Claude Workstream Kit epistemic report

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

Source-and-claim boundary

The question is how work-state claims acquire warrant before they guide another session, become instructions, or authorize closure. Scope and exclusions are see boundary; source identity and layers are see Source register. All source findings below use SRC-1 at the frozen commit. Direct inspection establishes code behavior or model-facing instructions, not model compliance. No target was executed.

The kit is deliberately an operational work-state system, not an all-purpose truth adjudicator. Its checks are heterogeneous: parser predicates establish facts about text; model verification compares task claims with artifacts; users approve substantive changes and closure. Their licenses must not be merged. Consequential claims are EPI-CLM-grounded-claims, EPI-CLM-gate-validation, and RT-CLM-rewrite-preservation.

Entry point or operation	Source path in SRC-1	Coverage and records; limit
Registered hooks	.claude/settings.json	Direct full read: SessionStart and Stop registrations only; supplied RT-RTE-session-reload, RT-RTE-stop-gloss supply hook findings. No registered tool server appears in these settings; no conclusion about external host tools.
Record CLI	.claude/scripts/workstream-record.py	Direct dispatch inspection: record/default-root, cites, refs, paths, decay, git, fires, bare-ids, self-check. Direct Learning-field and self-check implementation; EPI-RTE-marker-measure. Other handlers only function/search coverage here, supplemented by supplied hook/rewrite findings. Their exact matching completeness is unassessed, preventing a system-complete negative about omitted checking routes.
Rewrite CLI	.claude/scripts/workstream-rewrite.py, .claude/scripts/workstream_state.py	Direct dispatch, finish, Learning transformation/check, sentence/marker/fingerprint code and rotation admission tail; EPI-RTE-learning-condense, RT-RTE-rewrite-state. Detailed records/decisions and rotation dependency-sweep internals use supplied runtime findings, not a fresh full audit.
Compatibility entry	.claude/scripts/condense-completed-records.py	Direct full read: forwards to run_condense, including write flag. Same underlying route, not another evaluator.
Work, create, close and delegates	.claude/skills/workstream-work/SKILL.md, .claude/skills/workstream-create/SKILL.md, .claude/skills/workstream-close/SKILL.md, .claude/agents/verifier.md	Verifier directly read; other skill and agent findings supplied by RT-RTE-work-goal, RT-RTE-agent-packet, RT-RTE-state-lifecycle.
Capture, review, extract	.claude/skills/workstream-capture/SKILL.md, .claude/skills/workstream-review/SKILL.md, .claude/skills/workstream-extract/SKILL.md	Direct full reads; EPI-RTE-capture-synthesis, EPI-RTE-still-true, EPI-RTE-disposition-check. Linked failure histories not independently inspected: historical causal explanations remain attributed.
Status and cross-project intake	.claude/skills/workstream-status/SKILL.md, .claude/skills/handoff/SKILL.md	Direct full reads; EPI-RTE-status-statement, RT-RTE-state-lifecycle.
Installer, upgrade and context telemetry	install.sh, .claude/skills/workstream-upgrade/SKILL.md, .claude/scripts/status-line.sh	Supplied RT-RTE-install-upgrade and runtime context account only; exact deployment drift predicates not re-audited. Copying instructions is operational admission, not acceptance of their truth. Telemetry is a budget signal, not a semantic check.
Model-calling evaluator	tests/va3-acceptance.sh	Direct full read; RT-RTE-acceptance-driver. Other offline suites remain available but uninspected here: no claim of comprehensive test coverage or successful tests.
Public design claims	docs/design.md, README.md	Direct bounded search passages; claim records below. Statements about earlier runs are reported operation, not supplied traces.

Missing live consumer artifacts prevent individuating an actual accepted Learning or synthesis and checking its truth. Missing host implementation and traces prevent establishing actual goal arming, delegation isolation, user-gate compliance or downstream activation. Missing interventions prevent attributing safety, learning benefit or improvement to the kit. Exact provider state is see RT-CMP-claude: parameter updates, exact model version and endpoint resolution are uninspected. These limits do not negate the inspected deterministic routes or the declared evidence protocols.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-workstream-claims	Dated findings, explanations, completion and disposition claims	Durable premises for review and future work	SRC-1, .claude/skills/workstream-review/SKILL.md	Natural-language claims, not automatically accepted knowledge; actual instance truth not determinable.
EPI-OBJ-workstream-markers	Markers encode assertions of state; parser predicates about their occurrence are truth-apt separately	Count, signal and rewrite selection	SRC-1, .claude/scripts/workstream_state.py	Symbolic marker occurrence does not prove the asserted disposition.
EPI-OBJ-workstream-guidance	Plans/criteria are primarily normative; their premises and claims of satisfaction are truth-apt	Scope, ordering and user gates	SRC-1, .claude/skills/workstream-review/SKILL.md	Approval changes permission, not source truth.
RT-OBJ-active	see record	see record	SRC-1, .claude/hooks/session-start.sh	Pointer/prose mixture; cannot establish armed-goal state.
EPI-OBJ-record-learnings	Counts and classifications of Learning blocks under the parser grammar	Input to drain selection and status	SRC-1, .claude/scripts/workstream-record.py	Target is file form, not truth of entries or correctness of their claimed destinations.
RT-CMP-claude	Particular generated propositions not determinable from inaccessible weights	see record	SRC-1, .claude/agents/verifier.md	Model instructions and outputs must not be confused with parameter change or an oracle.

Authority-route ledger

Fields owned by supplied or declared records are linked rather than reproduced. Architectural status below is local to this ledger. A prompt-only semantic route is doctrine only: executable delivery does not establish execution of the requested evaluation.

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	Behavioral-authority path	Evidence source ID and local anchor	Claim IDs	Mismatch marker	Gap/limit
EPI-RTE-capture-synthesis	content transformation	doctrine only	EPI-OBJ-workstream-claims	truth-apt transformation: indeterminate	Session findings to durable synthesis	see record	see record	see record	Requested writing	Candidate only; no general acceptance	Routine capture can persist without per-claim user approval	see record	SRC-1, .claude/skills/workstream-capture/SKILL.md	none	none	Could copy, reshape, derive or conjecture; actual inputs/output needed to classify.
RT-RTE-state-lifecycle	retention	doctrine only	EPI-OBJ-workstream-claims	no content change	Current notes to committed checkpoint	see record	see record	see record	Model policy	Retention does not confer acceptance	Future work may use candidate notes	RT-BAP-resume-state	SRC-1, .claude/skills/workstream-capture/SKILL.md	none	none	Git retention depends on compliance and publication.
EPI-RTE-marker-measure	check/evidence production	implemented	EPI-OBJ-record-learnings	truth-apt transformation: entailed derivation	Learning blocks and dated markers	see record	see record	see record	Code classification, advisory to semantic decisions	Counts under named grammar only	Supplies drain/status inputs, never approves content	see record	SRC-1, .claude/scripts/workstream-record.py	EPI-CLM-grounded-claims	none	Syntactically terminal can be factually false.
EPI-RTE-still-true	check/evidence production	doctrine only	EPI-OBJ-workstream-claims	no content change	Premises versus current artifacts and commitments	see record	see record	see record	Requested comparative inspection	Local evidence, not document authority alone	Findings propose reliance changes	see record	SRC-1, .claude/skills/workstream-review/SKILL.md	EPI-CLM-grounded-claims	none	Missing consumer artifacts and runs.
EPI-RTE-still-true	disposition/acceptance	doctrine only	EPI-OBJ-workstream-claims	non-truth-apt policy/content update: mark supersession or revise scope	Stale findings to changed reliance	see record	see record	see record	Model policy; substantive restructure user gate	Evidence can withdraw current reliance; approval is not infallibility	Mark stale in place; revise tasks after required approval	see record	SRC-1, .claude/skills/workstream-review/SKILL.md	none	none	Not a universally enforced content admission gate.
EPI-RTE-disposition-check	check/evidence production	doctrine only	EPI-OBJ-workstream-claims	no content change	Claimed disposition versus governed destination	see record	see record	see record	Requested artifact check	Verifies routing/implementation, not every generalization's truth	Missing application requires edit or re-routing	see record	SRC-1, .claude/skills/workstream-extract/SKILL.md	EPI-CLM-grounded-claims	none	Marker counting is explicitly insufficient.
EPI-RTE-disposition-check	operational admission/selection/consumption	doctrine only	EPI-OBJ-workstream-claims	non-truth-apt policy/content update: apply, hand off or drop	Resolved entries to governed files or declined matter	see record	see record	see record	Requested edits plus separate handoff gate	Routing licenses intended operational use, not general truth	Guidance may change future work; dropping with rationale is allowed	see record	SRC-1, .claude/skills/workstream-extract/SKILL.md	none	none	Not automatically lifecycle integration of accepted ampliative knowledge.
EPI-RTE-learning-condense	content transformation	implemented	EPI-OBJ-workstream-claims	truth-apt transformation: indeterminate	Terminal Learning to selected sentences	see record	see record	see record	Deterministic selection	Literal retention checks do not prove preserved meaning	Shorter text can replace later-read state	see record	SRC-1, .claude/scripts/workstream-rewrite.py	RT-CLM-rewrite-preservation	Semantic preservation unestablished	Concrete trace below shows omission.
EPI-RTE-learning-condense	check/evidence production	implemented	EPI-OBJ-workstream-claims	no content change	Statement/target/disposition strings, count and fingerprint	see record	see record	see record	Local refusal before own write	Checked syntactic relation only	Blocks this CLI path when predicate fails	see record	SRC-1, .claude/scripts/workstream-rewrite.py	RT-CLM-rewrite-preservation	Rotation exception, see runtime	Manual edits bypass checks.
RT-RTE-rewrite-state	lineage/freshness/recovery	implemented	RT-OBJ-workstream	truth-apt transformation: indeterminate	Rotation and records/decisions reduction	see record	see record	see record	Code refuses selected defects; --write replaces text	Reference/structure checks, not semantic endorsement	Alters visible state; stale-claim override permits continuation	RT-BAP-resume-state	SRC-1, .claude/scripts/workstream-rewrite.py	RT-CLM-rewrite-preservation	General preservation claim too broad	Historical recovery requires prior commit/tag, not just a successful transform.
RT-RTE-agent-packet	check/evidence production	doctrine only	EPI-OBJ-workstream-claims	no content change	Worker claims against packet/spec and outputs	see record	see record	see record	Verifier advisory PASS/FAIL	Task-specific command/spec domain	Foreground must verify findings	see record	SRC-1, .claude/agents/verifier.md	EPI-CLM-grounded-claims	none	Fresh context does not establish evaluator independence or correctness.
RT-RTE-agent-packet	disposition/acceptance	doctrine only	EPI-OBJ-workstream-guidance	no content change	Task product for intended packet use	see record	see record	see record	Foreground admission; user veto at gates	Evidence-consuming task acceptance, conditional on compliance	Accept, retry, block or stop before next packet	see record	SRC-1, .claude/skills/workstream-work/SKILL.md	EPI-CLM-grounded-claims	none	Product completion is not acceptance of a new general explanation.
EPI-RTE-status-statement	content transformation	doctrine only	EPI-OBJ-record-learnings	truth-apt transformation: indeterminate	Records to cross-workstream statement	see record	see record	see record	Advisory chat	Attributed premises, no independently derived priority	Informs user selection; no state mutation requested	see record	SRC-1, .claude/skills/workstream-status/SKILL.md	none	none	Some joins follow recorded edges; semantic inference needs actual output.
EPI-RTE-status-statement	check/evidence production	doctrine only	EPI-OBJ-workstream-claims	no content change	Statement counts, citations and surrounding context	see record	see record	see record	Requested correction/drop before delivery	Occurrence and scoped interpretation, not external premise verification	Mismatched citations corrected or dropped	see record	SRC-1, .claude/skills/workstream-status/SKILL.md	none	none	Self-check fixtures calibrate patterns, not every real file or conclusion.
RT-RTE-state-lifecycle	disposition/acceptance	doctrine only	EPI-OBJ-workstream-guidance	no content change	Closure criteria with current evidence	see record	see record	see record	Required user approval by policy	Deliverable/exit scope only	Permits status done, archive and deletion, not arbitrary publication	see record	SRC-1, .claude/skills/workstream-close/SKILL.md	EPI-CLM-gate-validation	none	Preapproved unattended path differs from an interactive gate.
RT-RTE-state-lifecycle	operational admission/selection/consumption	doctrine only	EPI-OBJ-workstream-claims	truth-apt transformation: acquisition/import	Handoff claims to receiving backlog	see record	see record	see record	Receiver verification requested	Source warrant unknown until local check; provenance requested	Receiver may act, route or decline, then delete processed inbox	see record	SRC-1, .claude/skills/handoff/SKILL.md	none	none	Filing is not acceptance; sender approval is destination/visibility authority.
RT-RTE-session-reload	operational admission/selection/consumption	implemented	RT-OBJ-active	no content change	Retained pointer and signals to host context	see record	see record	see record	Code selection and stdout; model guidance thereafter	No semantic warrant from freshness	Can change what later model inspects	RT-BAP-resume-state	SRC-1, .claude/hooks/session-start.sh	none	none	Full workstream is separately read, not injected wholesale.
RT-RTE-stop-gloss	check/evidence production	implemented	EPI-OBJ-workstream-markers	no content change	Final answer bare-reference form	see record	see record	see record	Best-effort one-turn repair request	Gloss shape, not reference truth or task completion	Host can request another turn	see record	SRC-1, .claude/hooks/bare-id-guard.py	none	none	Fail-open and repeated-stop paths prevent universal enforcement.
RT-RTE-acceptance-driver	check/evidence production	implemented	EPI-OBJ-workstream-guidance	no content change	Greeting fixture and lifecycle artifacts	see record	see record	see record	Shell verdict; no automatic production update	Fixed scenario only	Reports PASS/FAIL/HELD to maintainer	see record	SRC-1, tests/va3-acceptance.sh	EPI-CLM-gate-validation	Outcome cannot identify mechanism	Unexecuted; pattern checks and conditional held classification are not universal safety tests.

The concrete Learning deduction is see EPI-RTE-learning-condense. No row establishes lifecycle integration after acceptance of a produced ampliative claim. Extraction and retention have operational authority, but actual candidate generation, evidence-consuming semantic acceptance and downstream integration are not shown as a connected observed episode.

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-grounded-claims	see record	SRC-1, docs/design.md; doctrine/design	Artifact checks, user gates and fresh verifier requested	EPI-RTE-marker-measure, RT-RTE-acceptance-driver; semantic checks remain doctrine	None supplied	No comparison identifying safety effects; marker evidence is weaker than truth	Coherent grounding protocol with executable file measures	Grounding not universally enforced; ordinary capture can retain unchecked candidates.
EPI-CLM-gate-validation	see record	SRC-1, docs/design.md; reported operation	Skills reserve gates to users	RT-RTE-acceptance-driver	Reported earlier tests only	Fixed scenario and prompt-supplied approvals; no attribution to text alone	Driver can test selected lifecycle outcomes	No independently supplied all-constraints trace; native goal not exercised by driver.
RT-CLM-rewrite-preservation	see record	SRC-1, .claude/skills/workstream-extract/SKILL.md; doctrine/design	Extraction orders checking before trimming	RT-RTE-rewrite-state, EPI-RTE-learning-condense	None supplied	Local predicates only	Checked syntactic retention on standard path	Rotation intentionally changes headings/checkboxes; Git backup is caller-dependent.

Bounded conclusion

Retained operational claims can guide later sessions before semantic acceptance. Capture asks the model to persist findings and synthesis, and reload/work routes deliver or request that state. Workstream claims, symbolic disposition markers and plans have different authority paths. Neither persistence nor an APPLIED marker establishes truth or application to another file. The supplied runtime and direct marker implementation support that separation.

The kit's strongest semantic checks are evidence-consuming protocols, not a single enforcing evaluator. Review compares dated claims with current artifacts, marks supersession and asks user approval for substantive restructuring. Extract verifies actual governed destinations, permits dropping with rationale and distinguishes deferral from disposition. Work acceptance compares products to packet criteria; closure additionally needs user approval against exit criteria. Each can change operational reliance in its local scope. Those protocols are doctrine only as semantic behavior in this evidence set; their wording alone is not a run.

Executable evidence is narrow but real. A dated marker can entail a parser classification; fingerprint and sentence-presence predicates can block a script write. Their checked domain is text form. Condensation's checked deletion may remove a qualification, and rotation has separate checks and explicit stale-claim override. These checks cannot license preserved meaning, external premise truth or reliable behavior. No warranted general explanatory conclusion follows from a preserved heading or a passing smoke test.

Imports and summaries retain uncertain warrant locally. Handoff carries source context and requests receiver verification before a claim becomes a task description; filing and deletion do not equal acceptance. Status limits sources, labels unverified premises inline and requests citation checking. Its chat informs user decisions but is not automatically retained or accepted. Capture's synthesis relation is indeterminate without actual inputs and output: the instruction's transfer test does not prove novelty, non-entailment, correctness or acceptance.

No actual accepted ampliative claim, associated test result and post-acceptance integration episode is established. This is not a claim that the model cannot conjecture. No source-supported parameter-learning route is established within the inspected kit, while provider behavior remains uninspected. Retained state and applied guidance can alter future action without a parameter update; improved capacity, exercised self-improvement and causal benefit remain independently unestablished. The design's grounding and autonomous-safety claims have protocol and fixture support, not a system-wide warrant or demonstrated causal effect. Uninspected record handlers and offline suites prevent a system-complete negative about other checking routes.

Shared records

Operative objects

EPI-OBJ-workstream-claims

Label: Workstream claims and findings

Part of: RT-OBJ-workstream

Source-native identity: Decisions, Learnings, Open Questions and completion/disposition prose inside workstream.md. Representational form: natural language. Storage substrate: consumer Markdown and intended Git history. This part differs from symbolic markers and normative task guidance in what must be checked: current artifacts and source evidence, not regex occurrence. SRC-1, .claude/skills/workstream-review/SKILL.md.

Move 1 asks what is new and unrecorded. This asks what is recorded and no longer true — the pass nothing else runs, because staleness has no event: nothing changes at the moment a Decision becomes false, so the detector has to be comparative, and wider than this file. --- .claude/skills/workstream-review/SKILL.md

EPI-OBJ-workstream-markers

Label: Workstream disposition markers

Part of: RT-OBJ-workstream

Source-native identity: dated terminal/deferred markers in Learning blocks, distinct from their semantic assertions. Representational form: symbolic for the parser consumer. Storage substrate: the same Markdown bytes and intended Git history. Consumers select/count/trim via finite patterns; they do not inspect the claimed destination. SRC-1, .claude/scripts/workstream_state.py.

if marker_counts(body, TERMINAL_RE): return 'terminal' --- .claude/scripts/workstream_state.py

EPI-OBJ-workstream-guidance

Label: Workstream plan and criteria

Part of: RT-OBJ-workstream

Source-native identity: backlog wording, critical-path guidance and Deletion Criteria. Representational form: natural-language policies interpreted by model/user, distinct from fixed-rule marker consumption. Storage substrate: consumer Markdown/Git. Facts invoked as premises remain truth-apt; choosing scope or approving a restructuring is normative. SRC-1, .claude/skills/workstream-review/SKILL.md.

Do not restructure on an assumption you have not surfaced; an option set the user reframes in free text is the signal the framing, not the options, was wrong. --- .claude/skills/workstream-review/SKILL.md

EPI-OBJ-record-learnings

Label: Record Learnings field

Distinct identity: derived JSON fields, not the retained source RT-OBJ-workstream or RT-OBJ-active. Source-native identity: count, terminal, deferred and undispositioned in record.learnings. Form: symbolic JSON. Substrate: process memory/stdout, no automatic retention established. SRC-1, .claude/scripts/workstream-record.py.

return {"count": count, "terminal": terminal, "deferred": deferred, "undispositioned": undispositioned} --- .claude/scripts/workstream-record.py

Routes

EPI-RTE-capture-synthesis

Label: Capture synthesis proposal

Part of: RT-RTE-state-lifecycle

Endpoints/progression: session findings and durable files → foreground model synthesis → Decision/Learning extension or supersession. Owner: foreground model following capture policy. Routine capture admits notes without per-claim user approval; existing substantive/user gates still apply. Trigger: session boundary. Intended use is future work; no universal evidence-consuming semantic acceptance criterion is supplied for new synthesis. Recovery uses scoped commits and supersession. The source does not individuate a produced theory; conditions of consumption, criticism and iteration for an actual theory instance remain uninspected, not inferred from the word synthesis. SRC-1, .claude/skills/workstream-capture/SKILL.md.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: requested routed findings and synthesis-level pattern.
• Later read-back: committed Decisions/Learnings are requested inputs to later work/review.
• Delegated visibility: uninspected — no live packet or host propagation supplied for this sweep.
• Selection predicate: findings not yet durable; synthesis asks what changes another piece of work.
• Invalidation or expiry: extend/supersede existing entries; later review may mark stale claims.
• Activation or effect: requested retained content can guide later model planning through file context; activation unobserved.
• Evidence limits: no actual input/output pair establishes entailed derivation or ampliative conjecture; generic model generation is insufficient.

Synthesis is a different question from detection, not a harder pass at it; the test is whether anything in the sweep's output would change how a DIFFERENT piece of work is done -- if nothing would, the step is still owed. --- .claude/skills/workstream-capture/SKILL.md

EPI-RTE-marker-measure

Label: Learning marker measurement

Distinct identity: deterministic record-field production, not the write route RT-RTE-rewrite-state or the host delivery route RT-RTE-session-reload. Endpoints: folded workstream Learning blocks → learnings_field/disposition → JSON classifications and counts. Owner: Python predicates. Target/domain: entries in the Learnings section and dated sentence/clause-start markers, not destination truth. There is no semantic answer oracle. Self-check compares predicates against hand-coded positives/negatives; it does not evaluate work completion. SRC-1, .claude/scripts/workstream-record.py, .claude/scripts/workstream_state.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: count, terminal number and deferred/undispositioned entries.
• Later read-back: inapplicable — this measurement does not persist changed state; later runs recompute from files.
• Delegated visibility: inapplicable — the inspected functions invoke no delegate.
• Selection predicate: Learning-shaped item block starting within the Learnings section; marker grammar prioritizes terminal over deferred.
• Invalidation or expiry: recomputed on next read; changed bytes may change classification.
• Activation or effect: classifications feed status and drain signals by code/prompt paths; actual behavior change unobserved.
• Evidence limits: predicate truth is not the truth of the marked claim or disposition.

if kind == 'item' and LEARNING_RE.match(raw[0]) and start in learnings_lines: count += 1 d = disposition(text) --- .claude/scripts/workstream-record.py

EPI-RTE-still-true

Label: Still-true comparison and withdrawal

Part of: RT-RTE-state-lifecycle

Endpoints: dated claims/plans → comparison with current repo, sibling workstreams and named commitments → stale finding, in-place supersession, approved substantive restructure. Foreground model proposes and investigates; user decides substantive restructure, while mechanical repairs proceed. Target is current applicability/truth of recorded premises, distinct from whether a premise still matters. Reference documents provide comparison material, not an oracle. Trigger: drift, staleness or request. Guidance persists as dated state/commits; rollback relies on history, not a software veto. SRC-1, .claude/skills/workstream-review/SKILL.md.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: drift/stale findings and requested user decisions.
• Later read-back: supersession marks and changed backlog/ACTIVE.md are committed for later model use.
• Delegated visibility: optional set-level adversarial subagent receives the set without workstream priors; host isolation uninspected.
• Selection predicate: dated Decisions/completed tasks, criteria, duplicate facts and framing assumptions; widen beyond the single file.
• Invalidation or expiry: mark superseded text when found, re-derive order and recheck at later drift.
• Activation or effect: model policy requests changed reliance and revised task selection; substantive changes need user approval.
• Evidence limits: current truth and revised reliance are not observed; no particular retained theory instance demonstrates conditions 1–4.

Read state against the project's DESIGNATED-AUTHORITY documents — the ADR, the design note, the spec — in both directions — a criterion and an ADR can contradict each other with neither referencing the other, and authority is not accuracy. --- .claude/skills/workstream-review/SKILL.md

Mark every stale hit in place at the moment you find it (~~superseded by ...~~, or a dated note saying what overtook it) rather than collecting them for a later pass. --- .claude/skills/workstream-review/SKILL.md

EPI-RTE-disposition-check

Label: Extraction destination verification

Part of: RT-RTE-state-lifecycle

Endpoints: resolved Learning/OQ and claimed disposition → inspect governed file/commit/destination → apply, hand off or drop with rationale. Foreground model checks and edits; handoff separately requires user destination/visibility confirmation. Trigger: drain symptoms, review or closure. Check target is whether the substance actually reached its governed destination; acceptance of a general proposition's truth is a different question. A future task is not an implemented result. Rejection is available through drop/reroute; commits preserve changes and rationale. SRC-1, .claude/skills/workstream-extract/SKILL.md.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: requested extraction/disposition account and edits.
• Later read-back: named permanent files and disposition markers guide future work and review.
• Delegated visibility: cross-project handoffs must supply self-contained context; actual receipt unobserved.
• Selection predicate: every resolved Learning and settled OQ; still-live entries remain; sweep already-applied entries too.
• Invalidation or expiry: verify destinations each pass; installed/generated copies remain provisional until governing source changed.
• Activation or effect: requested governed-file edits can change later guidance; retained marker alone cannot establish activation.
• Evidence limits: destination verification is not universal evidence for the explanatory content of a Learning; no observed accepted ampliative claim is supplied.

A file this project does not GOVERN — an installed copy, a generated file, a vendored template: provisional until the change reaches whatever governs the file; verify it there. --- .claude/skills/workstream-extract/SKILL.md

Only after Move 1, and the order is load-bearing: trimming before extracting destroys the insight you were about to route, and the trim looks successful either way. --- .claude/skills/workstream-extract/SKILL.md

EPI-RTE-learning-condense

Label: Terminal Learning sentence condensation

Part of: RT-RTE-rewrite-state

Endpoints: CLI learnings request → terminal-marker selection → sentence selection → learning_loss, count and finish fingerprint checks → dry report or file replacement. Owner: Python code, with caller supplying --write. Trigger/proposal: extract recommends trimming terminal entries; caller chooses applicability, no per-content human check enforced. Oracle is literal presence/count/structure, not source truth or semantic preservation. Refusal precedes this CLI's write; direct file writes bypass it; recovery needs prior Git retention. SRC-1, .claude/scripts/workstream-rewrite.py, .claude/scripts/workstream_state.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: entry/byte report, refusal, NO CHANGE or WRITTEN.
• Later read-back: with --write, selected sentences replace source prose for later hook/model invocations.
• Delegated visibility: inapplicable — inspected dispatch/transform/write path invokes no delegate.
• Selection predicate: terminal Learning, recognized prefix, multiple sentences, no later addendum; retain first statement, provenance plus following statement when applicable, target sentences and latest terminal-marker sentence.
• Invalidation or expiry: later writes replace text; count or preservation mismatch refuses before finish/write.
• Activation or effect: executable text reduction and local write veto; semantic effect on later consumers unobserved.
• Evidence limits: literal statement/target/disposition checks do not protect every qualification or mechanism sentence.

Concrete deduction, not target execution: inside an otherwise conforming ## Learnings section, take - L1 (2026-01-01): The probe works. Only on Linux. Extra explanation is long enough to remove. APPLIED 2026-01-02. The sentence splitter separates at punctuation plus whitespace. APPLIED at a sentence start with a date makes disposition terminal. With no provenance, integration-target sentence or addendum, condense_learning keeps sentence zero and the disposition, yielding joined text - L1 (2026-01-01): The probe works. APPLIED 2026-01-02. learning_loss finds its required statement and last marker intact; Learning count and heading/checkbox fingerprint are unchanged. The qualifier is dropped. This demonstrates a permitted loss of scope, not an actual deployed false statement. For arbitrary entries the relation remains indeterminate; it is not certified non-ampliative reshaping.

keep = set(range(2 if PROVENANCE_RE.match(sentences[0]) else 1)) keep.update(i for i, x in enumerate(sentences) if TARGET_RE.search(x)) keep = {i for i in keep if sentences[i] != disp and not ADDENDUM_RE.match(sentences[i])} condensed = ' '.join([sentences[i] for i in sorted(keep)] + [disp]) --- .claude/scripts/workstream-rewrite.py

if fingerprint(new_lines) != fingerprint(old_lines): sys.stderr.write("%s: structure changed (headings or checkbox IDs/states); refusing to write\n" % PROG) return 1 --- .claude/scripts/workstream-rewrite.py

EPI-RTE-status-statement

Label: Status synthesis and citation check

Distinct identity: on-demand cross-workstream chat production/checking, not RT-RTE-session-reload's hook output or RT-RTE-state-lifecycle's retained updates. Endpoints: bounded records/files → foreground statement → count/calibration checks → scout citation occurrences/context → correction/drop → user. Owner: foreground interprets; scout reports occurrences, not a verdict. Criteria: count agreement, calibrated patterns and matching cited text, with unverified premise labels. Scope excludes external premise verification and priority among independent chains. Operating mode: requested read-only project status. No automated product admission or rollback is present; report correction precedes delivery. SRC-1, .claude/skills/workstream-status/SKILL.md.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: roster, chains, gates, disagreements and coverage gaps in chat.
• Later read-back: inapplicable — status requests no file writes; optional user copying is separate.
• Delegated visibility: statement and records to fresh scout for cited claim text and surrounding context.
• Selection predicate: all workstreams including paused, ACTIVE, archive/inbox, Git reads and PROJECT-declared context; bounded exception for empty PROJECT descriptions.
• Invalidation or expiry: rerun measures for current status; do not recall old hook output.
• Activation or effect: advisory user selection; actionable findings name the owning skill rather than applying changes.
• Evidence limits: matching citation occurrence does not establish every inference; actual statement/output not supplied.

A premise the statement repeats is marked where it appears — asserted by <file>, unverified here, inline, never disclaimed twenty lines later. The skill does not verify it; that would make a read-only skill a verifying one. --- .claude/skills/workstream-status/SKILL.md

A citation that comes back as anything but MATCH is corrected or dropped before delivery; reading a cited line against a claim is the check a grep cannot make. --- .claude/skills/workstream-status/SKILL.md

Claims

EPI-CLM-grounded-claims

Label: Grounded claims and safe goal work

Distinct identity: semantic grounding/safety rationale, not RT-CLM-rewrite-preservation's rewrite predicate. Claimed operation: cited completion evidence plus durable, mechanically checkable state supports safe goal-driven autonomy. SRC-1, docs/design.md, doctrine/design. Local measurements and requested artifact checks support a scoped protocol; no causal or deployed-safety guarantee follows.

The kit replaces that with grounded claims: a checkbox closes only with cited evidence (a commit, a command's output, a count), which the user can check at the gate. --- docs/design.md

Durable, mechanically-checkable state is what makes goal-driven autonomy safe. --- docs/design.md

EPI-CLM-gate-validation

Label: Text-only gate validation report

Distinct identity: attributed lifecycle test outcome, not RT-CLM-rewrite-preservation. Claimed operation: prior autonomous acceptance sessions honored all user-authority constraints from skill text alone. SRC-1, docs/design.md, reported operation. The inspectable driver is RT-RTE-acceptance-driver, not those runs' evidence. Prompt preapproval and artifact assertions cannot independently isolate skill text as the cause.

This was validated directly: in the kit's acceptance tests, fully autonomous sessions ran the create/work/close lifecycle and honored every user-authority constraint — no auto-starting work after creation, no auto-passing checkpoint gates, no self-certified closure — from the skill text alone. --- docs/design.md

fixture's own Learnings and applies them to its installed skill copies. Read

Session C's output and the fixture's Learnings (in its git log, since the

closure archives the workstream file) before deleting the fixture -- both

findings from the first two runs were visible only there, while the summary

said ALL CHECKS PASS.

--- tests/va3-acceptance.sh

The driver's own comment supplies a countercase to equating its green summary with defect-free guidance. It is reported operation only. Reading saved fixture evidence and separately updating the kit could support a later improvement route; this boundary supplies no resulting update and exercised downstream causal chain.