Tracecraft epistemic report

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

Source-and-claim boundary

The question is whether Tracecraft checks retained content for truth, or instead preserves caller content and gives coordination state operational force. Scope, exclusions and source identity are see boundary and its Source register, SRC-1. External model hosts, backend internals and actual deployment grants remain excluded. Their missing evidence prevents conclusions about model reasoning, downstream compliance and universal service guarantees. No external model parameters are exposed by these SDK paths; external-host learning is not determinable from them.

This is a sparse overlay on the supplied runtime records. Direct inspection below checks dispatch and selected transformations/evaluators against that provisional account. No target execution occurred. Raw benchmark results and generated reports remain available but uninspected: numerical performance, observed invariants and causal benefit are not established here. Actual memory/message/artifact/transcript instances were not supplied, so individual propositions, their original sources and their truth are not determinable. Architecture can nevertheless be traced.

Entry point or operation	SRC-1 source path	Coverage and evidence basis	Limit and conclusion prevented
CLI registration	sdk/tracecraft/cli/__init__.py	Direct complete read; registered operations match supplied runtime	No additional model tool registry or content evaluator exposed in this dispatch
Init, backend and direct writes	sdk/tracecraft/store.py, sdk/tracecraft/s3.py, sdk/tracecraft/hf.py, sdk/tracecraft/cli/init_cmd.py	Supplied RT-RTE-init-config, RT-CMP-get-store	Service authentication and deployed conditional-write behavior unobserved; direct writes bypass CLI checks
Memory set/get/history/list	sdk/tracecraft/cli/memory.py	Direct complete read; RT-RTE-memory-write	No value-level semantic criterion in this path; attribution strings are not authenticated authorship
Send/inbox/delete and cursor	sdk/tracecraft/cli/messages.py, sdk/tracecraft/protocol.py	Supplied RT-RTE-mail-exchange	No independent inspection of all cursor branches here; no system-complete delivery guarantee
Claim/complete/status/wait and git evidence	sdk/tracecraft/cli/steps.py	Direct complete read; RT-RTE-step-control, EPI-RTE-git-files	Completion is not a correctness judgment; git failure becomes an ambiguous empty list
Agents and aggregate status	sdk/tracecraft/cli/agents.py, sdk/tracecraft/cli/status.py	Supplied runtime account and RT-BAP-step-polling	Staleness is heartbeat age, not measured liveness; aggregate state is not transactional
Artifact upload/download/list	sdk/tracecraft/cli/artifacts.py	Supplied RT-RTE-artifact-share	Opaque bytes; truth-aptness and later interpretation not determinable
Mirror, follow, follow-all and discovery	sdk/tracecraft/cli/session.py, sdk/tracecraft/harness/__init__.py, sdk/tracecraft/harness/hermes.py	Direct flush read and Hermes row serialization; other discovery/control branches supplied by RT-RTE-session-mirror	Host content production and project isolation for non-Claude adapters not established
Redact	sdk/tracecraft/redact.py	Direct complete read; EPI-RTE-regex-scrub	Pattern membership is not secret identity or transcript truth
Session show/stop/compact	sdk/tracecraft/cli/session.py	Direct complete operation reads; RT-RTE-session-compact	Stop resets cursor, not retention or process cancellation; interrupted compact can duplicate later display
Shipped verification instruction	examples/hermes-agent/SKILL.md	Direct complete read; EPI-RTE-read-verification, RT-BAP-agent-prompt	Requested read-back is not observed agent checking or semantic review
Benchmark evaluators	benchmarks/contention_bench.py, benchmarks/messaging_bench.py	Direct selected trial/read-back functions; EPI-RTE-bucket-evaluation	Preflight/sweep/report and frozen results not independently assessed; no numerical or causal finding
Test/release checks	sdk/tests/test_v025_features.py, .github/workflows/ci.yml, .github/workflows/release.yml	Direct memory test sample; workflow findings supplied by RT-RTE-package-release	Not exhaustive test adequacy; no observed passing CI or publication

Consequential warrant claims are see RT-CLM-mirror-idempotence, RT-CLM-secret-redaction, EPI-CLM-git-truth and EPI-CLM-history-loss. These concern preservation, confidentiality and local evidence, not a declared general theory-production engine. The selected reads do not support an unqualified system-complete negative about all possible caller content or uninspected evaluation reports.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-step-state	Caller declaration of work state; symbolic status also controls tooling	see record	SRC-1 sdk/tracecraft/cli/steps.py	Complete names no substantive criterion, intended-use evidence or independent judge
EPI-OBJ-handoff	Optional note assertions, optional git-derived file list; next-action is a directive, not necessarily truth-apt	see record	SRC-1 sdk/tracecraft/cli/steps.py	Mixed object split by consumers below; no supplied instance
RT-RTE-memory-write	Caller string may assert facts; key/writer/time are symbolic envelope	see record	SRC-1 sdk/tracecraft/cli/memory.py	No semantic restrictions; metadata-only get preserves attribution, default get omits it
RT-RTE-mail-exchange	Caller message may assert facts or request actions	see record	SRC-1 sdk/tracecraft/cli/messages.py	Acquired communication, not evidence of produced accepted knowledge
RT-RTE-artifact-share	Not determinable for arbitrary file bytes	see record	SRC-1 sdk/tracecraft/cli/artifacts.py	No artifact contents or parser supplied
EPI-OBJ-session-parts	Records of host utterances/tools may contain assertions, outputs or directives	see record	SRC-1 sdk/tracecraft/cli/session.py	A transcript records an assertion without endorsing it; redaction removes original details
EPI-RTE-bucket-evaluation	Computed owner-consistency and delivered-body-count propositions in test domains	see record	SRC-1 benchmarks/contention_bench.py, benchmarks/messaging_bench.py	Evaluator source, not observed outcome; no truth oracle for message bodies

Authority-route ledger

All rows are architectural findings. Observed activation is uninspected unless stated otherwise. Fields owned by supplied records are referenced rather than restated as new declarations. The same cited route appears in distinct functional rows deliberately: checking, disposition, retention and consumption have different licenses.

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-memory-write	content transformation	implemented	see record	truth-apt transformation: acquisition/import	Caller value to envelope	Structural key validation, not value truth	see record	Stored text and attribution	Software vetoes invalid key only	Source warrant unknown; value not endorsed	Allows valid-key write	see record	SRC-1 sdk/tracecraft/cli/memory.py	none	none	Human/model authorship unresolved
RT-RTE-memory-write	retention	implemented	see record	no content change	History then live value	Timestamp/unique key, ordinary overwrite	see record	Snapshot plus current value	Persistence and explicit reads	Availability and claimed attribution only	Later get/history deliver text	see record	SRC-1 sdk/tracecraft/cli/memory.py	EPI-CLM-history-loss	Immutable is not enforced	Direct writes/deletes and partial failure remain possible
RT-RTE-mail-exchange, RT-RTE-artifact-share	operational admission/selection/consumption	implemented	see records	no content change	Retained mail/files to requesting reader	see records	see records	Printed or downloaded content	Delivery, not content execution	No semantic warrant added	Reader can receive selected retained data	see records	SRC-1 sdk/tracecraft/cli/messages.py, sdk/tracecraft/cli/artifacts.py	none	none	Truth-aptness of artifacts indeterminate; mail source warrant unknown
RT-RTE-step-control	check/evidence production	implemented	RT-OBJ-step-state	no content change	Claim conflict and completion ownership	Conditional service write or agent-string comparison	see record	Conflict/mismatch or admissible completion	Structural/cooperative veto	License over claim admission only	Rejects competing claim or mismatched completion unless override	see record	SRC-1 sdk/tracecraft/cli/steps.py	none	none	HF arbitration differs; not work evaluation
RT-RTE-step-control	disposition/acceptance	implemented	RT-OBJ-step-state, EPI-OBJ-handoff	truth-apt transformation: acquisition/import	Caller outcome declaration	Caller flags; no semantic evaluator	see record	Complete, blocked, needs_review	Persists declaration	Operational admission, not evidence-consuming epistemic acceptance	Makes symbolic state available to polling	see record	SRC-1 sdk/tracecraft/cli/steps.py	none	none	No separate reviewer required to clear review state
RT-BAP-step-polling	operational admission/selection/consumption	implemented	RT-OBJ-step-state	no content change	Dependency readiness	Fixed status comparison	see record	Return, wait or error	Enforcing within waiter	Status equality only, not actual completed-work truth	Complete permits return; blocked fails	see record	SRC-1 sdk/tracecraft/cli/steps.py	none	none	No control over external work; complete precedes handoff write
EPI-RTE-git-files	check/evidence production	implemented	EPI-OBJ-handoff	truth-apt transformation: acquisition/import	Local tracked diff paths versus HEAD	Git exit status and name output	see record	List or empty fallback	Optional field writer	Conditional local diff evidence, not all activity	Handoff receives changed_files	see record	SRC-1 sdk/tracecraft/cli/steps.py	EPI-CLM-git-truth	Empty conflates failure and no tracked change	Untracked files, committed changes and content correctness outside license
RT-RTE-session-mirror	content transformation	implemented	EPI-OBJ-session-parts	truth-apt transformation: acquisition/import	Host records into delta stream	see record	see record	File bytes or serialized Hermes rows	Imports records	Original assertion warrant unknown; attribution retained only to extent host data supplies it	Makes host records available for retention	see record	SRC-1 sdk/tracecraft/harness/hermes.py, sdk/tracecraft/cli/session.py	none	none	Imported host claims are not Tracecraft-produced conclusions
EPI-RTE-regex-scrub	content transformation	implemented	EPI-OBJ-session-parts	truth-apt transformation: indeterminate	Byte-pattern replacement	Regex membership, not proposition truth	see record	Redacted bytes plus counts	Replaces matching substrings before mirror upload	Original matched details unavailable; untouched assertion warrant still unknown	Consumer sees modified stream	see record	SRC-1 sdk/tracecraft/redact.py	RT-CLM-secret-redaction	No privacy/truth equivalence	Cannot establish semantic preservation for arbitrary transcripts
RT-RTE-session-mirror	retention	implemented	EPI-OBJ-session-parts	no content change	Transformed delta to bucket parts/meta/cursor	see record	see record	Durable transformed trace	Storage and later show	No semantic endorsement	Explicit show can retrieve parts	see record	SRC-1 sdk/tracecraft/cli/session.py	RT-CLM-mirror-idempotence	Replay possible after cursor failure	Stop does not withdraw retained assertions
RT-RTE-session-compact	content transformation	implemented	EPI-OBJ-session-parts	truth-apt transformation: non-ampliative reshaping	Ordered part bytes to concatenation	Fixed byte-copy code	see record	Same byte sequence on completed normal path	Storage-layout replacement	No new semantic warrant	Show reads merged content after originals removed	see record	SRC-1 sdk/tracecraft/cli/session.py	none	none	Partial completion changes replay multiplicity; lineage metadata loses per-part detail
EPI-RTE-read-verification	check/evidence production	doctrine only	Caller retained state	no content change	Write/read correspondence	Host agent interpreting matching reads	see record	Potential write verification	Advisory instruction only	At most write/read agreement; not truth of value	Requests subsequent read or recipient confirmation	see record	SRC-1 examples/hermes-agent/SKILL.md	none	none	No checked result or uptake supplied
EPI-RTE-bucket-evaluation	check/evidence production	implemented	Trial results	truth-apt transformation: entailed derivation	API outcomes versus durable owner; intended sends versus distinct bodies	Fixed arithmetic/equality under trial inputs	see record	Invariant boolean, duplicate/loss counts	Trial reporting only	Conditional conclusions in configured benchmark domain	Invalidates trial on unexpected errors; records result	see record	SRC-1 benchmarks/contention_bench.py, benchmarks/messaging_bench.py	none	none	No route from these results to acceptance of message meaning or automatic SDK policy revision

No lifecycle integration after semantic acceptance is evidenced by these inspected writes/readers: they are retention and operational use before any such acceptance. This is a bounded account of these mechanisms, not a negative about what an external agent might implement. CI/package disposition is separately see RT-RTE-package-release: packaging checks have publication force, but neither sampled storage assertions nor release smoke checks judge the truth of retained claims.

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-mirror-idempotence	see record	SRC-1 sdk/tracecraft/cli/session.py; doctrine/design	see record	RT-RTE-session-mirror	Uninspected	No intervention evidence	Cursor suppresses normal repeated ingestion	Part success before cursor write permits repeated bytes
RT-CLM-secret-redaction	see record	SRC-1 sdk/tracecraft/redact.py; doctrine/design and implementation	Explicit denylist limits	EPI-RTE-regex-scrub	Uninspected	No disclosure experiment	Recognizable shapes replaced and counted	Counts do not establish all secrets removed or original truth preserved
EPI-CLM-git-truth	see record	SRC-1 sdk/tracecraft/cli/steps.py; doctrine/design	Optional non-hand-typed list	EPI-RTE-git-files	Uninspected	No support for quoted frequency claim	Imported tracked diff paths when subprocess succeeds	Not exhaustive work evidence; failure/no-change both empty
EPI-CLM-history-loss	see record	SRC-1 sdk/tracecraft/cli/memory.py; doctrine/design	Append snapshot before overwrite	RT-RTE-memory-write	Uninspected	No crash/durability experiment	Normal CLI set creates historical copy	No enforced immutability or unconditional no-loss guarantee

Bounded conclusion

Tracecraft retains and retrieves caller assertions, messages, handoff text, arbitrary artifacts and host records without licensing their substantive truth. The memory writer records attribution and time, but default get prints only the value; historical availability and freshness are not endorsement. Messages and artifacts supply selected data to later invocations; external-agent interpretation and compliance remain unobserved. Evidence is see RT-RTE-memory-write, RT-RTE-mail-exchange and RT-RTE-artifact-share.

Step status has concrete enforcing force at the dependency waiter, not merely advisory text. Complete permits return, blocked raises and review-needed waits. The completion writer admits the caller's outcome without an evidence-consuming criterion for work correctness. Clearing review does not require a separate reviewer. The handoff's directive has no implemented scheduler consumer here, and its optional changed-files field imports a narrower local Git measurement. These paths give operational authority without semantic acceptance: see RT-RTE-step-control, RT-BAP-step-polling and EPI-RTE-git-files.

Mirroring acquires host-produced content whose source warrant is unknown. Redaction deliberately degrades access to original matched details; it does not verify the remaining assertions. Compaction preserves ordered bytes on the completed normal path, not meaning by a new inference, and interrupted cleanup can duplicate displayed material. Show selects tail lines, not a resumed model conversation. Stop removes local cursor and stamps metadata, not epistemic withdrawal. See EPI-RTE-regex-scrub, RT-RTE-session-mirror and RT-RTE-session-compact.

There are narrow truth-apt checks: benchmark equality/arithmetic derives propositions from supplied trial outcomes and durable reads; the memory test compares returned v2 to a reference outcome. Their premises would require execution evidence for actual result claims. They do not check message meaning, explanation quality, transfer or model reasoning. The Hermes read-verification instruction requests a possible check but supplies no operative checked result. No demonstrated learning benefit, reflective criticism loop or exercised self-improvement follows from retaining raw traces or reporting storage results. These independent properties remain unestablished within these routes, not disproved for excluded hosts. See EPI-RTE-bucket-evaluation, EPI-RTE-read-verification and RT-RTE-package-release.

Shared records

Operative objects

EPI-OBJ-handoff

Label: Structured handoff

Part of: RT-RTE-step-control

Material output of that route, not a replacement for its claim/status control account. Source-native identity: steps/<sid>/handoff.json; bucket JSON contains natural-language note/next-action, symbolic state/recipient/schema and optional symbolic changed_files. No counterpart object for the handoff is declared in supplied runtime; RT-OBJ-step-state is the separate claim/status pair. The status poller reads that pair rather than the handoff. SRC-1 implementation, sdk/tracecraft/cli/steps.py:

store.put_json(f"steps/{sid}/handoff.json", handoff) --- sdk/tracecraft/cli/steps.py

EPI-OBJ-session-parts

Label: Transcript parts

Part of: RT-RTE-session-mirror

Material retained output of the route, not another declaration of mirroring. Source-native identity: numbered part-*.jsonl objects; substrate is configured bucket. Form is serialized host records containing natural language and symbolic tool/row metadata; opaque content remains possible. RT-RTE-session-compact later rearranges these same objects. No counterpart operative-object declaration in supplied runtime. SRC-1 implementation, sdk/tracecraft/cli/session.py:

part_key = f"{_session_prefix(harness_name, sess.session_id)}part-{seq:05d}-{uniq}.jsonl" --- sdk/tracecraft/cli/session.py

Hermes is a source-supported alternative to literal byte copying: rows are serialized, with content decoding while tool_calls/reasoning_details remain strings. It does not produce a summary or test the assertions in those rows. SRC-1 implementation, sdk/tracecraft/harness/hermes.py:

lines.append(json.dumps(d, default=str, ensure_ascii=False)) --- sdk/tracecraft/harness/hermes.py

Routes

EPI-RTE-git-files

Label: Optional git evidence

Part of: RT-RTE-step-control

Distinct material evidence-producing subroute. Caller opts into --changed-files-from-git; _git_changed_files calls Git, parses names and inserts changed_files into handoff. Caller proposes completion; code chooses the reference measurement; no human approval of each path or answer oracle for work correctness is established. Open coordination request. Failure yields [] rather than a veto. Guidance and persisted result are see EPI-CLM-git-truth. No automatic policy adaptation or rollback follows; a later completion can overwrite handoff.

• Immediate return: List of stripped nonempty diff output lines, or [] on nonzero exit/exception.
• Later read-back: Persisted changed_files can be read through direct store access to handoff.
• Delegated visibility: Bucket readers can see the optional field.
• Selection predicate: Explicit flag; fixed git diff --name-only HEAD in caller's working directory.
• Invalidation or expiry: Handoff overwrite; field is not refreshed by later commits or file changes.
• Activation or effect: Inserts machine-derived paths into handoff; downstream interpretation unobserved.
• Evidence limits: No execution supplied; excludes untracked paths and cannot certify completed work; empty is ambiguous.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 implementation, sdk/tracecraft/cli/steps.py:

["git", "diff", "--name-only", "HEAD"], --- sdk/tracecraft/cli/steps.py

if out.returncode != 0: return [] --- sdk/tracecraft/cli/steps.py

EPI-RTE-regex-scrub

Label: Regex scrub and counts

Part of: RT-RTE-session-mirror

A transformation subroute, not semantic acceptance. Source-native redact consumes bytes, applies denylist patterns sequentially and returns replacement bytes/counts to mirror. Code decides substring removal; no human/model judgment or secret/truth oracle. Caller selects activation through mirror options. Open ingestion. Counts and altered bytes persist; source bytes are not overwritten by this routine, and restoration would require original host material. RT-CLM-secret-redaction owns the claim.

• Immediate return: Replacement bytes and nonzero pattern counts.
• Later read-back: Mirror parts carry altered bytes; metadata stores cumulative counts.
• Delegated visibility: Bucket readers receive scrubbed parts and counts, not removed originals through this path.
• Selection predicate: Byte regex membership, unless --no-redact bypasses the call.
• Invalidation or expiry: No rule-driven re-scrub of already retained parts; later source/package revision changes future calls.
• Activation or effect: Replaces matched substrings before mirror upload; no behavior-shaping model consumer inspected.
• Evidence limits: No run supplied; arbitrary secrets and split-chunk matches not covered; counts do not license original truth.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 implementation, sdk/tracecraft/redact.py:

("aws_access_key", re.compile(rb"AKIA[0-9A-Z]{16}")), --- sdk/tracecraft/redact.py

out, n = pat.subn(f"[REDACTED:{name}]".encode(), out) if n: counts[name] = n --- sdk/tracecraft/redact.py

Concrete source deduction, not execution: input bytes key=AKIA1234567890ABCDEF match the 16-character uppercase/digit suffix. The AWS rule changes them to key=[REDACTED:aws_access_key] and records aws_access_key count 1. Later patterns do not match this replacement. This establishes loss of the literal credential-like value, not that the input was a real credential. In an arbitrary sentence or structured record such replacement can change a proposition or its parseability, so semantic preservation/entailment is indeterminate; no ampliative conjecture is inferred just from that uncertainty.

EPI-RTE-read-verification

Label: Requested read verification

Distinct doctrine-only checking request, compared with RT-BAP-agent-prompt, which owns the broader instruction delivery path, and RT-RTE-memory-write/RT-RTE-mail-exchange, which own executable reads/writes. The Hermes host agent is instructed to compare write/read results. No automatic comparator, recorded criterion outcome or acceptance disposition is implemented by this Markdown. Human/model compliance is unresolved; an expected prior write could be a reference for correspondence, not an answer oracle for substantive truth. Open coordination use, not a bounded experiment.

• Immediate return: inapplicable — static skill instruction returns no checked result.
• Later read-back: Requested memory get, recipient inbox or step-status reads retained state.
• Delegated visibility: Send verification explicitly asks a recipient to read inbox.
• Selection predicate: Corresponding read command after the named action.
• Invalidation or expiry: inapplicable — no persisted check result or acceptance state is specified.
• Activation or effect: Advisory request only; actual comparison and changed reliance unobserved.
• Evidence limits: No host loading/compliance trace; read success alone does not certify message meaning.
• doctrine conclusion status: claimed
• operation conclusion status: uninspected

SRC-1 doctrine/design, examples/hermes-agent/SKILL.md:

After any action, verify with the corresponding read command: - After memory set → memory get - After send → ask the recipient to check inbox - After claim → step-status --- examples/hermes-agent/SKILL.md

EPI-RTE-bucket-evaluation

Label: Bucket benchmark evaluation

No counterpart benchmark route is declared in supplied runtime; RT-RTE-step-control and RT-RTE-mail-exchange are product routes whose primitives the evaluators exercise, not these test drivers. Two fixed evaluators remain separate in license: contention compares API winners to durable owner; messaging compares intended sends to distinct retained bodies. Script caller chooses conditions; software performs writes, counts and judges equality. Human initiation/interpretation unresolved. Reference outcomes are one consistent winner and sent distinct bodies, not truth oracles for payload text. Bounded experimental mode. Exceptions invalidate contention trials; results report rather than admit product content. Contention deletes its trial key after evaluation; deletion failure is swallowed, and the inspected function does not confirm deletion despite its comment. No feedback consumer modifies the SDK policy in these functions.

• Immediate return: Contention Trial fields including invariant_held/valid; messaging sent, delivered_keys, distinct_messages, lost and errors.
• Later read-back: Fresh backend client retrieves claim owner or messages within the trial; publication/retention of resulting reports uninspected here.
• Delegated visibility: Test objects use configured bucket access; distribution of results uninspected.
• Selection predicate: Unique trial claim key or configured recipient mailbox; selected concurrency/key scheme and optional stagger.
• Invalidation or expiry: Trial cleanup attempts claim deletion; no epistemic rollback or product content acceptance state.
• Activation or effect: Deterministic result computation; no actual trial or downstream policy revision observed.
• Evidence limits: Source only; service correctness, supplied benchmark result authenticity and causal comparisons not assessed. Key counts are not end-user inbox delivery or message truth.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 implementation, benchmarks/contention_bench.py:

t.invariant_held = ( t.valid and win_count == 1 and t.stored_owner == t.declared_winner ) --- benchmarks/contention_bench.py

SRC-1 implementation, benchmarks/messaging_bench.py:

"distinct_messages": len(bodies), "lost": sent - len(bodies), --- benchmarks/messaging_bench.py

The stagger control affords testing a simultaneous-contention hypothesis, but without inspected intervention results it supports no causal finding. The treatment changes launch timing, not independently isolated backend components. Sampled SDK test supplies a different fixed reference outcome, not a semantic oracle. SRC-1 implementation, sdk/tests/test_v025_features.py:

r = env.invoke(cli, ["memory", "get", "k"]) assert r.output.strip() == "v2" --- sdk/tests/test_v025_features.py

Claims

EPI-CLM-git-truth

Label: Git source of truth

Part of: RT-RTE-step-control

A design/warrant claim attached to that route's optional evidence producer, not the runtime's mechanism declaration. SRC-1 doctrine/design, sdk/tracecraft/cli/steps.py, claims local Git is authoritative for changed paths and self-reported lists are unreliable. No evidence for the frequency estimate is supplied by this implementation.

Git is the source of truth for what changed — we never let an agent type the file list by hand (self-reported change lists are wrong ~half the time and go stale on the next commit). --- sdk/tracecraft/cli/steps.py

EPI-CLM-history-loss

Label: Immutable no-loss history

Part of: RT-RTE-memory-write

Design claim about its retention subfunction, not another declaration of the writer. SRC-1 doctrine/design, sdk/tracecraft/cli/memory.py, claims every set appends an immutable snapshot and no version is ever lost. Implementation gives snapshot-before-live ordering, not object immutability, transactional success or universal backend durability. Direct wrapper mutation is the counterroute supplied in runtime's forcing cases.

Every set also appends an immutable snapshot under memory/_history/, so the current value is last-write-wins but no version is ever lost. See memory history. --- sdk/tracecraft/cli/memory.py

Normal ordering evidence is see RT-RTE-memory-write. No semantic repair of that supplied route is needed: it already limits immutable-history claims to a nontransactional, directly mutable implementation.