toki/iop

toki 02753c5225 chore(epic): thin-run 준비 결과를 검증한다

2026-08-14 07:02:30 +09:00

23 KiB

Raw Blame History

Plan - Run and Score the Thin-Agent Matrix Once

For the Implementing Agent

Execute the pre-attempt gate and the nine rows exactly as specified. A started producer command consumes that row even on timeout, failure, or malformed output; never retry, resume, recover, or replace it. Fill implementation-owned sections in CODE_REVIEW-cloud-G08.md, then leave the active pair for official review.

Background

The checkpoint pair established the intended atomic benchmark and the first refinement added explicit evidence paths. That refinement remained non-executable: it could form /v1/v1/models, unset the token before producer work, and ran Claude outside the row workspace. This replacement preserves the original nine-route, single-attempt, blind-score intent while closing those command and evidence defects.

Archive Evidence Snapshot

Pre-refine intent: checkpoint e09aa66c3cdb829366463c10f8bc5f5801e3136e, with one atomic pair covering all four Milestone Task ids.
Replaced unstarted refinement: plan_local_G08_1.log, code_review_cloud_G08_1.log; no verdict or implementation evidence.
Earlier unstarted pair: plan_local_G08_0.log, code_review_cloud_G08_0.log; no verdict.
Preserved invariants: fixed prompt and nine tuples, empty row workspaces, one producer invocation per tuple, no benchmark runner/state machine, opaque single-pass scoring, failures and missing usage are not zero.

Analysis

Files Read

AGENTS.md
agent-ops/rules/project/rules.md
agent-ops/rules/common/rules-roadmap.md
agent-ops/rules/common/rules-agent-spec.md
agent-ops/rules/project/domain/testing/rules.md
agent-ops/skills/common/router.md
agent-ops/skills/common/plan/SKILL.md
agent-ops/skills/common/refine-plans/SKILL.md
agent-ops/skills/common/finalize-task-routing/SKILL.md
agent-test/local/rules.md
agent-test/dev/rules.md
agent-test/dev/iop-thin-agent-model-comparison.md
agent-test/dev/iop-benchmark-route-minimal-html-smoke.md
agent-roadmap/current.md
agent-roadmap/phase/knowledge-tool-optimization-extension/milestones/thin-agent-model-comparison-benchmark.md
agent-spec/index.md
agent-contract/index.md
docs/dev-opencode-settings-guide.md
scripts/e2e-hot-path-agents.sh
scripts/e2e-single-request-claude.sh
the four task-local logs named above

SDD Criteria

SDD is not required. The Milestone records a test-only observation of existing caller/product paths with no API, state-machine, retry, or schema change. This pair contributes exactly single-attempt-matrix, minimal-result-table, single-pass-scorecard, and bounded-conclusion.

Verification Context

No handoff was supplied. Repository-native dev rules select toki@toki-labs.com, /Users/toki/agent-work/iop-dev, Edge port 18083, the existing SOPS principal token, and the managed CA. Before creating run_root, normalize the decrypted URL once to api_root by removing one trailing slash and one trailing /v1; use ${api_root}/v1/models, ANTHROPIC_BASE_URL=$api_root, and OpenCode baseURL=${api_root}/v1. Keep the shell-local token through row 09 and unset it immediately afterward.

External preflight must prove the smoke-qualified clean dev runtime identity, required caller versions/flags, ports, credential inputs, timeout tool, and authenticated five-model catalog. It does not deploy or mutate tracked runtime config. Local /config/.local/bin/chromium renders each scorable source once at each fixed viewport. Secrets and raw caller streams remain ignored evidence.

Test Coverage Gaps

Live availability has no unit substitute; the authenticated catalog gate and immutable row ledgers are the evidence.
Caller usage schemas differ; use an explicit caller field from the sole stream or 미제공.
Visual judgment is manual by design; opaque input, fixed renders, locked anchors, arithmetic, and direct evidence bound it.

Symbol References

None; no product symbol changes.

Split Judgment

Keep one atomic pair. The result table, post-attempt opaque bijection, scorecard, and conclusion must bind to the same immutable nine-attempt set; independent children could expose route identity early or weaken the no-retry boundary. The explicit row protocol keeps large_indivisible_context=false.

Scope Rationale

Writable tracked output is only the comparison document and active review evidence. Writable ignored evidence is the enumerated agent-test/runs/bench-lite-01 files below. Product source/config, caller installation or user config, roadmap/spec/contract, scripts, runner/manifest/state-store, and prior smoke evidence are excluded.

Final Routing

evaluation_mode: first-pass for this semantic replacement
finalizer: finalize-task-policy.sh, mode pair
closures: scope/context/verification/evidence/ownership/decision are true for build and review
build scores: scope 1, state 2, blast 1, evidence 2, verification 2 = G08; local-fit; PLAN-local-G08.md
review scores: scope 1, state 2, blast 1, evidence 2, verification 2 = G08; official-review; CODE_REVIEW-cloud-G08.md
positive loop risk: variant_product (1); large_indivisible_context=false; rework 0; integrity failure false

Implementation Checklist

Pass the authenticated catalog/runtime gate without creating a producer workspace.
Create nine empty row workspaces and execute each fixed caller/model tuple exactly once in its row workspace, with no retry/resume/recovery.
Fill the nine-row result table from immutable evidence, using caller-provided usage or 미제공.
After all attempts, create one shuffled opaque bijection and copy/extract each scorable exact source without route facts.
Render each scorable opaque source once at desktop and once at mobile, then score it once with locked anchors and direct evidence.
Write a bounded conclusion comparing only successful scorable results and separating operational facts from quality.
Run the final count, isolation, placeholder, retry, secret, arithmetic, and scope checks.
Fill implementation-owned sections in CODE_REVIEW-*-G??.md with actual implementation notes and verification output.

[TEST-1] Consume the Immutable Nine-Row Matrix

Problem: agent-test/dev/iop-thin-agent-model-comparison.md:28-41 requires nine empty-workspace single attempts. The prior refinement appended /v1/models to an unnormalized URL, released the token before attempts, and invoked Claude from the repository root rather than the row workspace; those defects can fail the gate or invalidate that comparison.

Solution: Run one remote zsh session. Complete the gate in Final Verification, keeping api_root, iop_token, ca, run_root, and the prompt alive. Create all row workspaces before row 01. Execute these tuples in order: row 01 Claude/claude-sonnet-5; 02 Claude/gemini-3.6-flash; 03 OpenCode/gemini-3.6-flash; 04 Claude/gpt-5.6-luna; 05 Codex/gpt-5.6-luna; 06 Claude/gemini-hybrid; 07 OpenCode/gemini-hybrid; 08 Claude/gpt-hybrid; 09 Codex/gpt-hybrid.

Before each command, require missing attempt.txt and producer.jsonl, require an empty workspace, and write caller/model/version/start plus attempt_count=1, retry=0, resume=0. Wrap the sole invocation with /opt/homebrew/bin/gtimeout 900, redirect its only stream to producer.jsonl, and append end/elapsed/exit. Use the following exact caller forms, substituting only the tuple values:

# Claude rows: invoke from the row workspace.
( cd "$workspace" && ANTHROPIC_BASE_URL="$api_root" ANTHROPIC_AUTH_TOKEN="$iop_token" NODE_EXTRA_CA_CERTS="$ca" CLAUDE_CODE_MAX_RETRIES=0 /opt/homebrew/bin/gtimeout 900 claude --print --output-format stream-json --verbose --bare --no-session-persistence --dangerously-skip-permissions --model "$model" "$(cat "$run_root/prompt.txt")" ) > "$run_root/$row/producer.jsonl" 2>&1

# OpenCode rows: --dir and command-scoped config bind the row workspace.
IOP_BENCH_TOKEN="$iop_token" OPENCODE_CONFIG_CONTENT="$(jq -cn --arg base "$api_root/v1" --arg model "$model" '{permission:{read:"allow",write:"allow",edit:"allow",glob:"allow",bash:"allow"},provider:{iop:{npm:"@ai-sdk/openai-compatible",options:{baseURL:$base,apiKey:"{env:IOP_BENCH_TOKEN}"},models:{($model):{name:$model}}}}}')" NODE_EXTRA_CA_CERTS="$ca" /opt/homebrew/bin/gtimeout 900 opencode run --pure --auto --model "iop/$model" --agent build --format json --dir "$workspace" "$(cat "$run_root/prompt.txt")" > "$run_root/$row/producer.jsonl" 2>&1

# Codex rows: isolated auth/config home and explicit row cwd.
codex_home="$(mktemp -d)"
cp /Users/toki/.codex/auth.json "$codex_home/auth.json"
cp /Users/toki/.codex/iop-direct.config.toml "$codex_home/iop-direct.config.toml"
CODEX_HOME="$codex_home" IOP_CODEX_API_KEY="$iop_token" CODEX_CA_CERTIFICATE="$ca" /opt/homebrew/bin/gtimeout 900 codex exec --ephemeral --json --sandbox workspace-write --skip-git-repo-check --cd "$workspace" --profile iop-direct --model "$model" --output-last-message "$run_root/$row/terminal.txt" "$(cat "$run_root/prompt.txt")" > "$run_root/$row/producer.jsonl" 2>&1
rm -rf "$codex_home"

Capture status with set +e/status=$?/set -e around each exact invocation; cleanup is not a retry. For direct rows 01-05, accept workspace/index.html only with exactly one terminal BENCH_LITE_01_DONE. For preset rows 06-09, extract terminal text once from that row's sole stream; if its caller-native terminal event is absent or ambiguous, mark 채점 불가. A terminal is scorable only when this strict one-shot extractor finds exactly one html fence: ruby -e 's=File.binread(ARGV[0]);m=s.scan(/```html\r?\n(.*?)\r?\n```/m);abort("expected exactly one html fence") unless m.length==1;File.binwrite(ARGV[1],m[0][0])' terminal.txt index.html.

After row 09, unset iop_token, transfer the ignored run directory once to this checkout, then create opaque-map.txt exactly once with a shuffled row/E01-E09 bijection. Copy only scorable exact HTML into the mapped E*/index.html; write only its SHA-256 to source.txt. Do not place route, caller, model, time, usage, or row id in any E* directory. Freeze all scores before joining the mapping.

Modified Files and Checklist:

agent-test/dev/iop-thin-agent-model-comparison.md: immutable result rows and observations.
agent-test/runs/bench-lite-01/prompt.txt: exact fixed prompt.
agent-test/runs/bench-lite-01/catalog.json: credential-free catalog body.
agent-test/runs/bench-lite-01/runtime.txt: redacted preflight identity.
agent-test/runs/bench-lite-01/opaque-map.txt: one post-attempt bijection.
Row and opaque evidence files enumerated in Modified Files Summary.

Test Strategy: No new test code or common runner. The measured behavior is the nine live caller invocations; immutable ledgers and sole streams are the regression evidence.

Verification: Run Final Verification. Expect gate success before run_root, nine ledgers/streams, one attempt marker per row, exact empty-workspace binding, and no retry/resume/recovery.

[TEST-2] Render, Score Once, and Conclude

Problem: agent-test/dev/iop-thin-agent-model-comparison.md:43-91 requires source plus desktop/mobile evidence under a single blind evaluation pass, while its conclusion must not turn missing or failed data into zero.

Solution: For each scorable E*/index.html, create fresh temporary Chromium profiles outside the repository and invoke Chromium once with --window-size=1440,900 --screenshot=desktop.png, then once with --window-size=390,844 --screenshot=mobile.png. Record viewport and sole exit status in render.txt; a failed render is not repeated. Score only the opaque source/renders, use the fixed A selectors and B-D anchors 0/1/3/5, record direct evidence and each deduction, and verify A+B+C+D. Join route facts only after all score blocks are frozen. Compare only successful scorable results; keep failure, unscorable output, and 미제공 separate.

Modified Files and Checklist:

agent-test/dev/iop-thin-agent-model-comparison.md: score rows, evidence blocks, allowed correction notes, bounded conclusion.
Opaque render evidence enumerated in Modified Files Summary.

Test Strategy: No automated judge/browser gate. Fixed source, two one-shot renders, locked anchors, reviewer inspection, and arithmetic are the required evidence.

Verification: Expect every scorable ID to have exactly one source, SHA record, two images, and one two-entry render ledger; every score has anchor/evidence and correct arithmetic.

Modified Files Summary

File	Items
`agent-test/dev/iop-thin-agent-model-comparison.md`	TEST-1, TEST-2
`agent-test/runs/bench-lite-01/prompt.txt`	TEST-1
`agent-test/runs/bench-lite-01/catalog.json`	TEST-1
`agent-test/runs/bench-lite-01/runtime.txt`	TEST-1
`agent-test/runs/bench-lite-01/opaque-map.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-01/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-01/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-01/workspace/index.html`	TEST-1
`agent-test/runs/bench-lite-01/row-02/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-02/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-02/workspace/index.html`	TEST-1
`agent-test/runs/bench-lite-01/row-03/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-03/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-03/workspace/index.html`	TEST-1
`agent-test/runs/bench-lite-01/row-04/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-04/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-04/workspace/index.html`	TEST-1
`agent-test/runs/bench-lite-01/row-05/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-05/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-05/terminal.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-05/workspace/index.html`	TEST-1
`agent-test/runs/bench-lite-01/row-06/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-06/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-06/terminal.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-07/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-07/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-07/terminal.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-08/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-08/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-08/terminal.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-09/attempt.txt`	TEST-1
`agent-test/runs/bench-lite-01/row-09/producer.jsonl`	TEST-1
`agent-test/runs/bench-lite-01/row-09/terminal.txt`	TEST-1
`agent-task/m-thin-agent-model-comparison-benchmark/CODE_REVIEW-cloud-G08.md`	TEST-1, TEST-2
`agent-test/runs/bench-lite-01/E01/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E01/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E01/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E01/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E01/render.txt`	TEST-2
`agent-test/runs/bench-lite-01/E02/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E02/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E02/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E02/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E02/render.txt`	TEST-2
`agent-test/runs/bench-lite-01/E03/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E03/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E03/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E03/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E03/render.txt`	TEST-2
`agent-test/runs/bench-lite-01/E04/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E04/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E04/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E04/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E04/render.txt`	TEST-2
`agent-test/runs/bench-lite-01/E05/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E05/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E05/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E05/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E05/render.txt`	TEST-2
`agent-test/runs/bench-lite-01/E06/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E06/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E06/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E06/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E06/render.txt`	TEST-2
`agent-test/runs/bench-lite-01/E07/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E07/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E07/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E07/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E07/render.txt`	TEST-2
`agent-test/runs/bench-lite-01/E08/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E08/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E08/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E08/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E08/render.txt`	TEST-2
`agent-test/runs/bench-lite-01/E09/index.html`	TEST-1
`agent-test/runs/bench-lite-01/E09/source.txt`	TEST-1
`agent-test/runs/bench-lite-01/E09/desktop.png`	TEST-2
`agent-test/runs/bench-lite-01/E09/mobile.png`	TEST-2
`agent-test/runs/bench-lite-01/E09/render.txt`	TEST-2

Final Verification

Before creating run_root, run in one remote zsh shell:

set -euo pipefail
cd /Users/toki/agent-work/iop-dev
run_root="$PWD/agent-test/runs/bench-lite-01"
test ! -e "$run_root"
ca="$PWD/build/dev-runtime/.secrets/credential-plane/ca.pem"
secret=/Users/toki/.config/iop/secrets/dev-openai-toki.sops.yaml
export SOPS_AGE_KEY_FILE=/Users/toki/.config/sops/age/keys.txt
base_url="$(/opt/homebrew/bin/sops -d --extract '["base_url"]' "$secret")"
api_root="${base_url%/}"; api_root="${api_root%/v1}"
iop_token="$(/opt/homebrew/bin/sops -d --extract '["tokens"]["toki-dev-cline"]' "$secret")"
test -n "$api_root"; test -n "$iop_token"; test -f "$ca"
test -z "$(git status --short)"; test "$(git branch --show-current)" = dev
test "$(git rev-parse HEAD)" = 16b7aba95a282b6c5d1e88d3b1849eaa1208b28a
command -v /opt/homebrew/bin/gtimeout jq ruby claude opencode codex
claude --help | rg -- '--print|--output-format|--verbose|--no-session-persistence|--bare'
opencode run --help | rg -- '--pure|--model|--agent|--format|--dir'
codex exec --help | rg -- '--ephemeral|--json|--sandbox|--cd|--profile|--model|--output-last-message'
nc -z 127.0.0.1 18083; nc -z 127.0.0.1 19093
tmp_catalog="$(mktemp)"
http_code="$(curl --cacert "$ca" -sS -o "$tmp_catalog" -w '%{http_code}' -H "Authorization: Bearer $iop_token" "$api_root/v1/models")"
test "$http_code" = 200
for model in claude-sonnet-5 gemini-3.6-flash gpt-5.6-luna gemini-hybrid gpt-hybrid; do jq -e --arg model "$model" '.data[] | select(.id == $model)' "$tmp_catalog" >/dev/null; done
mkdir "$run_root"; mv "$tmp_catalog" "$run_root/catalog.json"
printf 'branch=dev\nhead=%s\nports=18083,19093\ncatalog_http=200\n' "$(git rev-parse HEAD)" > "$run_root/runtime.txt"

Copy the exact fixed prompt into prompt.txt, create all nine row workspaces, and execute the expanded caller forms in TEST-1. Preserve redacted expanded commands, exit statuses, and ledgers in the active review. Unset iop_token only after row 09.

For each scorable opaque ID, run once with a fresh profile per viewport and record both statuses:

opaque_id=E01; opaque_dir="$PWD/agent-test/runs/bench-lite-01/$opaque_id"; render_log="$opaque_dir/render.txt"
test ! -e "$render_log"; : > "$render_log"
profile="$(mktemp -d)"; printf 'viewport=1440x900\n' >> "$render_log"
set +e; /config/.local/bin/chromium --headless --disable-gpu --hide-scrollbars --run-all-compositor-stages-before-draw --user-data-dir="$profile" --window-size=1440,900 --screenshot="$opaque_dir/desktop.png" "file://$opaque_dir/index.html"; status=$?; set -e
printf 'exit_status=%s\n' "$status" >> "$render_log"
profile="$(mktemp -d)"; printf 'viewport=390x844\n' >> "$render_log"
set +e; /config/.local/bin/chromium --headless --disable-gpu --hide-scrollbars --run-all-compositor-stages-before-draw --user-data-dir="$profile" --window-size=390,844 --screenshot="$opaque_dir/mobile.png" "file://$opaque_dir/index.html"; status=$?; set -e
printf 'exit_status=%s\n' "$status" >> "$render_log"

Run fresh local checks; cached output is not applicable:

test "$(find agent-test/runs/bench-lite-01 -mindepth 2 -maxdepth 2 -name attempt.txt -type f | wc -l)" -eq 9
test "$(find agent-test/runs/bench-lite-01 -mindepth 2 -maxdepth 2 -name producer.jsonl -type f | wc -l)" -eq 9
test "$(rg -l '^attempt_count=1$' agent-test/runs/bench-lite-01/row-*/attempt.txt | wc -l)" -eq 9
test "$(rg -l '^workspace_initial_entries=0$' agent-test/runs/bench-lite-01/row-*/attempt.txt | wc -l)" -eq 9
test "$(cut -d' ' -f1 agent-test/runs/bench-lite-01/opaque-map.txt | LC_ALL=C sort -u | wc -l)" -eq 9
test "$(cut -d' ' -f2 agent-test/runs/bench-lite-01/opaque-map.txt | LC_ALL=C sort -u | wc -l)" -eq 9
scorable="$(find agent-test/runs/bench-lite-01 -mindepth 2 -maxdepth 2 -path '*/E*/index.html' -type f | wc -l)"
for name in desktop.png mobile.png render.txt source.txt; do test "$(find agent-test/runs/bench-lite-01 -mindepth 2 -maxdepth 2 -path "*/E*/$name" -type f | wc -l)" -eq "$scorable"; done
! rg -n '미실행|미측정|미확인|미부여-[0-9]|미채점' agent-test/dev/iop-thin-agent-model-comparison.md
! rg -n '(retry|resume|recovery)[[:space:]]*[:=][[:space:]]*(true|yes|[1-9])' agent-test/runs/bench-lite-01
! rg -n --hidden '(sk-|Bearer [A-Za-z0-9._-]{16,}|api[_-]?key[[:space:]]*[:=][[:space:]]*[A-Za-z0-9._-]{16,})' agent-test/dev/iop-thin-agent-model-comparison.md agent-test/runs/bench-lite-01
! rg -n 'Claude|OpenCode|Codex|claude-sonnet|gemini|gpt|hybrid|경과|usage|row-[0-9]' agent-test/runs/bench-lite-01/E*/source.txt agent-test/runs/bench-lite-01/E*/render.txt
git diff --check -- agent-test/dev/iop-thin-agent-model-comparison.md agent-task/m-thin-agent-model-comparison-benchmark
git diff --name-only -- . ':(exclude)agent-test/dev/iop-thin-agent-model-comparison.md' ':(exclude)agent-task/m-thin-agent-model-comparison-benchmark/**'

The last command must print nothing. Reviewer inspection must confirm gate-before-workspace ordering, exact tuple/workspace binding, one producer invocation per row, explicit usage or 미제공, opaque isolation through score freeze, fixed-anchor arithmetic, and a conclusion that excludes failed/unscorable rows from quality comparison.

After completing all code changes, fill implementation-owned sections in CODE_REVIEW-cloud-G08.md.

23 KiB Raw Blame History