Flopwire epistemic report

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

Source-and-claim boundary

This assesses how retained conversations and messages become evidence for later agents, and what checks license that reliance. Scope, external dependencies and exclusions are those of boundary; source identity and evidence layers are in its Source register. All new source reads below are SRC-1 at the frozen commit. They supply implementation or attributed doctrine/design, not observed runs. The supplied runtime orders the investigation but does not bound it.

The material distinction is between evidence that a transcript reports something and evidence that its subject is true. History retrieval imports claims; lexical ranking does not accept them. Commit observation extraction adds a narrower programmatic check with addressed command/result provenance. No source supplied here authenticates arbitrary transcript statements against the outside world. This is a limit on those routes, not a system-complete negative about all shipped operations.

Entry point or operation	Source path (SRC-1)	Coverage and records	Limit and conclusion prevented
CLI dispatch and MCP registry	cmd/flopwire/main.go, cmd/flopwire/mcp.go	Direct dispatch/registry reads; supplied runtime covers retrieval, agent, serve, messaging, hook, upgrade and trial	Registration is not observed invocation
Agent extraction and shared upload	internal/agent/index.go, internal/devicesync/syncer.go	Supplied RT-RTE-transcript-index, RT-RTE-shared-upload	All parser and remote admission implementations not independently audited here; no universal source-fidelity conclusion
Search, grep, sessions, read, raw	cmd/flopwire/mcp.go, internal/retrieval/format/render.go, internal/retrieval/format/text.go	Registry, rendering call sites and complete text transformations directly read; RT-RTE-history-retrieval, EPI-RTE-text-render	Ranking quality and complete chronology unobserved
Digest and outline	internal/digest/digest.go, internal/retrieval/format/summary.go	Direct digest declarations/Fold and full summary renderer; EPI-OBJ-digest, EPI-RTE-digest-fold	Other fold helpers not fully inspected; no semantic-summary accuracy guarantee
Commit change/trace and MCP change	cmd/flopwire/change.go, internal/digest/evidence.go, internal/digest/evidence_stream.go, internal/digest/commits.go, internal/ingest/commit_evidence.go, internal/retrieval/change.go, internal/retrieval/local/change.go	Direct complete extraction state machine and projection, selected command parser and lookup ranges; EPI-RTE-commit-check, EPI-RTE-commit-store, EPI-RTE-change-lookup	Shell grammar/helper coverage partial; no exhaustive spoof resistance or sole-authorship claim
Hook and alternate delivery; retained threads	plugins/codex/flopwire/hooks/hooks.json, plugins/codex/flopwire/skills/messaging/SKILL.md	Direct hook manifest and full skill; supplied RT-RTE-hook-delivery, RT-RTE-adapter-delivery	Bus backend/thread visibility not independently inspected; delivery is not agreement or authorization
Redact CLI and local withdrawal	cmd/flopwire/redact.go, internal/localindex/redact.go, internal/redact/redact.go	Direct complete CLI, local maskRow/maskDigest and scanner opening; EPI-RTE-index-redact	Remote archive rewrite, tombstone replay and all secret rules not fully audited; no total erasure or secret-removal guarantee
Policy, sender admission, binary installation	internal/api/api.go, internal/api/bus.go, internal/upgrade/upgrade.go	Supplied RT-RTE-policy-admission, RT-RTE-binary-upgrade	Operational checks have narrower domains than truth or improvement
Qualification and parser oracle	scripts/qualify-agent-workflow.py, testdata/oracle/README.md, internal/transcript/oracle/oracle.go	Direct qualification check body, oracle contract and Assert; RT-RTE-workflow-trial, EPI-RTE-parser-parity	Other probe/bench/acceptance/nightly and source-native reports remain unassessed here; no complete evaluator inventory or global negative
Backup/verify/restore, import, setup, admin/deletion and web administration	cmd/flopwire/main.go	Direct dispatch coverage; boundary and runtime identify shipped responsibilities	Handler internals not independently traced here. Prevents system-complete absence claims about acceptance, recovery and withdrawal

The consequential claim found is RT-CLM-recovery-evidence. Additional tool guidance is declared as EPI-CLM-commit-evidence. No live harness trace, independently verified task outcome or causal comparison was supplied. These gaps prevent claims of exercised reliance, reliable truth recovery or improved capacity. Distributed-parametric processing is external: see RT-CMP-codex-model. Its exact parameters, mutation and endpoint resolution remain uninspected; no conclusion about parametric learning follows.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-localindex	Imported utterances about decisions, commands and outcomes; symbolic location/state metadata	see record	SRC-1: internal/localindex/store.go; see record passages	A container; truth varies per utterance, and initial speaker warrant is unknown
EPI-OBJ-digest	Selected prompt/reply fragments and derived activity fields	see record	SRC-1: internal/digest/digest.go	Navigation does not certify intent, success or approval
EPI-OBJ-commit-evidence	A captured command/result supports a commit observation with a possibly abbreviated hash	see record	SRC-1: internal/digest/evidence.go	Transcript authenticity and actual repository state not independently checked
RT-OBJ-standing-instruction	Primarily directives; message bodies can carry arbitrary factual claims	see record	SRC-1: internal/busrender/busrender.go; see record	Directives are not truth acceptance; sender identity does not establish body truth
RT-CMP-codex-model	Internal truth-apt units not determinable from accessible parameters	see record	SRC-1: scripts/qualify-agent-workflow.py; see record	External opaque model; trial outputs are separate from parameters

Authority-route ledger

Route ID: RT-RTE-transcript-index Route function: content transformation Architectural status: implemented Object/candidate ID: RT-OBJ-localindex Content/update relation: truth-apt transformation: acquisition/import Transition or check target: Native transcript to indexed messages Evaluator/condition and domain: Parser and policy; structural eligibility, not factual truth Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Rows or parse error Implemented force: Enforcing parser/policy admission Epistemic authority and scope: Speaker/source warrant unknown; provenance retained in addresses, not external verification Operational authority: Makes parsed utterances queryable Behavioral-authority path: Later CLI/MCP → selected retained text → advisory evidence → later request Evidence source ID and local anchor: SRC-1: internal/agent/index.go; see cited route Claim IDs: none Mismatch marker: none Gap/limit: No supplied execution or causal evidence

Route ID: RT-RTE-transcript-index; RT-RTE-shared-upload Route function: retention Architectural status: implemented Object/candidate ID: RT-OBJ-localindex Content/update relation: no content change Transition or check target: Extracted rows and uploaded archive/index inputs Evaluator/condition and domain: See routes; policy and transport constraints Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Retained local rows; shared visibility after parsing Implemented force: Storage and eligibility enforcement Epistemic authority and scope: Retention licenses no additional truth claim Operational authority: Permits subsequent retrieval subject to scope Behavioral-authority path: Retriever → durable index/archive → permissive access → later calls Evidence source ID and local anchor: SRC-1: internal/localindex/store.go, internal/devicesync/syncer.go; see records Claim IDs: none Mismatch marker: none Gap/limit: No supplied execution or causal evidence

Route ID: RT-RTE-history-retrieval Route function: operational admission/selection/consumption Architectural status: implemented Object/candidate ID: RT-OBJ-localindex Content/update relation: truth-apt transformation: non-ampliative reshaping Transition or check target: Query to selected excerpts Evaluator/condition and domain: Lexical ranking/filtering; relevance domain Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Ranked excerpts, empty/partial page or error Implemented force: Ranking, not truth acceptance Epistemic authority and scope: Evidence of retained text in selected scope; no global absence from empty results Operational authority: Supplies model-visible evidence Behavioral-authority path: External agent → tool result → advisory/ranking → current and later retained context Evidence source ID and local anchor: SRC-1: cmd/flopwire/mcp.go; see record Claim IDs: RT-CLM-recovery-evidence Mismatch marker: none Gap/limit: No supplied execution or causal evidence

Route ID: EPI-RTE-text-render Route function: content transformation Architectural status: implemented Object/candidate ID: RT-OBJ-localindex Content/update relation: truth-apt transformation: non-ampliative reshaping Transition or check target: Text to safe display/window Evaluator/condition and domain: Code-point replacement and byte/line budget; display domain Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Cleaned or clipped excerpt Implemented force: Enforcing displayed selection Epistemic authority and scope: Preserves ordinary ASCII propositions but can remove decisive context; raw/JSON paths differ Operational authority: Bounds what reader sees Behavioral-authority path: Renderer → tool/hook text → advisory content → invocation Evidence source ID and local anchor: SRC-1: internal/retrieval/format/text.go, internal/retrieval/format/render.go Claim IDs: none Mismatch marker: none Gap/limit: No supplied execution or causal evidence

Route ID: EPI-RTE-digest-fold Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-digest Content/update relation: truth-apt transformation: non-ampliative reshaping Transition or check target: Prompt and reply fragments to bounded navigation fields Evaluator/condition and domain: Ordinal and clipping rules, not semantic intent assessment Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: First selected prompt and last reply snippets Implemented force: Selection/ranking guidance Epistemic authority and scope: Only selected source fragments; derived counters require their separate count rules Operational authority: Allows session navigation before full reads Behavioral-authority path: Session/outline renderer → digest text → advisory → later query Evidence source ID and local anchor: SRC-1: internal/digest/digest.go, internal/retrieval/format/summary.go Claim IDs: none Mismatch marker: none Gap/limit: This row covers textual fields; inspected code also derives counters/commit fields, whose full semantics are not established by this classification

Route ID: EPI-RTE-commit-check Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-commit-evidence Content/update relation: truth-apt transformation: entailed derivation Transition or check target: Captured command/result to observed hash and execution outcome Evaluator/condition and domain: Deterministic shell/call pairing and output grammar over one source generation/session Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Addressed observation or no event; whole-call failure can coexist with printed commit Implemented force: Rejects unsupported pairings and limits inferred continuity Epistemic authority and scope: Conditional derivation from captured bytes under parser rules; no independent Git authentication Operational authority: Emits eligible observations for projection Behavioral-authority path: Extractor → event list → enforcing evidence eligibility → next storage step Evidence source ID and local anchor: SRC-1: internal/digest/evidence_stream.go, internal/digest/evidence.go Claim IDs: EPI-CLM-commit-evidence Mismatch marker: none Gap/limit: Entailment is within the encoded transcript interpretation. External commit existence depends on truthful harness/output premises; mixed wrappers may be omitted

Route ID: EPI-RTE-commit-store Route function: retention Architectural status: implemented Object/candidate ID: EPI-OBJ-commit-evidence Content/update relation: no content change Transition or check target: Eligible observations to durable projection Evaluator/condition and domain: Available referenced rows, active/non-superseded source, budget and transaction revision Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Inserted/kept IDs, deleted obsolete fingerprints, dirty/limited state or rollback Implemented force: Enforcing projection consistency Epistemic authority and scope: Retains conditional observations; does not accept historical authorship Operational authority: Makes admitted events available to lookup Behavioral-authority path: Projection → SQL events → permissive access → subsequent lookup Evidence source ID and local anchor: SRC-1: internal/ingest/commit_evidence.go Claim IDs: none Mismatch marker: none Gap/limit: A storage acceptance criterion is not evidence-consuming acceptance of a broader historical claim

Route ID: EPI-RTE-change-lookup Route function: operational admission/selection/consumption Architectural status: implemented Object/candidate ID: EPI-OBJ-commit-evidence Content/update relation: truth-apt transformation: non-ampliative reshaping Transition or check target: Requested full hashes to exact/prefix evidence matches Evaluator/condition and domain: Repository visibility, prefix predicate and cursor revision; index domain Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Evidence addresses, coverage flags, restart request Implemented force: Enforces page/scope; ranks no truth verdict Epistemic authority and scope: Exact equality or prefix relation over stored observation; prefix does not establish uniqueness Operational authority: Permits follow-up reads and requires restart for changed cursor state Behavioral-authority path: Change tool → JSON addresses/flags → advisory evidence, enforcing pagination → next call Evidence source ID and local anchor: SRC-1: internal/retrieval/change.go, internal/retrieval/local/change.go, cmd/flopwire/change.go Claim IDs: EPI-CLM-commit-evidence Mismatch marker: none Gap/limit: No supplied execution or causal evidence

Route ID: EPI-RTE-index-redact Route function: lineage/freshness/recovery Architectural status: implemented Object/candidate ID: RT-OBJ-localindex; EPI-OBJ-digest; EPI-OBJ-commit-evidence Content/update relation: truth-apt transformation: non-ampliative reshaping Transition or check target: Selected local message content and dependent metadata to masked state Evaluator/condition and domain: Explicit requested target; masking/storage code, not a truth evaluator Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Masked row, scrubbed digest/title, FTS rebuild and refreshed change events Implemented force: Enforcing local withdrawal Epistemic authority and scope: Removes availability, not a finding that the content is false Operational authority: Changes subsequent search/read/change evidence Behavioral-authority path: Local write transaction → index/derived views → enforcing selection → later queries Evidence source ID and local anchor: SRC-1: internal/localindex/redact.go, cmd/flopwire/redact.go Claim IDs: none Mismatch marker: none Gap/limit: Native harness file and already delivered contexts remain outside this local withdrawal; remote rewrite not fully audited

Route ID: RT-RTE-hook-delivery; RT-RTE-adapter-delivery Route function: operational admission/selection/consumption Architectural status: implemented Object/candidate ID: RT-OBJ-standing-instruction Content/update relation: no content change Transition or check target: Delivered envelopes to recipient context Evaluator/condition and domain: Lease/event/budget rules; host delivery domain Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Rendered context or retry/undelivered state Implemented force: Software delivery; model handling advisory Epistemic authority and scope: No truth license from receipt or sender alone Operational authority: Requests replies/actions within existing authorization Behavioral-authority path: See RT-BAP-message-guidance; model/hook or host context → advisory policy → recipient turns Evidence source ID and local anchor: SRC-1: internal/busrender/busrender.go; see runtime routes Claim IDs: none Mismatch marker: none Gap/limit: No observed compliance; historical thread reads are not new direct requests

Route ID: RT-RTE-policy-admission; RT-RTE-binary-upgrade Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: RT-OBJ-standing-instruction Content/update relation: non-truth-apt policy/content update: configuration or executable replacement Transition or check target: Policy, sender or binary candidate Evaluator/condition and domain: See records; credential/config validation or checksum/runnable version criteria Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Reject or persist/replace Implemented force: Enforcing operational admission Epistemic authority and scope: No acceptance of truth or proof of improvement Operational authority: Changes future eligible collection/delivery or executable behavior Behavioral-authority path: See records; daemon/server or executable loader → symbolic config/binary → enforcing → subsequent use Evidence source ID and local anchor: SRC-1: internal/api/api.go, internal/upgrade/upgrade.go; see records Claim IDs: none Mismatch marker: none Gap/limit: Separate mechanisms grouped only for their non-epistemic consequence; individual criteria and rollback limits remain in cited records

Route ID: RT-RTE-workflow-trial Route function: check/evidence production Architectural status: implemented Object/candidate ID: RT-CMP-codex-model Content/update relation: truth-apt transformation: entailed derivation Transition or check target: Synthetic discovery, source reading, refusal and correlated token reply Evaluator/condition and domain: Script-owned expected token, exact session/thread IDs and transcript command checks Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Failure or completed trial report Implemented force: Check failure raises; no automatic product promotion Epistemic authority and scope: Conditional test result for explicit CLI resume workflow Operational authority: Stops failed trial/report; no production admission authority Behavioral-authority path: Script → report/exception → enforcing trial continuation, advisory maintenance → trial/review Evidence source ID and local anchor: SRC-1: scripts/qualify-agent-workflow.py; see record Claim IDs: none Mismatch marker: none Gap/limit: No execution evidence; script does not qualify hooks, MCP, automatic wake, arbitrary research correctness or component benefit

Route ID: EPI-RTE-parser-parity Route function: check/evidence production Architectural status: implemented Object/candidate ID: RT-OBJ-localindex Content/update relation: truth-apt transformation: entailed derivation Transition or check target: Parser conversation/message projection versus FAD synthetic expected output Evaluator/condition and domain: Programmatic comparison after declared divergence drops; fixture parity domain Activation and timing: On explicit request or newly eligible input; operation unobserved Possible or observed result: Differences cause test errors Implemented force: Enforcing test outcome Epistemic authority and scope: Agreement on projected fixture values only; reference itself has documented limitations Operational authority: Can fail parser tests; no runtime truth gate Behavioral-authority path: Oracle Assert → testing error → enforcing test failure → test invocation Evidence source ID and local anchor: SRC-1: internal/transcript/oracle/oracle.go, testdata/oracle/README.md Claim IDs: none Mismatch marker: none Gap/limit: No supplied execution or causal evidence

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
RT-CLM-recovery-evidence	see record	SRC-1: plugins/codex/flopwire/skills/messaging/SKILL.md; doctrine/design	see record and RT-BAP-recovery-guidance	RT-RTE-history-retrieval, EPI-RTE-text-render	none supplied	No comparative outcome evidence	Implemented evidence access plus requested source/contradiction checks	Interpretation and reliable claim support depend on external reader
EPI-CLM-commit-evidence	see record	SRC-1: cmd/flopwire/change.go; doctrine/design	see record	EPI-RTE-commit-check, EPI-RTE-commit-store, EPI-RTE-change-lookup	none supplied	No independent Git/causal verification; encoded pairing rules bound inference	Captured command/result provenance with explicit prefix and coverage limits	Sole authorship, complete enumeration and truthful underlying transcript unestablished

Bounded conclusion

Flopwire retains and imports conversation claims for later use; it does not thereby endorse their subject matter. Source warrant is unknown for arbitrary imported utterances. Addresses and follow-up reads preserve a path back to captured evidence, while clipping, outlines, digests and redaction reduce what a consumer sees. The shipped skill asks readers to recover decisive text and distinguish proposals from approval. Its authority is advisory at the external model, over invocations loading the guidance. See RT-RTE-history-retrieval, RT-BAP-recovery-guidance, EPI-RTE-text-render and EPI-RTE-digest-fold.

Commit extraction is a supported narrower derivation and check route. It links captured command/result records, rejects selected ambiguous command patterns and preserves whole-call outcome separately from observed commit output. Thus a failing shell call is not sufficient evidence that no commit occurred. Its warrant is conditional on truthful captured inputs and the inspected interpretation, not independently warranted Git facts. Projection retention and change lookup give those observations later evidential force; prefix matching still cannot license unique identity or sole authorship. See EPI-RTE-commit-check, EPI-RTE-commit-store and EPI-RTE-change-lookup.

There is no evidence-consuming acceptance decision for an ampliative historical explanation in the inspected retrieval, digest and commit routes. This route-local result does not assert such acceptance is absent from unassessed evaluators or source-native maintenance practice. No lifecycle integration after such acceptance is established. Indexing, storage, visibility admission and operational consumption are separate functions. Local redaction withdraws content and updates derived readers; it does not retract already supplied model context or classify an assertion as false. See EPI-RTE-index-redact.

The workflow trial and parser oracle check named synthetic targets, with operative failure consumers. Their licenses do not transfer to general research truth, reliable messaging across all hosts or improved capacity. No trial was executed in this analysis. Policy/sender gates and checksum/version tests change operational authority against specific criteria; they establish neither epistemic endorsement nor successful improvement. Delivered requests can change a recipient's tasks without a truth-apt update, but receipt is not agreement. See RT-RTE-workflow-trial, EPI-RTE-parser-parity, RT-RTE-policy-admission, RT-RTE-binary-upgrade and RT-BAP-message-guidance.

These findings establish retained experience access, bounded deterministic derivation and operational admission capacities. They do not establish learning as improved future capacity, content-directed theory criticism and iteration, reflection's two causal directions or exercised self-improvement. Those remain separate questions requiring route-specific evidence; opaque provider internals and absent run comparisons do not justify a bundled negative. The unassessed operations in the coverage table prevent a system-complete negative.

Shared records

Operative objects

EPI-OBJ-digest

Label: Digest

Part of: RT-OBJ-localindex

Source-native identity: digest.Digest, JSON stored with conversation state. Natural-language Intent/Last snippets and symbolic counts, paths, hashes and fold bookkeeping are distinct operative parts. Text is interpreted by readers; counters and references have fixed renderer/selector rules. This is a material part of the supplied runtime's local-index container, not a replacement container. SRC-1: internal/digest/digest.go, internal/retrieval/format/summary.go.

// Two kinds of fields make it up: // // - Fold reads the messages a batch writes, in order: the intent (first // user prompt), the last assistant message, files edited, PRs, commits // and issues. Re-folding a message is harmless: these are sets, or keep // the lowest (intent) or highest (last) ordinal seen. --- internal/digest/digest.go

EPI-OBJ-commit-evidence

Label: CommitEvidence

Source-native identity: digest.CommitEvidence, symbolic fields and evidence references represented as Go values and persisted JSON/SQL projection entries. Distinct from the digest's bounded commit list and the runtime's undifferentiated local-index container: this object spans local and server projections and retains command/result lineage. Closest supplied record is RT-OBJ-localindex; it supplies the local storage container, not the cross-backend observation referent. SRC-1: internal/digest/evidence.go, internal/ingest/commit_evidence.go.

// CommitEvidence is an observed commit, with the records supporting that observation. // Hash may be abbreviated. DirectoryKnown describes captured directory placement, // never repository identity. Outcome describes the whole execution, not commit creation. --- internal/digest/evidence.go

Routes

EPI-RTE-text-render

Label: Text rendering

Part of: RT-RTE-history-retrieval

Source-native Clean, Cut and renderer calls are symbolic Go transformations of text. This is the display subroute of the supplied history-retrieval route. The renderer calls Clean on message text; the bus also cleans and escapes message bodies. A concrete implementation deduction: Clean on the character sequence A, carriage return, newline, B, ESC yields A, newline, B, ␛. Cut on approved\nreversed with from=1 and maxChars=8 returns approved with Clipped=true: omission can change a historical interpretation. These are source deductions, not executed target examples. SRC-1: internal/retrieval/format/text.go, internal/retrieval/format/render.go, internal/busrender/busrender.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Cleaned display string or bounded excerpt
• Later read-back: inapplicable — this display function itself writes no retained result
• Delegated visibility: inapplicable — this subroute does not delegate a model task; host delivery remains in supplied runtime routes
• Selection predicate: Selected output format and line/byte budgets
• Invalidation or expiry: inapplicable — transient output; raw/JSON branches preserve different representations
• Activation or effect: Code changes displayed characters and visible context
• Evidence limits: Does not certify semantic equivalence for all Unicode/control-dependent inputs or full context

// characters (isBidiControl) become U+FFFD, a CR before a newline is dropped, and invalid // UTF-8 is replaced. JSON output never goes through Clean; it is exact. --- internal/retrieval/format/text.go

text := Clean(m.Text) --- internal/retrieval/format/render.go

e.Text = b.String() if e.From > 1 || e.To < e.Lines { e.Clipped = true } --- internal/retrieval/format/text.go

EPI-RTE-digest-fold

Label: Digest folding

Part of: RT-RTE-transcript-index

Indexer-time code folds message snippets and symbolic activity into Digest; rendering supplies navigation summaries. Closest supplied route is RT-RTE-transcript-index. This is its digest-specific subroute, with different output/checks from full extracted messages. User snippets use lowest eligible ordinal; assistant Last uses highest. The label Intent is not an evaluated explanation of intent. SRC-1: internal/digest/digest.go, internal/retrieval/format/summary.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Updated Digest fields
• Later read-back: Stored conversation digest reaches later sessions/outline renderers
• Delegated visibility: inapplicable — this subroute does not delegate a model task; host delivery remains in supplied runtime routes
• Selection predicate: Non-superseded, non-evidence-only rows; eligible user/assistant text and ordinal rules
• Invalidation or expiry: Reparse can replace same-ordinal text; local masking can scrub derived fields
• Activation or effect: Supplies shorter navigation text and activity fields
• Evidence limits: Textual selection shown directly; count/store helpers and all semantic implications are not exhaustively inspected

case transcript.KindAssistant: if t := oneLine(m.Text); t != "" { if d.State == nil || d.State.LastOrd == nil || m.Ordinal >= *d.State.LastOrd { d.state().LastOrd = ptr(m.Ordinal) d.Last = clip(t, LastLen) } } --- internal/digest/digest.go

EPI-RTE-commit-check

Label: Commit evidence extraction

Owner: deterministic CommitEvidenceState/ExtractCommitEvidence, over source-generation/session-ordered retained messages. It pairs addressed shell calls/results, tracks running handles, invalidates open HEAD continuity on uncertain/moving commands, and emits observation events. Distinct from RT-RTE-history-retrieval: this produces provenance-bearing observations before retrieval, not lexical hits. Different from ordinary RT-RTE-transcript-index parsing: it checks cross-message Git evidence. SRC-1: internal/digest/evidence.go, internal/digest/evidence_stream.go, internal/digest/commits.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: New CommitEvidence events or empty list
• Later read-back: Events can enter local/server projections, then change lookup
• Delegated visibility: inapplicable — this subroute does not delegate a model task; host delivery remains in supplied runtime routes
• Selection predicate: Typed call identity, shell grammar, safe output sources, printed hash or permitted successful HEAD reveal
• Invalidation or expiry: Superseded rows, HEAD movement and unsupported placement break inferred continuity
• Activation or effect: Emits conditional observation with command/result references and separate whole-call outcome
• Evidence limits: No external Git execution, source authentication or independent truth oracle; full shell grammar not audited

Proposer: retained captured commands/results. Decider/veto: code pairing and eligibility predicates; no per-observation human approval. Mode: open transcript evidence projection. Answer oracle: no externally provided expected commit; captured output is evidence input. This admits derived observations, not a revised explanatory theory.

// Other programs can print convincing Git output. Reject such compound calls. --- internal/digest/evidence.go

if p.g.printed && p.printedSafe { for _, hash := range printedEvidenceHashes(out) { e := base e.Hash = hash events = append(events, e) printed = true } } --- internal/digest/evidence_stream.go

// Commands enrichment is best effort, including heuristic Codex event joins. // Parse the addressed call itself to establish command ownership. --- internal/digest/evidence_stream.go

EPI-RTE-commit-store

Label: Commit projection retention

Server RefreshCommitEvidence reads retained live tool rows grouped by source/generation, invokes extraction and atomically replaces the derived projection while retaining unchanged fingerprints. Closest supplied RT-RTE-shared-upload describes shared ingestion, but this is a distinct derived-evidence persistence operation. SRC-1: internal/ingest/commit_evidence.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Transaction success, error or limited/dirty state
• Later read-back: Later shared Change queries join evidence and source/conversation visibility
• Delegated visibility: inapplicable — this subroute does not delegate a model task; host delivery remains in supplied runtime routes
• Selection predicate: Live non-superseded rows on active path, no tombstoned source; byte/row budgets and available references
• Invalidation or expiry: Deletes fingerprints not regenerated; revision mismatch requires rollback; limits leave dirty state
• Activation or effect: Updates later available observations and preserves provenance references
• Evidence limits: Inspected server transaction only; not proof of observed crash safety or all local projection semantics

Admission trigger: parsed/dirty conversation. Proposed change: rebuilt observation set. Software decides and can reject missing referenced messages or changed snapshot; no human approval or truth oracle. Recovery: entire transaction rollback on inconsistent projection. This criterion admits storage consistency for lookup, not historical explanatory acceptance.

// RefreshCommitEvidence atomically replaces one conversation's projection from // retained live rows. The caller holds its conversation row lock. Appended rows // can complete an older call; replay is source scoped, never cross-source. // Unchanged observations keep their IDs so appends do not disturb open pages. --- internal/ingest/commit_evidence.go

EPI-RTE-change-lookup

Label: Change lookup

Part of: RT-RTE-history-retrieval

CLI/MCP change lookup selects projected observations by full requested IDs and repository scope, maps exact/prefix matches, and returns addresses plus coverage/restart state. Material part of the supplied retrieval container route; no supersession needed because that container remains valid. SRC-1: cmd/flopwire/change.go, internal/retrieval/change.go, internal/retrieval/local/change.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: JSON evidence addresses, match kind, coverage flags or error
• Later read-back: Reads retained observations; request itself does not persist a revised historical claim
• Delegated visibility: inapplicable — this subroute does not delegate a model task; host delivery remains in supplied runtime routes
• Selection predicate: Repository/visibility scope, requested hash prefix equality, cursor/cutoff and budget
• Invalidation or expiry: Index/revision changes invalidate pagination; dirty/quarantined/limited states qualify coverage
• Activation or effect: Guides source reads and can refuse stale pagination
• Evidence limits: No complete-history, unique-prefix or sole-author claim; backend source inspected in bounded ranges

Selector proposer: user/model supplies IDs and repository; decider: code matches all qualifying stored prefixes. No human selection gate or expected-answer oracle. Open lookup mode. Reader still decides historical attribution after reading evidence.

// Change returns indexed observations, including all submitted IDs that share // an abbreviated observation. Prefix observations never establish uniqueness. --- internal/retrieval/change.go

if prior.Revision != cursor.Revision { out.RestartRequired = true return nil } --- internal/retrieval/change.go

EPI-RTE-index-redact

Label: Local index redaction

Explicit redact request routes through daemon or direct local store; local maskRow masks selected row text, digest/title and FTS entries and refreshes change events. Distinct from collection deny/rewrite described in RT-RTE-transcript-index: this withdraws selected content and dependent evidence. The CLI also requests remote redaction separately; its complete server archive mechanics remain unassessed. SRC-1: cmd/flopwire/redact.go, internal/localindex/redact.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Count or error; server and local results separate
• Later read-back: Subsequent local search/read/change sees masked/recomputed views
• Delegated visibility: inapplicable — this subroute does not delegate a model task; host delivery remains in supplied runtime routes
• Selection predicate: Explicit address/line range and copy options; local target resolution
• Invalidation or expiry: Masked text replaces original in index; native file unchanged; delivered contexts not retracted
• Activation or effect: Withdraws available text and dependent projection evidence
• Evidence limits: No total secret detection or cross-device erasure guarantee; tombstone replay/remote archive implementation not fully inspected

Proposer: request principal chooses target; code resolves/masks it, with remote credential/ownership distinctions exposed by CLI. Generic caller identity does not prove human action. Explicit maintenance mode; no correctness oracle. Failure can stop the operation; server success can precede a local error. Withdrawal expresses availability control, not epistemic falsification.

if err := w.maskDigest(t.conv, hidden); err != nil { return Tombstone{}, err } --- internal/localindex/redact.go

w.ftsInsert(t.id, p) if err := w.refreshChangeEvents(t.conv); err != nil { return Tombstone{}, err } --- internal/localindex/redact.go

// device's local index. The harness's own transcript file is not touched. --- cmd/flopwire/redact.go

EPI-RTE-parser-parity

Label: Parser parity oracle

Fixture oracle compares parser conversation/message projections to committed synthetic FAD outputs, after named divergence exclusions. Assert consumes differences as test errors. No supplied runtime record describes this comparator; closest RT-RTE-workflow-trial is a distinct real-model synthetic messaging trial. SRC-1: testdata/oracle/README.md, internal/transcript/oracle/oracle.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Test errors for differences
• Later read-back: inapplicable — this test comparator does not itself install retained runtime content
• Delegated visibility: inapplicable — this subroute does not delegate a model task; host delivery remains in supplied runtime routes
• Selection predicate: Explicit parser test invocation, fixture and projection/divergence rules
• Invalidation or expiry: Fixture/reference edits change comparison; excluded divergences limit target
• Activation or effect: Test failure/reporting, not runtime admission of utterance truth
• Evidence limits: No tests executed; reference parser has documented limits; no production outcome or causal comparison

Proposer: fixture/test author. Decider: projected differential comparator; maintainer chooses expected output and divergences. Answer oracle: committed FAD 0.3.1 expected values, subject to declared limitations. Bounded test mode, not open historical interpretation; failure rejects the test outcome, not automatically the installed binary.

// Assert reports Diff's problems on t, prefixed with the expected file name. func Assert(t testing.TB, f File, exp Conversation, conv transcript.Conversation, msgs []transcript.Message, rules Rules) { t.Helper() for _, p := range Diff(exp, conv, msgs, rules) { t.Errorf("%s: %s", f.Name, p) } }

--- internal/transcript/oracle/oracle.go

Every fixture here is synthetic. Never copy real transcript content into this tree. --- testdata/oracle/README.md

Claims

EPI-CLM-commit-evidence

Label: Captured commit evidence

The MCP change tool promises sessions with captured evidence and source addresses, while explicitly withholding prefix uniqueness. This narrows rather than contradicts RT-CLM-recovery-evidence: it addresses change provenance, not arbitrary research recovery. SRC-1: cmd/flopwire/change.go; doctrine/design tool description implemented by the lookup/extractor routes.

Read the evidence before attributing a change; a prefix observation remains a prefix match. --- cmd/flopwire/change.go

hint := "Read relevant command/result addresses, compare the evidence with the requested claim, and state remaining gaps." --- cmd/flopwire/change.go