycle integration was found in the complete profile module. This is a bounded inspection finding, not an absence finding about external omp. The only inspected failure disposition is shell return-status handling, whose target is command continuation rather than a truth-apt candidate. The planned provenance requirement is not itself an evaluator:

• Define how the assistant should present provenance, confidence, and fallback behavior when the agent response goes beyond the local docs. --- docs/work-items/assistant-03-agent-backed-documentation-guide.md

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-profile-contract	see claim	SRC-1, deployment-packages/local/lib/zsh/README.md; doctrine/design	see claim	RT-RTE-default-update, RT-RTE-profile-launch	None supplied	No comparison; external storage and parser required	Source supports conditional JSON preservation and forwarding, not unconditional durable preservation	Some documented GPT selectors differ from code; failures/concurrency bound preservation
RT-CLM-assistant-guide	see claim	SRC-1, docs/epics/assistant.md, docs/work-items/assistant-03-agent-backed-documentation-guide.md; doctrine/design	see claim; provenance/confidence/fallback remain tasks to define	No implemented assistant route established; profile launch is not the planned guide	None supplied	No comparison, acceptance procedure or implemented restriction in inspected mechanism	A planned constrained documentation mode, not exercised factual evaluation	No implementation/operation support for assistant warrant; plan is explicitly future-facing, not a failed deployed guarantee

Bounded conclusion

Archie implements operational model-role selection and a configuration retention request, not an inspected truth-apt acceptance pipeline. The fixed default update, helper-status gate and CLI presets affect the external runtime through symbolic arguments and config-setting calls; their scope, consumer and horizons are in RT-BAP-role-arguments. They confer no evidence-based authority on any answer. Their actual activation remains unobserved.

No truth-apt acquisition with preserved source warrant, accepted ampliative claim, checked entailment or post-acceptance integration is established in the selected shell mechanism. External caller input, model state and generated answers remain uninspected, not presumed non-truth-apt. Thus neither a system-wide negative nor a positive knowledge-production conclusion follows. The concrete JSON deduction concerns control configuration only.

The assistant's documentation corpus, contextual guidance, provenance and confidence presentation remain planned. Without an answer instance and evidence-consuming decision against a criterion for an intended use and scope, there is no basis to classify its future outputs as accepted knowledge. Presentation of provenance would not by itself establish source truth or acceptance. The plan and source-level configuration effects therefore retain separate statuses.

The profile path changes behavior requests without evaluated content: it is fixed caller-selected adaptation, not evidence of learning, reflection, autonomous theory building or self-improvement. External model processing and any parameter changes are independently uninspected as recorded in RT-CMP-model-services; symbolic selector changes do not establish weight updates. Missing executions also prevent benefit or causal attribution. None of these limits erases the implemented configuration path or implies a single epistemic verdict for Archie and its dependencies.

Shared records

Claims

EPI-CLM-profile-contract

Label: Persistent profile configuration contract

The profile reference claims that functions persist their default, preserve non-default roles and forward appended arguments. This is an operational contract, not a knowledge-production claim. SRC-1, deployment-packages/local/lib/zsh/README.md, doctrine/design:

Persistent omp:* profile functions set their own default model role before starting omp; they preserve configured non-default roles and forward appended arguments. Shorthand aliases delegate to the corresponding profile function. --- deployment-packages/local/lib/zsh/README.md

The closest supplied claim, RT-CLM-assistant-guide, concerns a distinct planned assistant mode. This claim instead concerns the deployed shell-profile contract implemented by RT-RTE-default-update and RT-RTE-profile-launch; those routes supply the implementation quotations. No duplicate claim counterpart was found in the supplied runtime declarations. The universal preservation reading is limited to ordinary input shape, external storage success and absence of conflicting concurrent writes. This record does not assert an actual persisted instance or a new epistemic evaluator.