Offline/Online Run Manifest
Offline/Online Run Manifest
이 manifest는 ADR 0079의 v0-b milestone을 위한 실행 환경 provenance다. 목적은 offline과 online 평가(evaluation)를 같은 schema로 기록해, metric aggregate가 어떤 환경과 payload boundary에서 생성됐는지 review할 수 있게 하는 것이다.
이 문서는 성능(performance) 주장을 만들지 않는다. Private real-eval 원문,
질문, 답변, 근거(evidence), doc_id, chunk_id, filename, exact local path는
manifest에 쓰지 않는다.
Command
python3 scripts/agent_loop.py eval-run-manifest \
--mode offline \
--payload-class none \
--egress-mode none \
--provider local \
--model local-judge-v1 \
--judge-backend local-llm
python3 scripts/agent_loop.py eval-run-manifest \
--mode online \
--payload-class private-raw \
--egress-mode private-raw \
--provider openai \
--model gpt-example \
--judge-backend external-judge
Default output:
reports/agent_loop/offline_online_run_manifest.json
eval/run_eval.py also records the same environment, model, payload, and
privacy sections inside run_manifest. A config may provide an optional
top-level block:
run_environment:
mode: offline
provider: local
model: local-judge-v1
judge_backend: local-llm
hardware: local-gpu
payload_class: none
private_data_egress: none
The same values can be overridden for ad hoc runs with:
BIDMATE_EVAL_ENVIRONMENT
BIDMATE_EVAL_PROVIDER
BIDMATE_EVAL_MODEL
BIDMATE_EVAL_JUDGE_BACKEND
BIDMATE_EVAL_HARDWARE
BIDMATE_EVAL_PAYLOAD_CLASS
BIDMATE_EVAL_PRIVATE_DATA_EGRESS
Schema
Top-level fields:
| field | purpose |
|---|---|
schema_version |
Manifest schema version. Current value: 1. |
generated_at |
UTC generation timestamp. |
git_head |
Current git commit. |
branch |
Current git branch. |
environment |
Offline/online execution mode and network assumptions. |
model |
Provider/model/judge backend provenance. |
payload |
Payload class and private-data egress mode. |
provenance |
Surface, case family, config digest, and sanitized source command. |
cost_latency |
Optional online cost/latency scalars. |
privacy |
Commit-safety booleans. |
environment fields:
| field | offline | online |
|---|---|---|
mode |
offline |
online |
network |
closed |
non-closed |
external_api_allowed |
false |
true |
external_api_used |
derived from provider | derived from provider |
hardware |
sanitized scalar | sanitized scalar |
payload.private_data_egress values:
| value | meaning |
|---|---|
none |
No private data leaves the execution environment. Required for offline. |
metadata-only |
Only coarse metadata/provenance leaves the environment. |
public-only |
Only public fixture or synthetic public payload leaves the environment. |
private-raw |
Private raw payload egress is intentionally allowed and recorded. |
Privacy Contract
- Offline manifests must set
private_data_egresstonone. - Online manifests must record provider and model.
- Config paths are not serialized. Only
config_sha256is recorded. - Source commands are sanitized before serialization.
- Manifest generation fails if exact local paths or raw private fields are present.
Claim Boundary
This manifest is provenance evidence only. It supports later v0-c report review,
but it is not a private real-eval result, benchmark lift, regression fix, or RFP
quality claim. New private-eval task, PR, claim, and handoff evidence must use
real100_v2 aggregate-only results with matching provenance; legacy real100 /
v1 / 221-case / kordoc evidence remains archive-only unless the maintainer
explicitly re-enables another private-eval surface.