Claude Context Guard epistemic report

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

Source-and-claim boundary

The question is what licenses later reliance on project records, learned behaviours, helpers and code descriptions, rather than merely making them available. Scope and exclusions are see boundary; evidence identities and layers are see Source register. SRC-1 supplies executable shell implementation and model-facing doctrine/design. It supplies no populated project histories, tests of generated helpers, observed host invocation or causal comparison. Model compliance, actual acceptance, successful recovery and improved future action therefore remain unobserved. External research linked by the README is not evidence acquired in this run. Images and host/provider internals remain excluded as in the boundary.

This analyst directly read the installer, settings, both hooks, all six commands, all eight templates and README at the registered revision. The tracked dispatch surface agrees with the supplied runtime: six command prompts, two event hooks, installer and manual-copy entry points; no additional registered tool operation was found in those paths. The table distinguishes these reads from reliance on runtime records. Nothing in the relevant shipped text surface is left unassessed; this is not complete coverage of possible consumer-project or host operations.

Entry point or operation	Source path	Coverage and evidence basis
Install and manual copy	SRC-1, install.sh, README.md, .claude/settings.json	Direct reads; see RT-RTE-install-package. Copy admission and conditional hook activation, not truth validation.
First-run creation and recovery	SRC-1, .claude/commands/start.md	Direct read; see RT-RTE-start-session. Imports human details; later consumes current files and selected archive entries.
Incremental updates and code-index maintenance	SRC-1, templates/CLAUDE.md	Direct read; see RT-BAP-project-rules. Model-operated logging and description revision, not an executable acceptance gate.
Checkpoint, wrap-up, rotation and supersession	SRC-1, .claude/commands/save.md, .claude/commands/end.md	Direct reads; see RT-RTE-save-session, EPI-RTE-feature-check, EPI-RTE-feature-admission.
Scratch-script harvest, helper tests, approval and deletion	SRC-1, .claude/commands/lessons.md	Direct read; see RT-RTE-harvest-lessons, EPI-RTE-lesson-inference.
Code labels, description generation, filtered comparison and restoration	SRC-1, .claude/commands/itemise.md	Direct read; see RT-RTE-itemise-code, EPI-RTE-itemise-filter.
Integrity diagnosis, report retention, selected repairs	SRC-1, .claude/commands/audit.md	Direct read; see RT-RTE-audit-project. Includes archive IDs, dates, contradictions and source/sidecar parity.
Commit reminder	SRC-1, .claude/hooks/pre-commit-check.sh	Direct read; see RT-RTE-commit-reminder. Raw substring produces advice, not a ledger check.
PreCompact backup	SRC-1, .claude/hooks/pre-compact-save.sh	Direct read; see RT-RTE-compact-backup. Existing bytes copied, no knowledge acceptance or automatic later restore.
Retained-state formats	SRC-1, templates/COMMENTS.md, templates/SESSION_LOG.md, templates/RESUME_STATE.md, templates/TASK_REGISTRY.md, templates/DECISIONS.md, templates/LEARNED_BEHAVIOUR.md, templates/FEATURE_LIST.json	Direct reads; see object records below. Templates are not actual histories or QA results.
Host numerical state and compaction	SRC-1, README.md, .claude/settings.json	Supplied host findings: see RT-CMP-claude-code. No accessible parameters, exact model version, model processing trace or compaction policy; truth-apt numerical content and parameter learning are not determinable.

Consequential claims are see RT-CLM-progress-protection, EPI-CLM-diff-record, EPI-CLM-helper-safety, EPI-CLM-archive-truth, EPI-CLM-itemise-preservation and EPI-CLM-context-economy. Their scopes differ: recovering records is not producing justified factual knowledge; navigating code is not proving code descriptions true.

Epistemic-object inventory

Object identities, forms, substrates and quoted evidence are owned by the cited records; see below refers to those fields rather than redeclaring them. Candidate content is scoped to hypothetical consumer entries, not to the empty/example templates.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-comments	That the user said a particular sentence in a session; the sentence itself may be a directive, not a proposition	see record	SRC-1, templates/COMMENTS.md	Verbatim recording preserves attribution if followed, not the truth of user assertions.
EPI-OBJ-session-log	What occurred, what was completed, errors and fixes	see record	SRC-1, templates/SESSION_LOG.md	Model recollection/compression; no actual entry checked.
EPI-OBJ-resume-state	Current approach, in-flight work and next micro-step; clean/unclean marker	see record	SRC-1, templates/RESUME_STATE.md	A clean marker is a procedure result, not proof all work is preserved.
EPI-OBJ-task-registry	Task provenance, status, files and approach	see record	SRC-1, templates/TASK_REGISTRY.md	Done status is deliberately not feature QA.
EPI-OBJ-decisions	Rationale and source assertions; choices/categories also grant policy authority	see record	SRC-1, templates/DECISIONS.md	Binding force is independent of whether the rationale is true.
EPI-OBJ-learned-behaviour	Scoped failure, workaround and optional root-cause explanation	see record	SRC-1, templates/LEARNED_BEHAVIOUR.md	Ordinary save can admit entries without the lesson approval path.
EPI-OBJ-feature-list	A named feature passed human end-to-end verification, or has a reported failing case	see record	SRC-1, templates/FEATURE_LIST.json	No environment, test protocol or observed user judgment supplied.
EPI-OBJ-sidecar-index	What a numbered code block does and when its description was edited	see record	SRC-1, .claude/commands/itemise.md	Symbolic number parity cannot warrant natural-language meaning.
EPI-OBJ-helper-library	None inherent in executable behavior; associated assumptions/specifications may be truth-apt	see record	SRC-1, .claude/commands/lessons.md	No generated code supplied; whether a helper formally encodes a proposition is not determinable.
EPI-OBJ-audit-report	Detected consistency defects, passing checks and proposed fixes	see record	SRC-1, .claude/commands/audit.md	Report is retained diagnosis, not a independently validated account.
RT-OBJ-claude	Principally policy; some factual assumptions about host behavior and recovery	see record	SRC-1, templates/CLAUDE.md	Host delivery and conflicting instructions remain limits.
RT-CMP-claude-code	Not determinable from accessible numerical state	see record	SRC-1, README.md	Version/fixity/mutable endpoint and hidden reasoning uninspected; no inference of absent learning.

Authority-route ledger

Rows split consequential functions even where one supplied route is their valid container. Generic progression, return, read-back, selectors, expiry and evidence limits are see the declaring record. All prompt-directed rows are doctrine only: host execution of Bash instructions is requested, not inspected package automation. Implemented applies only to inspected shell logic. No row claims observed operation. None in epistemic license means the route itself licenses no new truth claim, not that its inputs must be false.

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-start-session	content transformation	doctrine only	EPI-OBJ-comments, RT-OBJ-claude	truth-apt transformation: acquisition/import	User details and custom/existing project records enter model context	see record	see record	Requested reads and user details; source warrant unknown	see record	Attribution can be preserved; endorsement not established	Setup and recovery context made available	see record and RT-BAP-project-rules	SRC-1, .claude/commands/start.md	EPI-CLM-context-economy	Partial loading is not full-context warrant	No authentic consumer data supplied
RT-RTE-save-session	content transformation	doctrine only	EPI-OBJ-session-log, EPI-OBJ-resume-state, EPI-OBJ-task-registry, EPI-OBJ-decisions, EPI-OBJ-learned-behaviour	truth-apt transformation: indeterminate	Conversation/events become summaries, rationale and causal explanations	see record	see record	Requested entries; no populated instance	see record	Summary fidelity and entailment unestablished; root-cause text can add hypotheses	Entries guide later work before distinct factual acceptance	see record and RT-BAP-project-rules	SRC-1, .claude/commands/save.md	none	None; retention is not acceptance	Need inputs and authored entries to classify individual edges
RT-RTE-save-session	check/evidence production	doctrine only	EPI-OBJ-comments, EPI-OBJ-task-registry, EPI-OBJ-decisions, EPI-OBJ-learned-behaviour	no content change	Missing entries and trigger coverage	Foreground model reviews conversation and named checklists; bookkeeping domain	see record	Requested completeness checklist or return to updating	see record	Internal coverage evidence only; not independent factual testing	May delay Git step until each file addressed	Foreground model, checklist prompt, requested continuation gate, this save	SRC-1, .claude/commands/save.md	none	Self-check is not independent verification	Accurate conversation interpretation unobserved
RT-RTE-save-session	retention	doctrine only	EPI-OBJ-comments, EPI-OBJ-session-log, EPI-OBJ-resume-state, EPI-OBJ-task-registry, EPI-OBJ-decisions, EPI-OBJ-learned-behaviour	no content change	File writes and optional explicit-path Git durability	see record	see record	Requested durable state; no observed write	see record	No additional warrant from persistence	Later sessions can select retained state	see record	SRC-1, .claude/commands/save.md	RT-CLM-progress-protection	Retention without acceptance	No multi-file transaction
RT-RTE-save-session	lineage/freshness/recovery	doctrine only	EPI-OBJ-decisions, EPI-OBJ-learned-behaviour, EPI-OBJ-task-registry	truth-apt transformation: non-ampliative reshaping	Full bodies move intact; index titles/categories point to them	see record	see record	Requested rotation, supersession link, resolved LB marker or backlog disposition	see record	Archival preserves existing warrant at best, creates none	Active constraints remain binding; resolved entries need not be read; superseded decisions point to replacements	Later model, main indexes then archive body, requested policy, later sessions	SRC-1, .claude/commands/end.md	EPI-CLM-archive-truth, EPI-CLM-context-economy	Movement does not establish truth	Title relevance selection can omit a needed condition; template category conflict see runtime
RT-RTE-start-session	operational admission/selection/consumption	doctrine only	EPI-OBJ-resume-state, EPI-OBJ-decisions, EPI-OBJ-learned-behaviour, EPI-OBJ-feature-list, EPI-OBJ-task-registry	no content change	Handoff, pending tasks and relevant rules enter later work	see record	see record	Requested summary, retrieval and wait for user go-ahead	see record	Loaded data retain unknown original warrant; human continuation is not per-claim acceptance	Guides next task and forbids contradicting constraints	see record and RT-BAP-project-rules	SRC-1, .claude/commands/start.md	EPI-CLM-context-economy, EPI-CLM-archive-truth	Operational force without factual endorsement	Relevance and compliance are model judgments
EPI-RTE-feature-check	check/evidence production	doctrine only	EPI-OBJ-feature-list	no content change	Named user-facing feature works end-to-end	see record	see record	Human pass verification or broken-case report, not task completion	see record	Bounded user verification, not all-environment correctness	Supplies input to status admission	see record	SRC-1, .claude/commands/save.md	none	none	Verification scope/protocol not frozen
EPI-RTE-feature-admission	disposition/acceptance	doctrine only	EPI-OBJ-feature-list	non-truth-apt policy/content update: QA status admission or withdrawal	Record pass only after human verification; reset on break report	see record	see record	Requested true or false flag with failure notes	see record	Criterion/intended use are end-to-end user QA tracking; actual accepted instance absent	Later model treats verified versus failing/untested features differently	see record	SRC-1, .claude/commands/save.md	none	Acceptance rule, not observed acceptance	User authority and file update remain requested
EPI-RTE-lesson-inference	content transformation	doctrine only	EPI-OBJ-learned-behaviour, EPI-OBJ-helper-library	truth-apt transformation: indeterminate	Script edits become a challenged belief and classified lesson	see record	see record	Candidate FACT, HELPER, INSTRUMENT FAULT or JUDGEMENT	see record	A diff alone establishes changed text, not falsity, cause or transfer	Produces proposals, not authority to install yet	see record	SRC-1, .claude/commands/lessons.md	EPI-CLM-diff-record	Diff-as-falsification overstates evidence	Need script inputs, outcomes and candidate to decide entailment/non-entailment
RT-RTE-harvest-lessons	check/evidence production	doctrine only	EPI-OBJ-helper-library	no content change	Proposed helper rejects known-broken input and accepts known-good input	see record	see record	Requested two-way self-test proof summary	see record	Contrast licenses only tested cases with trustworthy case labels	Intended guard on installed helper quality	Foreground model, helper self-tests and report, requested protocol, harvest and future callers	SRC-1, .claude/commands/lessons.md	EPI-CLM-helper-safety	Two cases do not prove universal safety	No tests/results; test is after approval/install, no explicit failed-test rollback
RT-RTE-harvest-lessons	disposition/acceptance	doctrine only	EPI-OBJ-learned-behaviour, EPI-OBJ-helper-library, EPI-OBJ-decisions	no content change	User approves class/destination proposals before write, install or delete	see record	see record	Approved subset or no new lesson	see record	Human selection licenses installation, not a specified evidence-consuming truth criterion for every FACT/Why	Unapproved proposals cannot be installed by this prompt	see record	SRC-1, .claude/commands/lessons.md	EPI-CLM-helper-safety	Approval is not universal epistemic acceptance	Human reasoning not supplied; no actual response
RT-RTE-harvest-lessons	retention	doctrine only	EPI-OBJ-learned-behaviour, EPI-OBJ-helper-library, EPI-OBJ-decisions	no content change	Approved content installed before named scratch-script cleanup	see record	see record	Requested LB/helper/decision plus deletion/keep report	see record	Installed result does not gain warrant by replacing its evidence	Removes mined scripts; retains referenced and untracked/unmined files	see record	SRC-1, .claude/commands/lessons.md	EPI-CLM-diff-record	No mandatory retained input/outcome evidence bundle	Untracked mined evidence deletion may be irreversible
RT-RTE-harvest-lessons	behavior/policy adaptation	doctrine only	EPI-OBJ-helper-library, EPI-OBJ-learned-behaviour, EPI-OBJ-decisions	non-truth-apt policy/content update: helper/constraint admission	Approved lesson becomes later guidance or exported executable guard	see record	see record	Requested installed helper, Enforced by link or decision	see record	Helper behavior and linked prose have different licenses; no inherent proposition identified in absent code	Can constrain future callers or later agents	see record and RT-BAP-project-rules	SRC-1, .claude/commands/lessons.md	EPI-CLM-helper-safety	Export is not future call integration	Consumer use and improved action unobserved
RT-RTE-itemise-code	content transformation	doctrine only	EPI-OBJ-sidecar-index	truth-apt transformation: indeterminate	Code body becomes a job description or calls annotation; long units left blank	see record	see record	Requested new descriptions; existing descriptions preserved verbatim	see record	Semantic interpretation not automatically checked; particular descriptions might be faithful reshaping or new claims	Descriptions steer later code selection	see record	SRC-1, .claude/commands/itemise.md	EPI-CLM-context-economy	Not a semantic proof	Need generated description/body pair; dates/number parity cannot decide truth
RT-RTE-itemise-code	check/evidence production	doctrine only	EPI-OBJ-sidecar-index	no content change	Source marker numbers equal sidecar row numbers; each label lands on one job	see record	see record	Requested parity pass/fail and model granularity self-check	see record	Set equality licenses identifier coverage only, not description fidelity, order or granularity	Parity failure delays completion; granularity asks for subdivision	see record	SRC-1, .claude/commands/itemise.md	EPI-CLM-itemise-preservation	Different checks cover different properties	All checks are prompt operations, no execution supplied
EPI-RTE-itemise-filter	check/evidence production	doctrine only	EPI-OBJ-sidecar-index	non-truth-apt policy/content update: comparison-stream normalization	Comparison inputs normalized by marker filtering and CR deletion	see record	see record	No-output diff or difference; concrete false-negative below	see record	Equality in the filter's quotient, not byte or semantic equality	Supplies pass/fail to backup disposition	see record	SRC-1, .claude/commands/itemise.md	EPI-CLM-itemise-preservation	Filter can hide executable string content	Command is embedded doctrine, not an automatic installed checker
RT-RTE-itemise-code	disposition/acceptance	doctrine only	EPI-OBJ-sidecar-index	no content change	Complete parity and normalized diff license completion; diff failure restores source	see record	see record	Requested success, backup deletion or restored source	see record	Operational acceptance for labeling; cannot warrant all-code preservation	Keeps successful edits, restores failed source, ends backup recovery	see record	SRC-1, .claude/commands/itemise.md	EPI-CLM-itemise-preservation	Sidecar rollback not paired explicitly	False pass can discard recovery copy
RT-RTE-audit-project	check/evidence production	doctrine only	EPI-OBJ-audit-report, EPI-OBJ-sidecar-index, EPI-OBJ-decisions, EPI-OBJ-task-registry	truth-apt transformation: indeterminate	Task provenance, archive pointer/ID consistency, contradictions, Git state, dates and parity	see record	see record	Requested severity-tagged findings and passing checks	see record	Internal consistency/mtime evidence, not external truth or feature correctness; contradiction diagnoses require model interpretation	Surfaces issues; does not auto-rewrite descriptions	see record	SRC-1, .claude/commands/audit.md	EPI-CLM-archive-truth	Resume warning conflicts with normal save behavior, see runtime	No audit accuracy evidence; date mismatch is only a hint
RT-RTE-audit-project	disposition/acceptance	doctrine only	EPI-OBJ-audit-report, EPI-OBJ-sidecar-index	non-truth-apt policy/content update: admit user-selected repairs	User chooses each issue to fix or ignore	see record	see record	Requested selected repairs only; ignored issues not failures	see record	User permission, not automatically a truth judgment on diagnosis	Permits chosen fixes, withholds unchosen ones	see record	SRC-1, .claude/commands/audit.md	none	Ignored diagnosis remains without repair	Rollback/recheck not specified
RT-RTE-audit-project	retention	doctrine only	EPI-OBJ-audit-report	no content change	Displayed report saved to timestamped file	see record	see record	Requested durable audit	see record	Retention adds no independent warrant	Latest audit may guide a later agent via template	see record and RT-BAP-project-rules	SRC-1, .claude/commands/audit.md	none	Report persistence is not lifecycle integration of accepted claims	Startup command lacks explicit latest-audit read
RT-RTE-commit-reminder	operational admission/selection/consumption	implemented	RT-OBJ-settings	non-truth-apt policy/content update: per-event reminder	Raw Bash payload contains git commit	see record	see record	Checklist and unconditional zero exit	see record	None; no completeness evidence produced	Commit allowed whether records updated or not	see record	SRC-1, .claude/hooks/pre-commit-check.sh	none	Not a verifying or enforcing gate	Host/settings activation unobserved
RT-RTE-compact-backup	lineage/freshness/recovery	implemented	EPI-OBJ-session-log, EPI-OBJ-task-registry, EPI-OBJ-comments, EPI-OBJ-decisions, EPI-OBJ-feature-list	no content change	Copy existing five root files to timestamped directory	see record	see record	Attempted copies and count; no restore consumer	see record	Byte-copy intent preserves existing content, not unsaved progress or warrant	Provides recovery affordance, never blocks compaction	see record	SRC-1, .claude/hooks/pre-compact-save.sh	RT-CLM-progress-protection	Five-file/stale-state counterexample to all progress	Unchecked copy failure can still increment count

No lifecycle-integration row is established for accepted ampliative factual claims in these inspected paths. /lessons does connect approved entries to helpers and decisions, but its generic human permission rule does not specify evidence-consuming factual acceptance; ordinary save admits LB prose without that gate. Feature QA specifies a criterion, but tracking that result is not by itself explanatory integration. This bounded finding leaves open what an actual consumer might do; it does not convert missing deployments into an absence of checking.

System-claim versus route comparison

The claims and their source passages are see the cited claim records. There is no supplied observed-run or causal-experiment support for any row.

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-progress-protection	see record	SRC-1, README.md; doctrine/design	Incremental logging, checkpoint and backup policy	RT-RTE-compact-backup	Uninspected	No intervention/comparison; stale/omitted state cannot be copied	Narrow backup mechanism plus requested maintenance	All-progress/no-loss exceeds file coverage and unchecked copy result
EPI-CLM-diff-record	see record	SRC-1, .claude/commands/lessons.md; doctrine/design	Classifies changed beliefs, including instrument faults; excludes cosmetics	None; EPI-RTE-lesson-inference is doctrine only	Uninspected	No inputs/outcomes or attribution of edits	Useful candidate-mining instruction	Changed parameter may reflect changed requirements, not falsified belief
EPI-CLM-helper-safety	see record	SRC-1, .claude/commands/lessons.md; doctrine/design	Prefer helpers, two-way tests, approval/install before cleanup	None; see RT-RTE-harvest-lessons	Uninspected	Two examples do not exhaust inputs; future caller may not use helper	Intended executable guard and linked warning	Universal impossibility/never-forgotten claims unsupported
EPI-CLM-archive-truth	see record	SRC-1, .claude/commands/end.md; doctrine/design	Intact bodies, indexes, resolved and superseded markings	None; see RT-RTE-save-session, RT-RTE-audit-project	Uninspected	No archive integrity or recall experiment	Archiving is intended not to revoke active constraints	Continued policy force is supportable design; factual truth does not follow from movement
EPI-CLM-itemise-preservation	see record	SRC-1, .claude/commands/itemise.md; doctrine/design	Insert-only instruction, backups, parity, granularity and filtered comparison	None; EPI-RTE-itemise-filter analyzes requested command	Uninspected	Concrete multiline-string countercase below; Edit can replace matched text	Requested restricted labeling with an operational restore path	Neither tool name nor normalized comparison proves every character/semantics preserved
EPI-CLM-context-economy	see record	SRC-1, README.md; doctrine/design	Last-five-session bodies, index scans, on-demand archive and code reads	None; see RT-RTE-start-session, RT-RTE-itemise-code	Uninspected	Growing indexes/backlog, arbitrary body sizes/custom files, conflicting plan read; no token measurement	Intended selective access can avoid some full bodies	Predictable age-independent small cost, full recovery and measured savings not established

Bounded conclusion

Records receive operational authority more readily than epistemic warrant. /save, /end and incremental instructions retain conversation-derived status, summaries, decisions and gotchas. Comments are requested verbatim imports: that preserves attribution, not the truth of their propositions. Other outputs require interpretation; without input/output pairs, non-ampliative summary, derivation and conjecture cannot be distinguished claim by claim. Scoped root-cause and workaround fields can host criticizable content, but neither formatting nor retention supplies truth acceptance. See RT-RTE-save-session, EPI-OBJ-comments, EPI-OBJ-learned-behaviour.

Two admission regimes must not be collapsed. Feature QA names a human end-to-end criterion, an intended use (QA tracking), and a withdrawal input (reported breakage). It can license a bounded pass record if followed; no actual accepted instance or protocol-wide correctness is evidenced. /lessons instead asks permission to install proposals, while helper tests check two labeled cases after approval/install. That is a real requested check, not an absent evaluator, but it does not license every FACT or causal Why. Ordinary gotcha logging bypasses lesson approval. See EPI-RTE-feature-check, EPI-RTE-feature-admission, EPI-RTE-lesson-inference, RT-RTE-harvest-lessons.

The later authority paths are concrete but policy-level. /start reloads main records and indexes, uses a handoff marker, and requests relevant full archive bodies at task time. Decision supersession changes the active rule; resolved LB markings reduce reading without destroying the original. Constraints and lessons influence a later model through file reads; helpers can enforce guards only at actual calls. Approval, exporting and loading do not establish content-dependent behavior or successful learning. The numerical host component's version, mutability and internal truth-apt state remain uninspected; opaque reasoning is not evidence for or against hidden checking. See RT-BAP-project-rules, RT-RTE-start-session, RT-CMP-claude-code.

Checks license their own domains. Audit pointer/ID parity and code-marker parity assess internal consistency. Date comparison is expressly a stale-description hint, not a truth check. Contradiction/completeness checks are model interpretations with user-selected repairs. Filtered diff checks normalized text only; its false-negative witness prevents a universal preservation guarantee. The commit hook merely advises; the backup hook duplicates selected existing bytes without checking truth or restoreability. See RT-RTE-audit-project, EPI-RTE-itemise-filter, RT-RTE-commit-reminder, RT-RTE-compact-backup.

Lesson theory properties remain separate. The supplied theory account is see RT-RTE-harvest-lessons: localized LB/helper candidates are afforded; content-dependent consumption, criticism and retained next-round influence are claimed, not observed. This analyst's diff countercase qualifies the evidence that would satisfy criticism, without negating the instruction to distinguish instrument faults. User approval limits autonomy of admission. Retained debug beliefs can represent consumer-agent operation, but actual bidirectional reflection, improved capacity and criticism of Context Guard's own method remain unestablished. The potential improvement objective is avoiding recurrent consumer debugging errors across sessions; proposals, installed guidance and later channels are intended links, while actual update attribution and subsequent change-dependent behavior are missing. No exercised self-improvement or successful improvement follows. These uncertainties coexist with the described retention and checking capacities; they do not yield a single system-wide epistemic score.

Shared records

Operative objects

These objects are distinct consumer artifacts, not additional declarations of RT-OBJ-claude or RT-OBJ-settings. The closest supplied route container is cited per object; a route's description of its output is not a competing declaration of that output object. No supplied object counterpart was found after comparison. Generated objects are specified, not observed.

EPI-OBJ-comments

Label: Verbatim comments

Source-native identity: consumer COMMENTS.md and COMMENTS_pageN.md. Representational form: natural-language utterances plus textual timestamps/session attribution. Storage substrate: project files, optionally Git. Closest supplied record: RT-RTE-save-session, which writes but is not this artifact. Intended later consumer: /start and /audit; actioned comments can move out of the main file. Evidence: SRC-1, templates/COMMENTS.md, .claude/commands/end.md.

Format: Timestamp | Session | Verbatim comment

--- templates/COMMENTS.md

EPI-OBJ-session-log

Label: Session history

Source-native identity: consumer SESSION_LOG.md and its archive pages. Representational form: natural-language event summaries and status claims. Storage substrate: project files, optionally Git. Closest supplied record: RT-RTE-save-session. Historical entries differ from replaceable current state below. Evidence: SRC-1, templates/SESSION_LOG.md, .claude/commands/save.md.

Format: Session ID | Date | Summary of work done | Tasks completed | Tasks remaining

--- templates/SESSION_LOG.md

EPI-OBJ-resume-state

Label: Current in-flight state

Source-native identity: consumer RESUME_STATE.md. Representational form: natural-language handoff with a textual boolean marker interpreted by the model. Storage substrate: overwriteable project file, optionally Git. Closest supplied record: RT-RTE-save-session; unlike historical entries, /end clears this object. Evidence: SRC-1, templates/RESUME_STATE.md, .claude/commands/start.md.

Overwritten by /save. Wiped by /end. Read first by /start.

This file holds ONLY current in-flight state, not history. History lives in SESSION_LOG.md.

--- templates/RESUME_STATE.md

EPI-OBJ-task-registry

Label: Task provenance and status

Source-native identity: consumer TASK_REGISTRY.md and pages. Representational form: natural-language task/approach notes in a textual table with conventional IDs/status markers; host-model interpretation supplies consequences. Storage substrate: files, optionally Git. Closest supplied record: RT-RTE-save-session. Task done is distinct from feature verified. Evidence: SRC-1, templates/TASK_REGISTRY.md, .claude/commands/save.md.

CHECK THIS FILE BEFORE CREATING NEW TASKS. DO NOT DUPLICATE

--- templates/TASK_REGISTRY.md

- On completion, amend with: Files: path1, path2 | Approach: one-sentence pattern/library | Governed by: D-xx

--- templates/TASK_REGISTRY.md

EPI-OBJ-decisions

Label: Decisions and rationale

Source-native identity: consumer DECISIONS.md and pages/index. Representational form: natural-language choice/rationale/source and model-interpreted category/links. Storage substrate: files, optionally Git. Closest supplied record: RT-RTE-save-session. Unlike RT-OBJ-claude, these are use-derived per-decision constraints, not the template that tells the agent to respect them. Evidence: SRC-1, templates/DECISIONS.md, .claude/commands/end.md.

Decision: [The choice that was made]

Rationale: [The reasoning that led to it]

Source: [User comment | Plan | Session | External requirement]

--- templates/DECISIONS.md

• If a new decision supersedes an earlier one, mark the old Category: superseded with a pointer to the new D-number. --- .claude/commands/end.md

EPI-OBJ-learned-behaviour

Label: Tactical knowledge entries

Source-native identity: consumer LEARNED_BEHAVIOUR.md, pages/index and optional helper references. Representational form: natural-language scoped failure/workaround/explanation; resolved is model-interpreted disposition text. Storage substrate: files, optionally Git. Closest supplied records: RT-RTE-save-session and RT-RTE-harvest-lessons, two writers of the same object, not two object identities. No actual lesson examined. Evidence: SRC-1, templates/LEARNED_BEHAVIOUR.md, .claude/commands/lessons.md.

Context: Where this surfaces (platform, plugin, version) Gotcha: What fails and how (symptom) Workaround: What actually works Why: Root cause if known --- templates/LEARNED_BEHAVIOUR.md

1. An entry whose underlying platform or library has been removed, or upgraded past the bug, can be marked resolved in the index, so it stays findable but no longer needs reading. --- .claude/commands/end.md

EPI-OBJ-feature-list

Label: Feature QA status

Source-native identity: consumer FEATURE_LIST.json, initialized with an empty features array. Representational form: symbolic JSON fields whose intended QA semantics require human/model interpretation; natural-language names/failure notes are separate claim content. Storage substrate: files, optionally Git. Closest supplied record: RT-RTE-save-session. Truth-apt human-verification attribution is separate from operational pass classification. Evidence: SRC-1, templates/FEATURE_LIST.json, .claude/commands/save.md.

"verifiedBy": "user | null", "verifiedAt": "YYYY-MM-DD | null", "notes": "Failing case description, linked task IDs, or empty string" --- templates/FEATURE_LIST.json

EPI-OBJ-sidecar-index

Label: Code sidecar descriptions

Source-native identity: generated <source_filename>.index.md paired with numbered source markers. Representational form: natural-language job descriptions plus symbolic section addresses and date strings; semantic selection by a model, marker matching by grep if executed. Storage substrate: project source/Markdown files, optionally Git. Closest supplied record: RT-RTE-itemise-code. No generated sidecar supplied; descriptions do not have the same warrant as source code. Evidence: SRC-1, .claude/commands/itemise.md, templates/CLAUDE.md.

1. If the section is fewer than 50 lines, write a one-line description following the ruleset.

2. If the section is 50 lines or longer, leave the description as _(blank, fill on first edit)_. This is a deliberate token-saving choice: long sections need careful summarisation, and an inaccurate auto-generated description is worse than no description (it misleads agents and corrupts their model of the codebase). Calibration scaffolding is not authoritative, so let the next coding agent fill it in when they touch the section for the first time. --- .claude/commands/itemise.md

3. Preserve every existing description verbatim. Do not rewrite even if it doesn't match the auto-generated style. Hand-written descriptions are sacred and exist for a reason. --- .claude/commands/itemise.md

EPI-OBJ-helper-library

Label: Installed lesson helpers

Source-native identity: functions in a consumer project's helper library, exported by /lessons, optionally named by LB Enforced by. Representational form: symbolic executable code; associated explanatory prose belongs to the LB object. Storage substrate: consumer code files, optionally Git. Closest supplied record: RT-RTE-harvest-lessons. No helper implementation available, so exact assumptions, consumer call paths and proposition encoding are not determinable. Evidence: SRC-1, .claude/commands/lessons.md.

• LEARNED_BEHAVIOUR entries: standard format (## LB-NNN: title (Session N, dd/mm/yy) with Context / Gotcha / Workaround / Why). Where a helper now enforces the lesson, add an **Enforced by:** line naming it. That tells a future reader the rule is already load-bearing in code and does not depend on them remembering it. --- .claude/commands/lessons.md

The supplied RT-RTE-harvest-lessons retains the helper export/two-way self-test passage. That instruction does not itself show the helper is called.

EPI-OBJ-audit-report

Label: Persistent integrity diagnosis

Source-native identity: generated audits/YYYY-MM-DD_HHMMSS.md. Representational form: natural-language issue/severity/details/fix and passing checks. Storage substrate: project Markdown file. Closest supplied record: RT-RTE-audit-project. Report creation precedes selected repairs; retained diagnosis is not proof fixes succeeded. Evidence: SRC-1, .claude/commands/audit.md, templates/CLAUDE.md.

Save the full audit report (exactly as displayed above) to: audits/YYYY-MM-DD_HHMMSS.md --- .claude/commands/audit.md

Routes

EPI-RTE-feature-check

Label: Human end-to-end feature check Part of: RT-RTE-save-session

Endpoints/progression/owner: user tests a named feature end-to-end or reports a broken case → foreground model obtains that report → QA admission consumes it. This is the evaluation part of the supplied save/end route, not its file update. Context: user-facing feature under the user's actual test conditions, unspecified in the package. Evaluator and answer reference: human expected end-to-end behavior; the package freezes neither an oracle nor test suite. User supplies judgment, model is recorder; model task-completion status cannot substitute. Operating mode: open project QA, triggered by human verification or a broken report, not a controlled experiment.

• design conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: user verification or failing-case report, if supplied.
• Later read-back: inapplicable — this part produces the judgment; status retention and later reads belong to admission and the parent route.
• Delegated visibility: inapplicable — no separate worker specified for human verification.
• Selection predicate: a named feature manually verified end-to-end or reported broken by the user.
• Invalidation or expiry: later broken report can defeat the recorded positive result; no version/environment expiry prescribed.
• Activation or effect: requested judgment becomes input to QA status update, not an automatic product pass.
• Evidence limits: no human test transcript, feature implementation, environment or result supplied.

Evidence: SRC-1, .claude/commands/save.md; the load-bearing human-verification passage is see RT-RTE-save-session. Distinct from admission: this part assesses behavior; admission authorizes the record change.

EPI-RTE-feature-admission

Label: Admit or withdraw feature status Part of: RT-RTE-save-session

Endpoints/progression/owner: human verification/break report → foreground model evaluates the admission trigger → changes passes and failure notes → later /start QA summary. Proposed change: bounded status claim, not task completion. Human controls the qualifying evidence and can withhold it; model performs the update. Criterion: user manually verified end-to-end success; intended use and scope: the project's QA tracker for that feature, not universal correctness. A broken report withdraws positive reliance. Operating mode: open QA tracking; reference outcome is the human check above. Guidance is the FEATURE_LIST semantic rule. Recovery: later reports can reverse status; no automatic code rollback is specified. Parent route supplies durability, not independent validation.

• design conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: requested true/false status and failing-case notes.
• Later read-back: QA status reaches later /start and model summaries through FEATURE_LIST.
• Delegated visibility: uninspected — no host propagation of QA status to other agents supplied.
• Selection predicate: human verification permits true; human break report requires false; task completion alone is insufficient.
• Invalidation or expiry: later break report resets status; no automatic expiry on code change specified.
• Activation or effect: requested status-dependent later QA interpretation; actual changed reliance unobserved.
• Evidence limits: no recorded acceptance instance or downstream behavior supplied; verifiedBy/verifiedAt are template fields, not enforced provenance.

Evidence: SRC-1, .claude/commands/save.md, templates/FEATURE_LIST.json. Separate from EPI-RTE-feature-check: admission changes the status artifact, not the test outcome.

• If the user reports a feature broken, flip passes: false with a notes description of the failing case. --- .claude/commands/save.md

EPI-RTE-lesson-inference

Label: Infer lessons from script changes Part of: RT-RTE-harvest-lessons

Endpoints/progression/owner: foreground model selects current-session script series → consecutive diffs → interprets a prior belief and its alleged failure → deduplicates/classifies → presents proposals to human. Context includes existing LB/decision indexes and helper library. This part produces the candidate, unlike the parent's approval/install/check/cleanup stages. Proposed change: scoped FACT, HELPER, INSTRUMENT FAULT or JUDGEMENT. Model proposes and classifies; human can veto the later installation. Operating mode: open debugging harvest, no frozen expected-answer oracle. Static guidance prefers helpers, ignores cosmetic changes and treats check failure separately from product failure. What persists, if approved: Context/Gotcha/Workaround/Why, helper links/code and constraint rationale; input/outcome traces need not survive script deletion. Rollback/recovery: no candidate rollback needed before approval; parent's Git/cleanup limits apply after installation.

The content/update relation is indeterminate without actual scripts/outcomes and authored lesson. A changed selector can correct a mistaken assumption, reflect changed requirements, or simply target another page. This source-supported alternative prevents treating every diff as an entailed falsification. The prompt's own cosmetic exclusion and INSTRUMENT FAULT category confirm that changed text does not uniquely identify the blame target. No actual ampliative accepted claim is evidenced.

• design conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: proposed classified lessons or nothing-new result.
• Later read-back: inapplicable — proposals alone are not retained by this part; approved outputs reach later consumers through the parent.
• Delegated visibility: inapplicable — foreground harvest has no delegated mining step.
• Selection predicate: session prefix/mtime, name-stem series, changed substantive selectors/waits/endpoints/guards, novelty against existing knowledge.
• Invalidation or expiry: duplicate/cosmetic candidates rejected; user may reject before writes; no retained proposal expiry specified.
• Activation or effect: requested proposal generation, no installation authority until parent approval.
• Evidence limits: no actual series, outcomes or lesson instance; diff is not a test result or causal contrast.

Evidence: SRC-1, .claude/commands/lessons.md. Theory account and theory-builder conditions 1–4 are see RT-RTE-harvest-lessons, the containing route: this inference part applies its static criticism guidance but supplies no separately observed formulation, consumption, criticism or iteration. Addressability is afforded by localized LB fields/helper functions; persistence and improved future capacity remain unobserved. learning: uninspected — no improvement/attribution evidence. Reflection is afforded for consumer debug beliefs, with actual two causal directions unobserved. Human approval is outside this inference part but remains inside the complete admission route, so it does not establish an autonomous theory builder. No criticism-fed revision of the shipped harvest method is shown. The complete route's dispositional self-change assessment is not upgraded by this candidate-generation part.

Theory-builder conditions 1–4 for this part: - condition 1 conclusion status: afforded

The requested whole LB/helper proposal can localize a claim; no actual formulation supplied.

• condition 2 conclusion status: claimed

Classification is requested to depend on candidate content; no decision trace supplied.

• condition 3 conclusion status: claimed

The prompt asks for the challenged belief and separates instrument fault; diff alone does not establish that the criticism occurred or was warranted.

• condition 4 conclusion status: claimed

The parent requests installation for later use; this candidate-generation part alone supplies no next-round evidence.

For each consecutive pair of files in a series, answer one question: what did the agent believe in file N that turned out to be wrong in file N+1? --- .claude/commands/lessons.md

Look for a changed selector, wait condition, endpoint, parameter name, assertion, path, encoding, or a guard that suddenly appeared. Ignore cosmetic churn: renamed variables and reformatting are not lessons. --- .claude/commands/lessons.md

| INSTRUMENT FAULT | a LEARNED_BEHAVIOUR entry, flagged as such | the check was wrong while the thing being checked was fine | --- .claude/commands/lessons.md

EPI-RTE-itemise-filter

Label: Normalized itemisation comparison Part of: RT-RTE-itemise-code

Endpoints/progression/owner: foreground model is told to run Bash → both source/backup streams pass through grep exclusion and CR deletion → diff -B emits differences or no output → parent keeps edits/deletes backups or restores. This is the comparison part, not the parent's marker generation/parity or disposition. Evaluator: deterministic text filter and diff, if executed. Target/domain: normalized line equality after removing regex-matching lines, carriage returns and blank-only differences. Oracle: backup under the same normalization. Operating mode: bounded labeling check; model invokes/interprets, program compares, human approved scope earlier. The producing command is inspected as embedded doctrine; it is not installed executable enforcement and was not run here.

Concrete deduction from the inspected expression, not target execution: take a Python file with the three lines s = """, # 1. alpha, """. Change only the middle line to # 1. beta. That middle line is string data, not a Python comment, yet both lines match the stated regex and are removed by grep -Ev. The two normalized streams both consist of the unchanged opening and closing string lines; CR deletion and diff -B cannot distinguish them. The requested no-output branch labels this PASS although the value of s changed. Marker syntax therefore does not identify comments in a language-aware domain. Independently, tr -d '\r' removes all CR bytes, not only CRLF endings. These source-level deductions qualify universal source-preservation warrant, without asserting a real corrupted file.

• design conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: requested diff output or no output, consumed as FAIL/PASS.
• Later read-back: inapplicable — comparison streams are current-operation inputs, not accumulated memory; parent keeps source/sidecar edits.
• Delegated visibility: uninspected — foreground is to verify a worker's edits, but actual worker/result handling is not supplied.
• Selection predicate: each approved itemised source and its pre-edit backup; both use identical filter.
• Invalidation or expiry: backup deletion on completion ends immediate comparison/recovery reference; re-itemisation makes a new backup.
• Activation or effect: intended completion/restoration decision in the parent, not independently observed force.
• Evidence limits: no command run; filter equality licenses only its quotient, not semantic or byte preservation; error-status handling is not prescribed by the output-only branch.

Evidence: SRC-1, .claude/commands/itemise.md. No theory is formulated by this deterministic comparison part. Normalization rearranges/removes comparison representation without deriving a new factual code claim; the semantic inference from no diff to no code change is not warranted across the advertised language domain.

PATTERN="^\s(\/\/|#|/*|<!--|--)\s+([0-9]+(.[0-9]+)(\s|.)|end of\s+[0-9]+)" diff -B <(grep -Ev "$PATTERN" "{filename}" | tr -d '\r') <(grep -Ev "$PATTERN" "{filename}.itemise-backup" | tr -d '\r') --- .claude/commands/itemise.md

• If diff produces no output: PASS, only comment lines (and cosmetic blank lines) were added
• If diff produces output: FAIL, actual code was changed --- .claude/commands/itemise.md

Claims

Each new claim has a distinct public/design assertion from RT-CLM-progress-protection, whose backup/recovery referent is unchanged. These claims concern inference, helper universality, archive truth, preservation checking and context cost respectively; none supersedes the supplied record.

EPI-CLM-diff-record

Label: Script diff records false belief

Claimed operation: consecutive script differences are a literal record of an agent belief that proved false and can be mined into permanent lessons. Source: SRC-1, .claude/commands/lessons.md, README.md; doctrine/design. The diff itself lacks outcome/blame identification; changed needs and cosmetics are alternatives. Design conclusion status: claimed. Operation conclusion status: uninspected.

The diff between consecutive versions of a throwaway script is a literal record of what the agent believed that turned out to be false. --- .claude/commands/lessons.md

EPI-CLM-helper-safety

Label: Helper prevents wrong answer

Claimed operation/warrant: a helper can make wrong answers impossible and avoid the failure of remembered prose. Source: SRC-1, .claude/commands/lessons.md; doctrine/design. Two-way testing and user approval support a narrower intended guard, not universal correctness or inevitable future calls. Design conclusion status: claimed. Benefit conclusion status: uninspected.

The governing principle: prefer a helper over prose. A rule that must be remembered fails. A helper function that gets called, and a LEARNED_BEHAVIOUR entry that a future /start surfaces, do not. When a lesson can be expressed as code that makes the wrong answer impossible, write the code. A prose note saying "remember to escape the path" decays, a safePath() that escapes it cannot be forgotten. --- .claude/commands/lessons.md

EPI-CLM-archive-truth

Label: Archived lesson remains true

Claimed warrant: an archived learned behaviour is still true; archived active decisions still govern. Source: SRC-1, .claude/commands/end.md; doctrine/design. Normative continuation differs from factual validity. Resolved markers and intact relocation can preserve available content but do not test its premises. Design conclusion status: claimed. Operation conclusion status: uninspected.

An archived learned behaviour is still true. The index line is what stops a future session re-discovering it the hard way. --- .claude/commands/end.md

Archiving does not revoke a decision. forever-active and active-constraint decisions govern from the archive exactly as they did from the main file, and the index is how future sessions find them. --- .claude/commands/end.md

EPI-CLM-itemise-preservation

Label: Itemisation preserves existing source

Claimed operation/warrant: labeling adds comments without changing source; Edit structurally prevents loss and the verification check catches any character change. Source: SRC-1, .claude/commands/itemise.md; doctrine/design, including attributed Session 1 operation rather than an acquired trace. No observed run/causal experiment supplied. The insert-only discipline is a real requested restriction; Edit's ability to replace old_string does not enforce inclusion of every old character, and the filter has a semantic false-negative witness. Design conclusion status: claimed. Operation conclusion status: uninspected.

Removing, altering, or rewriting ANY existing character in the file, whether code or comment, is a catastrophic failure. The verification step (Step 5) will catch this, but the Edit-based approach prevents it structurally. --- .claude/commands/itemise.md

EPI-CLM-context-economy

Label: Small predictable recovery context

Claimed operation/warrant: recovery uses only five-session content at a predictable small token cost independent of project age; sidecar-guided code reads scope to relevant code. Source: SRC-1, README.md; doctrine/design. Variable-sized recent bodies and growing indexes/backlog prevent the stated fixed-cost implication. Sidecar guidance may reduce reads, not prove relevance/completeness or measured savings. Design conclusion status: claimed. Benefit conclusion status: uninspected.

It reads the last five sessions' worth of content and nothing more: archives and plan files stay closed, so recovery costs a predictable, small number of tokens no matter how old the project is. --- README.md

Why this exists: a number on its own (e.g. 2.1) is a coordinate without a label. With the sidecar, an agent answering "where is the code that does X?" reads the small sidecar (cheap), picks the matching number, then greps the source for that number's start/end markers and reads only those bytes. Token usage scopes to exactly the relevant code instead of the whole file. --- README.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; the applicable project-file channel is see RT-BAP-project-rules, with per-route narrower consumers/force distinguished in the ledger.