Find a retained run

E6-observer-admission

Claim scope: verifier-behavior-on-author-fixture

Review state: retained-measured-record

Open Report, Provenance and Contract.

Recorded procedure

Per-claim results

Roles

Limits

• One declared broker file write and one local consumer admission
• No live-agent or production-isolation measurement
• No independently operated witness or consumer
• Consumer-state rollback can erase replay protection
• No exactly-once downstream-effect claim

answerExposure: Implementation, demo and retained expected outcomes were available to the atlas runner; not an answer-blind replication

comparisonOrder: Retrospective retained-fixture comparison; not a result frozen before expected-answer comparison

jep-core07-2026-10-02

Claim scope: verifier-behavior-on-public-producer-fixtures

Review state: submitted-awaiting-producer-response

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Pinned manifest coverage only, not complete JEP Core coverage.
• Acceptance effects are synthetic local SQLite rows.
• No power-loss or distributed failover experiment.

answerExposure: Published expected outcomes were available before implementation; not answer-blind.

comparisonOrder: Actual per-case outputs retained before comparison in memory; raw output and comparison published together, not separate revisions.

remora-e7-2026-10-02

Claim scope: verifier-behavior-on-public-producer-fixtures

Review state: producer-review-open

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Native INDEPENDENT means the external operator and separate implementation satisfy REMORA package criteria; it establishes no independent effect custody.
• Global runtime_capability_surface_completeness remains NOT_ESTABLISHED.
• Producer review is open; it does not establish producer acceptance, final verification or CI adoption.

answerExposure: Published expected outcomes were available before implementation; not answer-blind. REMORA reference-verifier text was also exposed during lint diagnostics after implementation.

comparisonOrder: Actual per-case outputs retained before comparison in memory; raw output and comparison published together, not separate revisions.

langgraph-durable-2026-10-02

Claim scope: author-operated-native-framework-control

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Graph-worker hard exits/reopened SQLite checkpoints only; target process remains alive.
• No target-process restart, host power loss, issuer authentication, general exactly-once guarantee or independent custody.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison published together; not separate raw-result-first revisions.

pydantic-failure-2026-10-02

Claim scope: author-operated-native-framework-control

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Real Pydantic framework with scripted FunctionModel; no model-quality or real-provider inference measurement.
• Selected loopback ticket effects; no caller authentication, outside adoption or independent custody.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison published together; not separate raw-result-first revisions.

model-operational-2026-10-02

Claim scope: author-operated-actual-weight-microtask-result

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Original 274795286-byte ZIP retained locally and available through expiring GitHub workflow artifact, not committed to Git or represented by the capsule.
• Capsule retains exact selected JSON/text members and excludes model weights/binaries; it cannot replay the full source/binary integrity check or rerun inference alone.
• Fixed configuration order confounds cache/warmup; author microtasks are not a benchmark estimate.
• CPU is whole-process delta; RSS process-lifetime peak; no task memory isolation, monetary cost or independent effects.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison published together; not separate raw-result-first revisions.

authority-recovery-2026-10-02

Claim scope: author-operated-native-recovery-authority-control

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Host-selected deterministic integer recovery clock; signed target grant remains separate.
• Graph workers hard-exit/reopen; target process stays alive; no host power loss, distributed exactly-once or independent custody.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison emitted and published together; public Atlas retention before indexing does not establish a blind or raw-result-first comparison.

target-process-recovery-2026-10-02

Claim scope: author-operated-native-target-process-control

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Selected target-process hard exits/reopen and local concurrency only; process IDs are unsigned runner testimony.
• Native SQLite/key/clock/HTTP remain same-operator; no power loss, production containment, general exactly-once or outside adoption.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison emitted and published together; public Atlas retention before indexing does not establish a blind or raw-result-first comparison.

model-paired-cpu-2026-10-02

Claim scope: author-operated-actual-weight-paired-microtask-result

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Producer-selected compact original retains native source/library/call bytes; full preparation weights/source/dependency wheels remain privately archived and provider-expiring.
• Two separately declared preparation attempts, initial failure before inference and corrected preparation; cumulative transfer is not one 512MiB budget.
• Author tasks/answers exposed; blocked interleaving and cache reset do not make a representative benchmark or independent operation.
• Whole-process CPU and shared lifetime RSS are not per-task allocations; no live decision effects.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison emitted and published together; public Atlas retention before indexing does not establish a blind or raw-result-first comparison.

openai-agents-ticket-2026-10-02

Claim scope: author-operated-scripted-native-sdk-control

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Actual OpenAI Agents SDK Runner/function_tool/TraceProcessor with scripted Model; nine model calls do not measure remote inference or model quality.
• Producer implementation, PR synthetic execution, final reader and separately pinned protected target revisions remain distinct.
• Same-operator synthetic HTTP ticket effect/key/clock; no independent custody, target restart or outside recurring adoption.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison emitted and published together; public Atlas retention before indexing does not establish a blind or raw-result-first comparison.

execsurface-state-2026-10-02

Claim scope: author-operated-native-metadata-and-endpoint-state-join

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Voluntary/peer/log-import/software-only predicate with incomplete scope and unknown typed collection health.
• Native schema-v2 metadata does not carry written bytes or write count; nine-byte content comes from separate snapshots and declared workload.
• Same operator controls generated fixture, key, store, clock and captures; no independent custody, principal authentication or outside adoption.
• Selected Linux x86_64 ptrace metadata only; no mmap/io_uring attribution or absence-of-intermediate-write claim.
• Native learn/check producer PASS is a separate outside-interval same-byte calibration, not authority, collection health or task quality.
• Release source/binary association is pinned but Rust build reproducibility is not established.

answerExposure: Author-written deterministic overwrite, companion and expected outcomes exposed; not blind.

comparisonOrder: Native companion output and verifier comparison emitted together; public raw retention before indexing does not establish a blind or native raw-result-first comparison.

model-format-cpu-2026-10-02

Claim scope: author-operated-actual-weight-paired-format-and-correctness-result

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Format/schema correctness separate from typed-exact target correctness; each model/cap/decoder/family row retained independently.
• Protocol/compiler/grammars frozen before inference; selected integer-length, key-order and spacing restrictions limit generalization.
• Exposed author tasks and same-operator native execution; no representative benchmark, live protected decision effects or independent custody.
• Four separately declared preparation envelopes; cumulative transfer exceeds one512MiB envelope.
• Original compact native provider archive retains sources/grammars/outputs and resources; full preparation weights/source/wheels remain privately archived.
• CPU is whole-process delta and RSS shared lifetime peak; not per-model/task allocation.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison emitted and published together; public Atlas retention before indexing does not establish a blind or raw-result-first comparison.

google-adk-ticket-2026-10-02

Claim scope: author-operated-native-google-adk-scripted-reference-result

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Real Google ADK nativeRunner/plugin/retry paths with scriptedBaseLlm; no remote inference or model-quality evaluation.
• Typed MCPisError throughFunctionTool is not actual MCP transport.
• Task complete/error/incomplete separate from signednativeRevision; after-effect error and early close retain committed effects.
• Author-operated target/store/key/clock/capture and exposed expected outcomes; no independent custody/outside adoption.
• Original synthetic execution source, producer implementation, reviewed reader and protected target baseline separately pinned.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison emitted and published together; public Atlas retention before indexing does not establish a blind or raw-result-first comparison.

installed-format-policy-2026-10-02

Claim scope: author-operated-installed-native-reader-upgrade-and-publication-policy-result

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Actual framework-free installed0.0.2→0.0.3 upgrade and selectedhostpolicy run, not a new inference experiment.
• Verified scoped-report publication separate from held policy1/6 correctness, no dispatch or recovery authority.
• Full original2268member provider artifact retained including previous upgrade, wheels, immutable fixture packets and rawgate streams; no omissions.
• Same operator selects installation/reader/policy/expected outcomes; no registry release, outside recurring use or independent custody.
• Consumer source/reviewed commit and actualnative192 producer source/run remain separate.

answerExposure: Author-written tasks/controls and expected outcomes exposed; not blind.

comparisonOrder: Native output and comparison emitted and published together; public Atlas retention before indexing does not establish a blind or raw-result-first comparison.

model-boundary-cpu-2026-10-02

Claim scope: author-operated-finite-typed-policy-grounded-boundary-result

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• All24 model/cap/decoder/family rows and384 original attempt scores retained unchanged; pair correctness separate from schema validity.
• 48 authored case identities include46 unique literal inputs and two repeated positive controls; exposed targets, not blind or representative.
• No protected model decision effects, independent operation/effect custody or outside recurring use.
• Five separate preparation envelopes; cumulative2167006362 response-body bytes cannot be presented as one512MiB acquisition.
• Complete original preparation privately archived; public871-member original compact maps to901 full members with30 explicit omissions.
• Total process CPU includes initialization; returned-call CPU and shared lifetime RSS keep their original scopes.
• Post hoc vocabulary diagnosis and cap comparisons do not alter the frozen rubric or published scores.

answerExposure: Author-written cases, public targets and previous public failures informed selection;48 identities include46 literal inputs and two repeated positive-control inputs. Not blind.

comparisonOrder: Native output and comparison emitted and published together; public Atlas retention before indexing does not establish a blind or raw-result-first comparison.

atomic-delegation-2026-10-02

Claim scope: native-local-certificate-selection-persistence

Review state: author-operated-upstream-review-pending

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Normal fresh reader-process restart, not crash-mid-write consistency.
• Self-contained local grant selection does not establish external issuer authority or server permission enforcement.
• Local key, clock and filesystem remain under one operator; child PIDs are unsigned testimony.
• No remote revocation, independently operated effect custody, model task quality or recurring outside use is established.
• Full author-assembled ZIP packages byte-exact original retained-run files; it is not a CI provider archive.

answerExposure: Author-written native cases and expected outcomes exposed; not blind.

comparisonOrder: Original native observations and test assertions emitted together, not separate public raw-result-first revisions.

installed-boundary-policy-2026-10-02

Claim scope: author-operated-installed-reader-publication-and-quality-policy-result

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Author-operated installed0.0.3-to0.0.4 reader and host-selected publication/quality policy; no new inference.
• Public selected capsule omits native packet duplicates and unrelated earlier workflow results; hashes cannot recreate omitted bytes.
• The complete provider original is retained separately; native compact archive and preparation originals have their own retention records.
• Evidence verification and complete measurement are distinct from selected quality acceptance, action authority, recurring outside adoption and independent effect custody.

answerExposure: Author-written cases, targets and host quality thresholds exposed; not blind.

comparisonOrder: Original CI execution and comparison emitted together. Selected raw Atlas retention precedes indexing at dc54f604d90278fc41c730313ffa9310a18d7ebc; it does not establish blind or independent comparison.

model-vocabulary-cpu-2026-10-03

Claim scope: authored-native-action-vocabulary-and-installed-quality-hold

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Finite authored CPU policy task; schema validity does not imply policy correctness.
• Installed review performs no model calls and grants no action authority.
• Same operator source-bound readback; outside producer acceptance, recurring adoption and independent effect custody remain unestablished.
• Public retention preserves all native packet members; full weights, dependency wheels, source archive and CI installed environment remain outside this capsule. Their hashes cannot reconstruct omitted bytes.
• Provider retention expiry is recorded; separate same-operator retention is not independent custody.

answerExposure: Authored policy inputs and targets exposed; vocabulary intervention informed by previous 384-call results; not blind.

comparisonOrder: Native inference follows frozen protocol, source-bound controls and protected merge; Atlas raw retention precedes indexing at e1122b6c08915207921ca575d98115cea4790062. Same-operator replay does not establish independent custody.

model-tool-arguments-cpu-2026-10-03

Claim scope: authored-native-tool-arguments-with-strict-quality-hold

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Finite exposed authored tool-argument and abstention cases; no representative benchmark or blind evaluation.
• Native inference, evidence publication and semantic quality are separate; all eight strict quality rows hold.
• No tools or protected publication/admission/dispatch/recovery effects execute from model answers.
• Atlas retains the entire compact native provider artifact, not the complete preparation artifact containing model weights, wheels, archive and installed environment; hashes cannot reconstruct omitted bytes.
• Provider artifact expiry is disclosed; durable author-held Git bytes do not establish independent effect custody, outside acceptance or recurring adoption.
• Atlas validation reads data only and performs no new inference or installation of packet code.

answerExposure: Authored inputs, targets and thresholds exposed; prospective new study follows earlier 384/128-call findings; not blind.

comparisonOrder: Prospective protocol and reviewed source precede the protected Observer main execution; exact native provider bytes are retained additively in Atlas afterward. Neither raw retention nor replay establishes independent execution.

aps-jcs-2026-10-04

Claim scope: raw-admission-serializer-and-signed-receipt-comparison

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Two runtimes repeat the same 1,223 identities; populations remain separate.
• Parsed serialization, strict raw parsing and signed serialized receipt verification remain separate.
• The 42 parsed acceptances on raw-refusal rows include profile differences; they are not all duplicate losses.
• RFC 8785 and opt-in I-JSON numeric and Unicode policies differ; all 23 strict-parser policy rows remain retained.
• Selected licensed first-party source is retained separately from native output. Registry packages restore from integrity lockfiles; Node and Rust runtimes are not vendored.
• Author-operated conformance; separately proposed APS dependency and formal pre-run study confirmation remain separate.

answerExposure: Frozen public conformance expectations and authored receipt controls were exposed before implementation.

comparisonOrder: Native outputs and comparisons were emitted together; original archives and selected source were published before Atlas indexing.

google-adk-responses-2026-10-04

Claim scope: native-user-response-routing-callback-effect-and-separate-verify-comparison

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Three finite three-turn cases on each fully qualified execution; resumability disabled.
• Actual ADK framework with a fixed public BaseLlm script, no provider inference or model-quality measurement.
• The LRO write is a same-author host completion, separate from the native start-only tool body.
• Two distinct Observer/witness keys remain author-held PEER custody with author-operated storage and clock.
• Actual Verify event_absence/v1 covers the six plugin-delivered native turns; whole-second projections retain original fractional timestamps separately.
• First failed reader attempt and intermediate correction remain separate original archives.
• The older 16-row comparison remains pinned separately; the prospective eight-task implementation-owned study has not started.
• Restore the originally empty denied-workspace directory before installed-reader replay; no file bytes change.

answerExposure: Fixed public model responses, authored controls and expected outcomes were exposed before execution.

comparisonOrder: Native outputs and assertions were emitted together. Failed, corrected, PR and merged-main originals were retained before Atlas indexing.

go-jcs-2026-10-04

Claim scope: raw-input-admission-and-pinned-go-canonical-byte-comparison

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Finite conformance inputs and exposed expected answers; no task or workload experiment.
• Six executions on two Go runtimes repeat the same 1,263 case identities: two final-PR, two branch-push and two merged-main runs; each stays separate.
• Go accepts objects or arrays; number samples use declared array wrappers, admitted scalar roots refuse separately.
• RFC 8785 and I-JSON policies differ; bare Go outcomes remain separate from admission policy.
• The original native archives retain source and reports. Installed executable hashes are recorded, but executables and the full installed runtime are outside those provider archives.
• The upstream multi-gigabyte number benchmark was not selected.

answerExposure: Public fixture expectations and finite controls were available before implementation.

comparisonOrder: Native outputs and comparisons were emitted together; all four originals were byte-retained before Atlas indexing.

ag2-push-authority-2026-10-04

Claim scope: native-task-and-push-url-admission-dispatch-target-and-observed-effect-comparison

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• 18 finite actual native tasks per qualified run; fixed answers and no provider inference.
• Explicit same-author post-completion sender call; no automatic task callback dispatch.
• Controlled ASGI target and native plaintext gRPC loopback; no external callback target.
• PEER keys, target and custody; no outside operator or host adoption.
• Original native JSON and wire bytes retained; the reader checks JSON semantics without decoding protobuf wire.
• Older 16 rows unchanged; prospective eight-task study not started.

answerExposure: Public fixed responses and finite expected outcomes were exposed before execution.

comparisonOrder: Original failures and passing native archives were retained before Lab indexing.

remora-boundary-2026-10-04

Claim scope: verifier-behavior-on-public-producer-fixtures

Review state: producer-review-pending-submission

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Run records state SECOND_IMPLEMENTATION, EXTERNAL operator and NOT_INDEPENDENT; under REMORA's rule they support REPRODUCED, not EXTERNALLY_VERIFIED.
• Authorization signatures use a public test-only HMAC key created and checked by the reader; no REMORA-issued authorization is authenticated.
• The effect hash rule is unsupported until the contract defines its preimage; no frozen case uses it.
• Fixture premises such as revocation visibility at presentation and the observed field values are taken as supplied.
• No production safety, authority, independent effect custody or lifecycle change follows from these records.

answerExposure: Expected outcomes are inside the frozen fixtures and the reader compares against them; not answer-blind. Whether reference-verifier text was read while the reader was written is not recorded.

comparisonOrder: Native outputs and comparisons were emitted together in one run.

aeoess-receipt-signature-2026-10-05

Claim scope: source-reported-outside-verifier-and-primitive-checks-on-public-fixtures

Review state: upstream-merged-retained-run

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Operator identity, human review and container custody are declared by the original source, not independently authenticated.
• Four context-bearing identifier checks remain outside the runner primitive coverage.
• Author harness interpretation and grading remain a separate layer.
• Repeated runs use the same 25 fixtures; no agent, production effect or independent custody measurement.
• Merged run record does not establish full-family admission or maintained host CI adoption.

answerExposure: Published expected outcomes and the author harness were available before the run; not answer-blind.

comparisonOrder: The original harness emitted verifier answers and comparison results together. No separate raw-result revision preceded comparison.

w3c-report-replay-2026-10-06

Claim scope: historical-report-accounting-and-separate-diagnostic-control-dependencies

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Synthetic conformance artifacts; the customer refund story is hypothetical.
• Retained-output accounting executes no captured reader or mutant code and performs no new model inference.
• Historical 272-case report and separate 232-fixture diagnostic population are not combined.
• Report validity, evidence conclusion and actual action completion are distinct.
• Reader binding check is a historical proposed rule; its draft normative status is not promoted.
• No independent operation, effect custody, production workload, customer payment or outside maintained job is established.

answerExposure: Published expected answers and author-selected predicates were exposed; not answer-blind.

comparisonOrder: Historical harness outcomes and comparison grades emitted together. Diagnostic copies remove one predicate each; outputs are retained separately from the historical report.

aps-corrected20-2026-10-10

Claim scope: independently-written-and-operated-reader-on-producer-authored-finite-synthetic-corpus

Review state: author-retained-measured-record

Open Report, Provenance and Contract.

Per-claim results

Roles

Limits

• Finite printable-ASCII corpus reader; not general JCS conformance.
• Signature verification uses cryptography/OpenSSL; no independent cryptographic-backend claim.
• Diagnostic agreement does not force the five normatively undetermined SDK outcomes.
• Resolver answers are supplied synthetic data, without historical validity windows or authenticated lookup.
• No APS package qualification, principal binding, live DID lookup, host adoption, federation enrollment or commercial agreement.
• Twenty public synthetic cases, not a representative deployment population.
• Three declared full-outcome differences remain visible; PASS means the bounded diagnostic comparison, not uniform verdict agreement.
• The five upstream draft annotations are separate from the two indeterminate source outcomes.
• No new package release, maintainer acceptance of this reader, host adoption, federation enrollment or commercial agreement.
• Both selected SDK invocations stopped at IPC creation; their corpus/ambiguity checks remain not exercised.
• Original dependency audit remains adverse; the finite reader result is a separate evidence layer.

answerExposure: The corrected public fixture, all upstream expected outcomes and README annotations were available before the reader and run.

comparisonOrder: The independently written reader emitted raw checks, its outcomes and comparisons in one report; not answer-blind or separately graded.

trace-m1-consumer-2026-10-11

Claim scope: signature-report-binding-and-native-claim-ceiling-on-public-synthetic-inputs

Review state: retained-independent-consumer-result

Open Report, Provenance and Contract.

Recorded procedure

Per-claim results

Roles

Limits

• Bounded ASCII JSON, explicit report digest, P1Y retention and frozen synthetic inputs
• Native reason codes are preserved; independent refusal codes are not equated
• Signature and binding checks do not establish truth, workload identity, authority or effects
• Unsigned external-evidence mapping is not a TRACE Trust Record or conformance level
• No REMORA or AgenTrust host adoption, deployment or commercial agreement is established

answerExposure: Author read producer source and expected cases before implementation; not an answer-blind study

comparisonOrder: Synthetic known-answer replay; retained native output compared to published fixture expectations