Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Job has been terminated due to a temporary spike in resource usage and may be restarted later.
Error code:   JobManagerCrashedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

id
string
goal
string
plan
string
outcome
string
meta
unknown
reward
unknown
steps
list
pbc-r01-field4-userid-reuse-c71a
catalog-svc CatalogItem reused proto field 4: string user_id became int64 account_id. Old gRPC clients fail with wire decode errors while protojson contract tests stay green. Restore wire compat without dropping account_id.
Prove field-4 LEN vs VARINT mismatch, reject keep-int64 and json_name-only dead-ends, restore user_id=4, add account_id=18, dual-read shim, buf breaking 0.
Restored string user_id=4 and added int64 account_id=18. json_name-only dead-end left wiretype 2 vs varint. Normalize maps usr_ada->99 so old gRPC frames work. protojson IgnoreUnknown was the false-green (extra user_id/debug dropped). buf breaking 0, go 8/8, contract 5/5. Residual: lookupAccount is a test map, not the ...
{ "factory": "proto-breaking-change-factory", "round": 1, "generator": "grok-4.6" }
{ "success": true, "tests_passed": 13, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: locate catalog.proto, buf.yaml, generated Go, and contract tests before editing field numbers.", "observation": "proto/catalog.proto proto/catalog.proto.bak buf.yaml buf.lock src/catalog/item.go src/catalog/compat.go src/gen/catalog.pb.go clients/go/ clients/py...
pbc-r01-pagetoken-wire-sunset-9f2b
fulfillment-api ListShipments changed field 2 from int32 page to string page_token and remapped empty-page NOT_FOUND to OUT_OF_RANGE. JSON gateway tests pass; generated Python stubs still send page=int. Do not break the Go server.
Capture LEN vs VARINT on field 2, restore page=2 plus page_token=7, add a gateway translator, leave Python stubs on a sunset ticket if codegen is frozen.
Restored int32 page=2 and added string page_token=7 so frozen python stubs encode field 2 as varint. json_name-only dead-end broke new token clients. Gateway copies ?page= and ?page_token=. JSON extra-field test remains false-green (additionalProperties). Empty-page error stays OUT_OF_RANGE=11; python still wants NOT_F...
{ "factory": "proto-breaking-change-factory", "round": 1, "generator": "grok-4.6" }
{ "success": false, "tests_passed": 9, "xfailed": 2, "cost_steps": 24, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: list fulfillment proto, gateway, python stubs, and error-code tests before touching field 2.", "observation": "proto/fulfillment.proto src/fulfillment/list.go src/fulfillment/errors.go src/gateway/query.go src/gen/fulfillment.pb.go src/gen/fulfillment.pb.gw.go cl...
pbc-r02-enum2-draft-sent-a8e1
billing-ledger Invoice.status reused proto enum number 2: DRAFT became SENT. Old gRPC clients still encode 2 as DRAFT; new handlers treat 2 as SENT and skip the review queue. protojson name tests stay green. Restore wire numbers without dropping SENT.
Prove enum value 2 rename, reject allow_alias and json-name-only dead-ends, restore DRAFT=2, add SENT=5, dual-read shim, buf breaking 0.
Restored InvoiceStatus DRAFT=2 and added SENT=5. allow_alias dead-end left number 2 as a single value (Go still one identifier). Old gRPC frames with status=2 enqueue review again; new clients use 5 to skip. protojson IgnoreUnknown plus name-only asserts were the false-green. buf breaking 0, go 9/9, contract 4/4. Resid...
{ "factory": "proto-breaking-change-factory", "round": 2, "generator": "grok-4.6" }
{ "success": true, "tests_passed": 13, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: locate invoice.proto, buf.yaml, generated Go, and status contract tests before editing enum numbers.", "observation": "proto/invoice.proto proto/invoice.proto.bak buf.yaml buf.lock src/billing/status.go src/billing/queue.go src/gen/invoice.pb.go clients/go/ cli...
pbc-r02-packed3-sku-qty-b3c7
warehouse-bin Bin.sku_qty dropped [packed=false] on repeated int32 field 3. Old mobile/C# clients still send unpacked repeated tags; new Go decodes field 3 as packed LEN. JSON list tests stay green. Restore unpacked wire without dropping packed v2. Rust bindings are frozen.
Prove packed vs unpacked on field 3, reject json_name-only, restore packed=false on 3, add packed sku_qty_v2=9, leave rust on WH-812 if codegen is frozen.
Restored [packed=false] on Bin.sku_qty=3 and added packed sku_qty_v2=9 so C#/mobile unpacked frames decode. json_name-only dead-end left wiretype 0 vs packed LEN. Go MergeQty prefers field 9. JSON extra-slot test remains false-green (additionalProperties). Rust prost frozen packed=true on tag 3 — WH-812 open, default c...
{ "factory": "proto-breaking-change-factory", "round": 2, "generator": "grok-4.6" }
{ "success": false, "tests_passed": 8, "ignored": 1, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: list warehouse proto, store, rust bindings, and packed golden tests before touching field 3.", "observation": "proto/warehouse.proto src/warehouse/bin.go src/warehouse/decode.go src/gen/warehouse.pb.go clients/rust/src/warehouse.rs clients/csharp/Bin.cs tests/con...
pbc-r03-oneof2-card-wallet-c4d2
payments-api Charge.instrument oneof reused field 2: CardAuth card became WalletAuth wallet. Old gRPC clients still LEN-encode field 2 as CardAuth; new handlers read it as WalletAuth and drop PAN last4. protojson oneof-name tests stay green. Restore card=2 without dropping wallet.
Prove oneof member number reuse, reject json_name-only, restore card=2, add wallet=8, dual-read shim, buf breaking 0.
Restored CardAuth card=2 and added WalletAuth wallet=8. json_name-only dead-end left last4 parsed as wallet.token. NormalizeInstrument accepts card-only so old gRPC frames work. protojson IgnoreUnknown was the false-green (extra card/debug dropped). buf breaking 0, go 8/8, contract 4/4. Residual: JSON last-key-wins if ...
{ "factory": "proto-breaking-change-factory", "round": 3, "generator": "grok-4.6" }
{ "success": true, "tests_passed": 12, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: locate charge.proto, buf.yaml, generated Go, and oneof contract tests before editing field 2.", "observation": "proto/charge.proto proto/charge.proto.bak buf.yaml buf.lock src/payments/charge.go src/payments/compat.go src/gen/charge.pb.go clients/go/ clients/py...
pbc-r03-reserved12-resurrect-e91a
notify-api Alert resurrected reserved field 12 as string debug_blob. Old clients still treat 12 as the retired int32 retry_after. JSON gateway tests pass; Java stubs frozen on reserved. Do not break the Go server. Leave Java on NT-220 if codegen is frozen.
Capture reserved-range resurrection, restore reserved 12 plus a new debug field, add a dual-read for leftover retry_after frames, xfail Java if freeze holds.
Restored reserved 12,13 and added debug_blob=16 so old retry_after varints are dropped instead of exploding as string UTF-8. json_name-only dead-end left wiretype 0 vs LEN and broke JSON keys. Emit no longer requires debug_blob. JSON extra retry_after remains false-green (additionalProperties). Java stubs frozen on int...
{ "factory": "proto-breaking-change-factory", "round": 3, "generator": "grok-4.6" }
{ "success": false, "tests_passed": 9, "skipped": 1, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: list notify proto, gateway, java stubs, and reserved-field tests before touching field 12.", "observation": "proto/alert.proto src/notify/alert.go src/notify/compat.go src/gateway/query.go src/gen/alert.pb.go clients/java/AlertOuterClass.java tests/contract/test_...
pbc-r04-singular5-tags-rep-d81c
search-index Document.tags changed field 5 from string to repeated string. Old gRPC clients still LEN-encode a single comma-joined tag string; new handlers range over tags and drop the legacy blob. protojson list tests stay green. Restore singular tags=5 without dropping repeated tag_list.
Prove cardinality change on field 5, reject json_name-only, restore string tags=5, add repeated tag_list=11, split shim, buf breaking 0.
Restored string tags=5 and added repeated tag_list=11. json_name-only dead-end left the comma-joined blob as one repeated element. MergeTags splits tags=5 and prefers tag_list. protojson list-only asserts were the false-green. buf breaking 0, go 8/8, contract 4/4. Residual: extra keywords JSON still ignored via IgnoreU...
{ "factory": "proto-breaking-change-factory", "round": 4, "generator": "grok-4.6" }
{ "success": true, "tests_passed": 12, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: locate document.proto, buf.yaml, generated Go, and tag contract tests before editing field 5.", "observation": "proto/document.proto proto/document.proto.bak buf.yaml buf.lock src/search/index.go src/search/compat.go src/gen/document.pb.go clients/go/ clients/p...
pbc-r04-jsonname-id-collision-f02e
devices-api Device.legacy_uuid field 7 set json_name=deviceId, colliding with Device.device_id field 1 default JSON name. Gateway protojson last-key-wins drops device_id; OpenAPI tests stay green. Swift stubs frozen on the collision. Restore distinct JSON names without dropping legacy_uuid.
Prove json_name collision, reject deleting field 1, restore deviceId vs legacyUuid, leave Swift on DV-118 if codegen is frozen.
Restored distinct JSON names (device_id -> deviceId, legacy_uuid -> legacyUuid) so protojson no longer last-key-wins into field 7. Deleting field 1 was a dead-end (old gRPC ids dropped). Upsert copies legacy_uuid when device_id is empty. OpenAPI extra hwid remains false-green. Swift stubs frozen with dual deviceId — DV...
{ "factory": "proto-breaking-change-factory", "round": 4, "generator": "grok-4.6" }
{ "success": false, "tests_passed": 10, "xfailed": 2, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: list device proto, gateway, swift stubs, and json_name tests before touching field 7.", "observation": "proto/device.proto src/devices/upsert.go src/gateway/json.go src/gen/device.pb.go clients/swift/Device.pb.swift tests/contract/test_device_json.py tests/wire/t...
pbc-r05-required2-email-opt-a19f
identity-svc Profile.email dropped proto2 required on field 2 during a proto3 syntax migrate. Old proto2 clients still require the field on decode; new proto3 writers omit empty email. JSON contract tests stay green. Restore explicit presence without dropping proto3 syntax.
Prove required->implicit presence loss, reject json_name-only, add proto3 optional email=2, dual-read empty vs unset, buf breaking 0.
Migrated Profile to proto3 optional email=2 so HasEmail() is back. json_name-only dead-end left omitted field 2. RequireEmail checks presence; Encode emits empty string so proto2 required readers still decode. JSON omit==empty was the false-green. WIRE buf 0, go 9/9, contract 4/4, Java 2/2. Residual: FILE-mode buf stil...
{ "factory": "proto-breaking-change-factory", "round": 5, "generator": "grok-4.6" }
{ "success": true, "tests_passed": 15, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: locate profile.proto, buf.yaml, generated Go, and email presence tests before editing field 2.", "observation": "proto/profile.proto proto/profile.proto.bak buf.yaml src/identity/upsert.go src/identity/compat.go src/gen/profile.pb.go clients/java/Profile.java te...
pbc-r05-map4-attrs-entry-c60b
flags-api Flag.attrs changed field 4 from map<string,string> to repeated Attr entries. Old gRPC clients still encode proto map entries (key=1,value=2). New handlers only read Attr{name,value} and drop the map. JSON object tests stay green. Kotlin stubs frozen. Restore the map number without dropping repeated attrs_v2.
Prove map vs repeated message on field 4, reject json_name-only, restore map attrs=4, add repeated Attr attrs_v2=10, leave Kotlin on FG-77 if codegen is frozen.
Restored map<string,string> attrs=4 and added repeated Attr attrs_v2=10. json_name-only dead-end left map entries decoded as Attr then dropped on empty src. MergeAttrs unions both encodings. JSON object-for-list stub was the false-green. Kotlin stubs frozen map-only — FG-77 open, 1 passed / 1 ignored. Do not reuse fiel...
{ "factory": "proto-breaking-change-factory", "round": 5, "generator": "grok-4.6" }
{ "success": false, "tests_passed": 9, "ignored": 1, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: list flags proto, store, kotlin stubs, and map golden tests before touching field 4.", "observation": "proto/flag.proto src/flags/put.go src/flags/compat.go src/gen/flag.pb.go clients/kotlin/Flag.kt tests/contract/test_attrs_json.py tests/wire/test_map4_golden.go...
pbc-r06-ts6-unixms-stamp-b7e4
audit-log Event.created_at reused field 6: int64 unix_ms became google.protobuf.Timestamp. Old gRPC clients still VARINT-encode millis; new handlers fail to unmarshal a message on field 6. protojson RFC3339 tests stay green. Restore unix_ms=6 without dropping Timestamp.
Prove VARINT vs LEN message on field 6, reject json_name-only, restore int64 created_ms=6, add google.protobuf.Timestamp created_at=14, dual-read shim, buf breaking 0.
Restored int64 created_ms=6 and added Timestamp created_at=14. json_name-only dead-end left wiretype 0 vs LEN. EventTime prefers field 14 then millis. protojson RFC3339 plus IgnoreUnknown created_ms was the false-green. buf breaking 0, go 8/8, contract 4/4. Residual: zero created_ms still indistinguishable from unset.
{ "factory": "proto-breaking-change-factory", "round": 6, "generator": "grok-4.6" }
{ "success": true, "tests_passed": 12, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: locate event.proto, buf.yaml, generated Go, and timestamp contract tests before editing field 6.", "observation": "proto/event.proto proto/event.proto.bak buf.yaml src/audit/write.go src/audit/compat.go src/gen/event.pb.go clients/python/event_pb2.py tests/contr...
pbc-r06-ext100-range-reuse-d3aa
plugin-host HostOptions resurrected extension range 100-199 as regular fields (int32 timeout_ms=100). Old plugins still write extension 100 as a length-delimited PluginTimeout. JSON gateway tests pass; C++ plugin SDK frozen. Restore the range without dropping the new first-class field.
Prove extension-range resurrection, reject json_name-only, restore extensions 100 to 199, add timeout_ms=21, leave C++ on PH-55 if codegen is frozen.
Restored extensions 100-199 and added timeout_ms=21. json_name-only dead-end left LEN vs VARINT on field 100. ResolveTimeout reads field 21 or PluginTimeout extension 100. JSON extra pluginTimeout remains false-green. C++ stubs frozen on scalar field 100 — PH-55 open, 1 disabled. Do not resurrect 100 as a scalar to gre...
{ "factory": "proto-breaking-change-factory", "round": 6, "generator": "grok-4.6" }
{ "success": false, "tests_passed": 8, "disabled": 1, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: list host proto, plugin SDK, and extension golden tests before touching field 100.", "observation": "proto/host.proto proto/plugin_timeout.proto src/host/options.go src/gen/host.pb.go clients/cpp/host_options.cc tests/contract/test_host_json.py tests/wire/test_ex...
pbc-r07-any3-payload-typed-e5c1
jobs-api Job.payload reused field 3: google.protobuf.Any became message TypedPayload. Old gRPC clients still wrap type_url+value; new handlers expect kind/body and drop the Any envelope. protojson @type tests stay green. Restore Any=3 without dropping TypedPayload.
Prove Any vs typed message on field 3, reject json_name-only, restore google.protobuf.Any payload=3, add TypedPayload typed=12, unpack shim, buf breaking 0.
Restored google.protobuf.Any payload=3 and added TypedPayload typed=12. json_name @type dead-end left type_url in kind. NormalizePayload unpacks Any into typed. protojson @type stub was the false-green. buf breaking 0, go 9/9, contract 4/4. Residual: kind is the type_url suffix only, not a validated jobs.v1 registry.
{ "factory": "proto-breaking-change-factory", "round": 7, "generator": "grok-4.6" }
{ "success": true, "tests_passed": 13, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: locate job.proto, buf.yaml, generated Go, and Any contract tests before editing field 3.", "observation": "proto/job.proto proto/job.proto.bak buf.yaml src/jobs/submit.go src/jobs/compat.go src/gen/job.pb.go clients/python/job_pb2.py tests/contract/test_any_json...
pbc-r07-sfixed8-crc-int32-a2d8
storage-api Block.crc reused field 8: sfixed32 became int32. Old clients still encode field 8 as 32-bit little-endian (wire type 5); new handlers expect VARINT. JSON number tests stay green. Rust bindings frozen. Restore sfixed32 without dropping zigzag-free int32 on a new number.
Prove sfixed32 vs int32 on field 8, reject json_name-only, restore sfixed32 crc=8, add int32 crc32=15, leave Rust on ST-90 if codegen is frozen.
Restored sfixed32 crc=8 and added int32 crc32=15 so old I32 frames decode. json_name-only dead-end left wiretype 5 vs varint. JSON number tests remain false-green (no wire type). Rust prost frozen sfixed32-only — ST-90 open, 1 passed / 1 ignored. Do not flip field 8 to int32 to green rust varint tests.
{ "factory": "proto-breaking-change-factory", "round": 7, "generator": "grok-4.6" }
{ "success": false, "tests_passed": 9, "ignored": 1, "cost_steps": 22, "buf_breaking": 0 }
[ { "n": 1, "decision_basis": "Plan: list block proto, rust bindings, and sfixed golden tests before touching field 8.", "observation": "proto/block.proto src/storage/put.go src/gen/block.pb.go clients/rust/src/block.rs tests/contract/test_crc_json.py tests/wire/test_sfixed8_golden.go tickets/ST-90....
End of preview.

Proto Breaking Change Trajectories

Rights & intended use: legacy public research corpus / portfolio artifact. Hosted frontier-model outputs are research-only inputs under project policy (synthetic-factory#161): intended_use: research_only, project_training_policy: blocked. Not training data for any model-weight update. Machine-readable record: rights.json.

Release status: The raw, uncurated payload is now published under data/raw/. It is available for inspection and reproducibility, but it is not training-ready.

Visibility: public raw-data repository.

Rights and intended use

This is a legacy public research corpus and portfolio artifact, not training data.

Under the Synthetic Factory project-policy decision recorded in rmems/synthetic-factory#161, outputs of hosted frontier models (here: Grok 4.6 (xAI)) are research-only inputs:

  • intended_use: research_only
  • project_training_policy: blocked

These records must not enter SFT, DPO, RL, distillation, continued pretraining, or any other model-weight update in this project. training_ready: false is a data-quality statement; even a future curated release would not make this corpus eligible for weight updates under project policy. That policy value does not change merely because a provider document later permits training — changing it would require a separate recorded project decision.

Research retention, evaluation, and redistribution rights are tracked separately per provider/channel/date and fail closed while unresolved (research_retention_status, research_evaluation_status, redistribution_status: unresolved). The machine-readable record is rights.json. Full contributor/role breakdown is preserved in ATTRIBUTION.md.

Historical release provenance

This dataset was originally published under Apache-2.0 during the 2026-08 Synthetic Data Factory runs. That historical grant is documented as released — this update reframes intended use going forward; it does not rewrite, backdate, or delete the release history or the payload.

Protobuf leftover-compat breaking-change episodes.

Intended model target

This is a general agentic research dataset in the Grok 4.6 agentic factory. It is independent of the Fable 5 collection and is not labeled as Spikenaut training data.

Generation attribution

The underlying synthetic data is generated by Grok 4.6 through the Synthetic Data Factory agentic lane. Factory slug: proto-breaking-change-factory. Source tree: outputs/raw/2026-08-19-agentic/proto-breaking-change-factory/.

Published raw payload

The release contains 3414 raw records across data/raw/batch-r01.jsonl through data/raw/batch-r1707.jsonl (~25345 KB), snapshotted from outputs/raw/2026-08-19-agentic/proto-breaking-change-factory/. Supporting notes are under data/metadata/NOTES-*.md. The factory source remains the write destination; this Hub copy is a public evidence snapshot, not the curated training export. Public visibility is not a training-readiness claim.

Kind comes from the payload, not the slug. All 3414 records are protobuf breaking-change episodes from proto-breaking-change-factory. No leftover-mill mix.

Schema and Dataset Viewer

Streaming preview can load early rows. Full Dataset Viewer / parquet indexing fails because reward optionally includes buf_breaking, xfailed, skipped, ignored, disabled, or pending. Step reflection is also optional. The card YAML above declares a union/json schema so the viewer can index without rewriting historical raw files. See synthetic-factory#49.

Every step includes a public decision_basis field (observable rationale, not a hidden chain-of-thought).

Planned curated release

Curated training publication remains blocked until a later audit and export pass. Do not treat this repository as a training corpus.

Execution and validation environment

All local curation, validation, dataset processing, packaging, and release engineering for this dataset were performed exclusively on the Ship of Theseus AI/HPC research workstation.

Ship of Theseus was the sole local execution environment for this project. It was not the hosted model-generation channel: the synthetic records were generated through the provider surface identified in rights.json and the dataset provenance metadata.

Links

License

This public raw release is licensed under the Apache License 2.0. That license grants reuse permissions; it does not make the records training-ready or factual real-world measurements.

Downloads last month
561

Collection including rmems/proto-breaking-change-trajectories