Which AI reviewer catches real bugs?

Needlefish measures which model and review setup finds real pull-request defects without blocking clean changes. Every lane reviews the same guarded scenarios three times, and every score links to raw evidence.

87 × 3review scenarios × independent draws
87current fixtures
3draws per lane
sealedholdouts included
v2anti-cheat generation

Current leaderboard

A lane is the complete combination of model, agent harness, provider route, and effort. Balanced Review Accuracy gives equal weight to anchored recall and usable specificity. Tier-1 recall and positive-fixture noise ≤ 0.12 are hard gates.

Current deployment

DeepSeek V4.1 Flash via DeepSeek via CLIProxyAPI is deployed and passes the current qualification gates.

Scroll horizontally to see every column.

RankModelBalanced score95% CIStatus
1
DeepSeek V4.1 Flash high
97.81%95.6%–100.0%Deployed
Exact model
deepseek-v4.1-flash-expires-on-0910
Harness
Pi 0.85.1 (operator-attested)
Runner ID
pi
Route
DeepSeek API through private managed proxy (operator-attested)
Provider
DeepSeek via CLIProxyAPI (operator-attested)
Effort
high
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner pi --model deepseek-v4.1-flash-expires-on-0910 --effort high --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --env PI_AUTH_MODE=proxy --env PI_PROVIDER=cliproxy-deepseek --draws 3 --concurrency 3 --holdout include --gate-class R --report eval/reports/2026-09-09-pi-deepseek-v41-flash-high-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 95.6%Usable specificity 100.0%Tier-1 100.0%Tier-2 97.2%Tier-3 90.7%False positive 0.0%Positive noise / review 0.027Invalid 0.0%Verdict match 97.3%Mean time 31sRecall · adjacent-independent-defects-one-file 100.0%Recall · ai-slop-error-swallow 77.8%Recall · ai-slop-error-swallow-defer-close 100.0%Recall · async-check-then-act 100.0%Recall · authorization-guard-weakened 100.0%Recall · behavior-file-moved-out-of-effective-path 100.0%Recall · boundary-flip-accepts-zero 100.0%Recall · boundary-flip-rejects-zero-override 100.0%Recall · bundle-basesha-inconsistent-with-diff-range 100.0%Recall · cache-key-missing-tenant 100.0%Recall · cross-file-unit-contract-break 100.0%Recall · default-sort-lexicographic 100.0%Recall · diff-against-base-tip-not-mergebase 100.0%Recall · duplicate-reinvented-config 100.0%Recall · error-swallow-payment-path 100.0%Recall · fail-open-check-conclusion+stale-checkout-vs-live-head-race 0.0%Recall · fail-open-finding-field-coercion 100.0%Recall · fail-open-malformed-review-json 100.0%Recall · fail-open-severity-downgrade 100.0%Recall · fallback-review-missing-commit-pin 66.7%Recall · gh-cli-missing-repo-flag-before-checkout 100.0%Recall · goroutine-leak-no-sync 100.0%Recall · hardcoded-credential 100.0%Recall · haystack-off-by-one 100.0%Recall · inverted-permission-check 100.0%Recall · lenient-candidate-parse-drops-real-finding 100.0%Recall · manual-dispatch-wrong-ref 100.0%Recall · migration-rename-misses-reader 100.0%Recall · missing-maxbuffer-large-diff-abort 100.0%Recall · missing-security-flag 100.0%Recall · multi-defect-single-diff 100.0%Recall · over-block-shared-predicate 100.0%Recall · over-block-viewer-dead-branch 100.0%Recall · pagination-rounds-down 100.0%Recall · pandas-in-place-mutation 100.0%Recall · removed-none-guard 100.0%Recall · review-options-not-forwarded-large-mode 100.0%Recall · rounding-carry-overflow 100.0%Recall · self-inconsistent-emitted-output 100.0%Recall · silent-hotspot-truncation-coverage-gap+residual-risks-dropped-in-map-reduce-merge+critic-missing-diff-context-large-mode 33.3%Recall · spec-impl-drift-doc-trust 100.0%Recall · spec-impl-drift-flag-ignored 100.0%Recall · spec-impl-drift-name-trust 100.0%Recall · spec-impl-drift-unwrap-trust 100.0%Recall · string-replace-dollar-token-corruption 100.0%Recall · subprocess-inherits-write-token 100.0%Recall · supply-chain-curl-pipe-sh 100.0%Recall · tenant-cache-bleed 100.0%Recall · timing-unsafe-signature-compare 100.0%Recall · type-error-generalized-prop 100.0%Recall · untrusted-code-on-persistent-runner 100.0%Recall · untrusted-self-review-tool-checkout 100.0%Recall · use-after-move 100.0%Recall · utc-to-local-time-drift 100.0%Recall · workflow-secret-exposure 100.0%Recall · write-moved-outside-transaction 100.0%定位 · adjacent-independent-defects-one-file 100.0%定位 · ai-slop-error-swallow 66.7%定位 · ai-slop-error-swallow-defer-close 100.0%定位 · async-check-then-act 100.0%定位 · authorization-guard-weakened 0.0%定位 · behavior-file-moved-out-of-effective-path 100.0%定位 · boundary-flip-accepts-zero 100.0%定位 · boundary-flip-rejects-zero-override 100.0%定位 · bundle-basesha-inconsistent-with-diff-range 100.0%定位 · cache-key-missing-tenant 100.0%定位 · cross-file-unit-contract-break 100.0%定位 · default-sort-lexicographic 100.0%定位 · diff-against-base-tip-not-mergebase 100.0%定位 · duplicate-reinvented-config 66.7%定位 · error-swallow-payment-path 100.0%定位 · fail-open-check-conclusion+stale-checkout-vs-live-head-race 100.0%定位 · fail-open-finding-field-coercion 100.0%定位 · fail-open-malformed-review-json 100.0%定位 · fail-open-severity-downgrade 100.0%定位 · fallback-review-missing-commit-pin 66.7%定位 · gh-cli-missing-repo-flag-before-checkout 100.0%定位 · goroutine-leak-no-sync 100.0%定位 · hardcoded-credential 100.0%定位 · haystack-off-by-one 100.0%定位 · inverted-permission-check 100.0%定位 · lenient-candidate-parse-drops-real-finding 100.0%定位 · manual-dispatch-wrong-ref 100.0%定位 · migration-rename-misses-reader 100.0%定位 · missing-maxbuffer-large-diff-abort 100.0%定位 · missing-security-flag 100.0%定位 · multi-defect-single-diff 100.0%定位 · over-block-shared-predicate 0.0%定位 · over-block-viewer-dead-branch 100.0%定位 · pagination-rounds-down 100.0%定位 · pandas-in-place-mutation 100.0%定位 · removed-none-guard 100.0%定位 · review-options-not-forwarded-large-mode 100.0%定位 · rounding-carry-overflow 100.0%定位 · self-inconsistent-emitted-output 100.0%定位 · silent-hotspot-truncation-coverage-gap+residual-risks-dropped-in-map-reduce-merge+critic-missing-diff-context-large-mode 100.0%定位 · spec-impl-drift-doc-trust 100.0%定位 · spec-impl-drift-flag-ignored 100.0%定位 · spec-impl-drift-name-trust 100.0%定位 · spec-impl-drift-unwrap-trust 100.0%定位 · string-replace-dollar-token-corruption 100.0%定位 · subprocess-inherits-write-token 100.0%定位 · supply-chain-curl-pipe-sh 100.0%定位 · tenant-cache-bleed 100.0%定位 · timing-unsafe-signature-compare 100.0%定位 · type-error-generalized-prop 100.0%定位 · untrusted-code-on-persistent-runner 100.0%定位 · untrusted-self-review-tool-checkout 100.0%定位 · use-after-move 100.0%定位 · utc-to-local-time-drift 100.0%定位 · workflow-secret-exposure 100.0%定位 · write-moved-outside-transaction 100.0%

Raw DeepSeek V4.1 Flash high report
1
Qwen 3.8 Flash max
95.63%92.7%–98.6%Below gate: Tier-1 miss
Exact model
qwen3.8-flash
Harness
Pi 0.85.1 (operator-attested)
Runner ID
pi
Route
Bailian API through private managed proxy (operator-attested)
Provider
Alibaba Bailian via CLIProxyAPI (operator-attested)
Effort
max
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner pi --model qwen3.8-flash --effort max --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --env PI_AUTH_MODE=proxy --env PI_PROVIDER=cliproxy-bailian --draws 3 --concurrency 3 --holdout include --gate-class R --report eval/reports/2026-09-09-pi-qwen38-flash-max-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 91.3%Usable specificity 100.0%Tier-1 95.2%Tier-2 95.4%Tier-3 81.5%False positive 0.0%Positive noise / review 0.055Invalid 1.1%Verdict match 93.5%Mean time 118sRecall · adjacent-independent-defects-one-file 100.0%Recall · ai-slop-error-swallow 100.0%Recall · ai-slop-error-swallow-defer-close 66.7%Recall · async-check-then-act 100.0%Recall · authorization-guard-weakened 100.0%Recall · behavior-file-moved-out-of-effective-path 100.0%Recall · boundary-flip-accepts-zero 100.0%Recall · boundary-flip-rejects-zero-override 66.7%Recall · bundle-basesha-inconsistent-with-diff-range 100.0%Recall · cache-key-missing-tenant 100.0%Recall · cross-file-unit-contract-break 100.0%Recall · default-sort-lexicographic 100.0%Recall · diff-against-base-tip-not-mergebase 100.0%Recall · duplicate-reinvented-config 100.0%Recall · error-swallow-payment-path 100.0%Recall · fail-open-check-conclusion+stale-checkout-vs-live-head-race 0.0%Recall · fail-open-finding-field-coercion 100.0%Recall · fail-open-malformed-review-json 100.0%Recall · fail-open-severity-downgrade 100.0%Recall · fallback-review-missing-commit-pin 66.7%Recall · gh-cli-missing-repo-flag-before-checkout 100.0%Recall · goroutine-leak-no-sync 100.0%Recall · hardcoded-credential 100.0%Recall · haystack-off-by-one 100.0%Recall · inverted-permission-check 66.7%Recall · lenient-candidate-parse-drops-real-finding 0.0%Recall · manual-dispatch-wrong-ref 100.0%Recall · migration-rename-misses-reader 100.0%Recall · missing-maxbuffer-large-diff-abort 100.0%Recall · missing-security-flag 100.0%Recall · multi-defect-single-diff 100.0%Recall · over-block-shared-predicate 100.0%Recall · over-block-viewer-dead-branch 100.0%Recall · pagination-rounds-down 66.7%Recall · pandas-in-place-mutation 100.0%Recall · removed-none-guard 100.0%Recall · review-options-not-forwarded-large-mode 100.0%Recall · rounding-carry-overflow 100.0%Recall · self-inconsistent-emitted-output 100.0%Recall · silent-hotspot-truncation-coverage-gap+residual-risks-dropped-in-map-reduce-merge+critic-missing-diff-context-large-mode 0.0%Recall · spec-impl-drift-doc-trust 100.0%Recall · spec-impl-drift-flag-ignored 100.0%Recall · spec-impl-drift-name-trust 100.0%Recall · spec-impl-drift-unwrap-trust 100.0%Recall · string-replace-dollar-token-corruption 66.7%Recall · subprocess-inherits-write-token 100.0%Recall · supply-chain-curl-pipe-sh 100.0%Recall · tenant-cache-bleed 100.0%Recall · timing-unsafe-signature-compare 100.0%Recall · type-error-generalized-prop 100.0%Recall · untrusted-code-on-persistent-runner 100.0%Recall · untrusted-self-review-tool-checkout 100.0%Recall · use-after-move 100.0%Recall · utc-to-local-time-drift 100.0%Recall · workflow-secret-exposure 66.7%Recall · write-moved-outside-transaction 100.0%定位 · adjacent-independent-defects-one-file 100.0%定位 · ai-slop-error-swallow 66.7%定位 · ai-slop-error-swallow-defer-close 66.7%定位 · async-check-then-act 100.0%定位 · authorization-guard-weakened 0.0%定位 · behavior-file-moved-out-of-effective-path 100.0%定位 · boundary-flip-accepts-zero 100.0%定位 · boundary-flip-rejects-zero-override 66.7%定位 · bundle-basesha-inconsistent-with-diff-range 100.0%定位 · cache-key-missing-tenant 100.0%定位 · cross-file-unit-contract-break 100.0%定位 · default-sort-lexicographic 100.0%定位 · diff-against-base-tip-not-mergebase 100.0%定位 · duplicate-reinvented-config 100.0%定位 · error-swallow-payment-path 100.0%定位 · fail-open-check-conclusion+stale-checkout-vs-live-head-race 100.0%定位 · fail-open-finding-field-coercion 100.0%定位 · fail-open-malformed-review-json 100.0%定位 · fail-open-severity-downgrade 100.0%定位 · fallback-review-missing-commit-pin 66.7%定位 · gh-cli-missing-repo-flag-before-checkout 100.0%定位 · goroutine-leak-no-sync 100.0%定位 · hardcoded-credential 100.0%定位 · haystack-off-by-one 100.0%定位 · inverted-permission-check 66.7%定位 · lenient-candidate-parse-drops-real-finding 0.0%定位 · manual-dispatch-wrong-ref 100.0%定位 · migration-rename-misses-reader 100.0%定位 · missing-maxbuffer-large-diff-abort 100.0%定位 · missing-security-flag 100.0%定位 · multi-defect-single-diff 100.0%定位 · over-block-shared-predicate 0.0%定位 · over-block-viewer-dead-branch 100.0%定位 · pagination-rounds-down 66.7%定位 · pandas-in-place-mutation 100.0%定位 · removed-none-guard 100.0%定位 · review-options-not-forwarded-large-mode 100.0%定位 · rounding-carry-overflow 100.0%定位 · self-inconsistent-emitted-output 100.0%定位 · silent-hotspot-truncation-coverage-gap+residual-risks-dropped-in-map-reduce-merge+critic-missing-diff-context-large-mode 100.0%定位 · spec-impl-drift-doc-trust 100.0%定位 · spec-impl-drift-flag-ignored 100.0%定位 · spec-impl-drift-name-trust 100.0%定位 · spec-impl-drift-unwrap-trust 100.0%定位 · string-replace-dollar-token-corruption 66.7%定位 · subprocess-inherits-write-token 100.0%定位 · supply-chain-curl-pipe-sh 100.0%定位 · tenant-cache-bleed 100.0%定位 · timing-unsafe-signature-compare 100.0%定位 · type-error-generalized-prop 33.3%定位 · untrusted-code-on-persistent-runner 100.0%定位 · untrusted-self-review-tool-checkout 100.0%定位 · use-after-move 100.0%定位 · utc-to-local-time-drift 100.0%定位 · workflow-secret-exposure 66.7%定位 · write-moved-outside-transaction 100.0%

Raw Qwen 3.8 Flash max report
1
Grok 4.6 xhigh
95.48%92.2%–98.8%Candidate
Exact model
grok-4.6
Harness
Grok CLI 1.0.13 (operator-attested)
Runner ID
grok
Route
Grok subscription (operator-attested)
Provider
xAI Grok (operator-attested)
Effort
xhigh
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner grok --model grok-4.6 --effort xhigh --provider 'xAI Grok' --route 'Grok subscription' --runner-version 'Grok CLI 1.0.13' --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --draws 3 --concurrency 3 --holdout include --gate-class R --report eval/reports/2026-09-06-grok-grok46-xhigh-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 92.3%Usable specificity 98.6%Tier-1 100.0%Tier-2 96.3%Tier-3 81.5%False positive 1.4%Positive noise / review 0.011Invalid 0.0%Verdict match 97.3%Mean time 230s

Raw Grok 4.6 xhigh report
4
GLM-5.3-Flash max
94.93%91.6%–98.3%Candidate
Exact model
glm-5.3-flash
Harness
Pi 0.85.1 (operator-attested)
Runner ID
pi
Route
Z.AI API through private managed proxy (operator-attested)
Provider
Z.AI via CLIProxyAPI (operator-attested)
Effort
max
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner pi --model glm-5.3-flash --effort max --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --env PI_AUTH_MODE=proxy --env PI_PROVIDER=zai-cliproxy --draws 3 --concurrency 3 --holdout include --gate-class R --report eval/reports/2026-09-09-pi-glm53-flash-max-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 91.3%Usable specificity 98.6%Tier-1 100.0%Tier-2 93.5%Tier-3 83.3%False positive 1.4%Positive noise / review 0.038Invalid 0.0%Verdict match 95.8%Mean time 208sRecall · adjacent-independent-defects-one-file 100.0%Recall · ai-slop-error-swallow 77.8%Recall · ai-slop-error-swallow-defer-close 100.0%Recall · async-check-then-act 100.0%Recall · authorization-guard-weakened 100.0%Recall · behavior-file-moved-out-of-effective-path 100.0%Recall · boundary-flip-accepts-zero 100.0%Recall · boundary-flip-rejects-zero-override 33.3%Recall · bundle-basesha-inconsistent-with-diff-range 66.7%Recall · cache-key-missing-tenant 100.0%Recall · cross-file-unit-contract-break 100.0%Recall · default-sort-lexicographic 100.0%Recall · diff-against-base-tip-not-mergebase 100.0%Recall · duplicate-reinvented-config 100.0%Recall · error-swallow-payment-path 100.0%Recall · fail-open-check-conclusion+stale-checkout-vs-live-head-race 33.3%Recall · fail-open-finding-field-coercion 100.0%Recall · fail-open-malformed-review-json 100.0%Recall · fail-open-severity-downgrade 100.0%Recall · fallback-review-missing-commit-pin 66.7%Recall · gh-cli-missing-repo-flag-before-checkout 100.0%Recall · goroutine-leak-no-sync 100.0%Recall · hardcoded-credential 100.0%Recall · haystack-off-by-one 100.0%Recall · inverted-permission-check 100.0%Recall · lenient-candidate-parse-drops-real-finding 0.0%Recall · manual-dispatch-wrong-ref 100.0%Recall · migration-rename-misses-reader 100.0%Recall · missing-maxbuffer-large-diff-abort 100.0%Recall · missing-security-flag 100.0%Recall · multi-defect-single-diff 100.0%Recall · over-block-shared-predicate 100.0%Recall · over-block-viewer-dead-branch 100.0%Recall · pagination-rounds-down 33.3%Recall · pandas-in-place-mutation 100.0%Recall · removed-none-guard 100.0%Recall · review-options-not-forwarded-large-mode 100.0%Recall · rounding-carry-overflow 100.0%Recall · self-inconsistent-emitted-output 100.0%Recall · silent-hotspot-truncation-coverage-gap+residual-risks-dropped-in-map-reduce-merge+critic-missing-diff-context-large-mode 0.0%Recall · spec-impl-drift-doc-trust 100.0%Recall · spec-impl-drift-flag-ignored 100.0%Recall · spec-impl-drift-name-trust 100.0%Recall · spec-impl-drift-unwrap-trust 100.0%Recall · string-replace-dollar-token-corruption 100.0%Recall · subprocess-inherits-write-token 100.0%Recall · supply-chain-curl-pipe-sh 100.0%Recall · tenant-cache-bleed 100.0%Recall · timing-unsafe-signature-compare 100.0%Recall · type-error-generalized-prop 100.0%Recall · untrusted-code-on-persistent-runner 100.0%Recall · untrusted-self-review-tool-checkout 100.0%Recall · use-after-move 100.0%Recall · utc-to-local-time-drift 100.0%Recall · workflow-secret-exposure 100.0%Recall · write-moved-outside-transaction 100.0%定位 · adjacent-independent-defects-one-file 100.0%定位 · ai-slop-error-swallow 55.6%定位 · ai-slop-error-swallow-defer-close 100.0%定位 · async-check-then-act 100.0%定位 · authorization-guard-weakened 0.0%定位 · behavior-file-moved-out-of-effective-path 100.0%定位 · boundary-flip-accepts-zero 100.0%定位 · boundary-flip-rejects-zero-override 33.3%定位 · bundle-basesha-inconsistent-with-diff-range 66.7%定位 · cache-key-missing-tenant 100.0%定位 · cross-file-unit-contract-break 100.0%定位 · default-sort-lexicographic 100.0%定位 · diff-against-base-tip-not-mergebase 100.0%定位 · duplicate-reinvented-config 66.7%定位 · error-swallow-payment-path 100.0%定位 · fail-open-check-conclusion+stale-checkout-vs-live-head-race 100.0%定位 · fail-open-finding-field-coercion 100.0%定位 · fail-open-malformed-review-json 100.0%定位 · fail-open-severity-downgrade 100.0%定位 · fallback-review-missing-commit-pin 66.7%定位 · gh-cli-missing-repo-flag-before-checkout 100.0%定位 · goroutine-leak-no-sync 100.0%定位 · hardcoded-credential 100.0%定位 · haystack-off-by-one 100.0%定位 · inverted-permission-check 100.0%定位 · lenient-candidate-parse-drops-real-finding 0.0%定位 · manual-dispatch-wrong-ref 100.0%定位 · migration-rename-misses-reader 100.0%定位 · missing-maxbuffer-large-diff-abort 100.0%定位 · missing-security-flag 100.0%定位 · multi-defect-single-diff 100.0%定位 · over-block-shared-predicate 0.0%定位 · over-block-viewer-dead-branch 100.0%定位 · pagination-rounds-down 33.3%定位 · pandas-in-place-mutation 100.0%定位 · removed-none-guard 100.0%定位 · review-options-not-forwarded-large-mode 100.0%定位 · rounding-carry-overflow 100.0%定位 · self-inconsistent-emitted-output 100.0%定位 · silent-hotspot-truncation-coverage-gap+residual-risks-dropped-in-map-reduce-merge+critic-missing-diff-context-large-mode 100.0%定位 · spec-impl-drift-doc-trust 100.0%定位 · spec-impl-drift-flag-ignored 100.0%定位 · spec-impl-drift-name-trust 100.0%定位 · spec-impl-drift-unwrap-trust 100.0%定位 · string-replace-dollar-token-corruption 100.0%定位 · subprocess-inherits-write-token 100.0%定位 · supply-chain-curl-pipe-sh 100.0%定位 · tenant-cache-bleed 100.0%定位 · timing-unsafe-signature-compare 100.0%定位 · type-error-generalized-prop 0.0%定位 · untrusted-code-on-persistent-runner 100.0%定位 · untrusted-self-review-tool-checkout 100.0%定位 · use-after-move 100.0%定位 · utc-to-local-time-drift 100.0%定位 · workflow-secret-exposure 100.0%定位 · write-moved-outside-transaction 100.0%

Raw GLM-5.3-Flash max report
4
GLM-5.3-Flash max
94.81%91.5%–98.1%Below gate: Tier-1 miss
Exact model
zai-direct/glm-5.3-flash
Harness
Pi 0.85.1 (operator-attested)
Runner ID
pi
Route
Z.AI coding-plan subscription (direct API) (operator-attested)
Provider
Z.AI (operator-attested)
Effort
max
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner pi --model zai-direct/glm-5.3-flash --effort max --provider Z.AI --route 'Z.AI coding-plan subscription (direct API)' --runner-version 'Pi 0.85.1' --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --env PI_AUTH_MODE=proxy --env PI_PROVIDER=zai-direct --draws 3 --concurrency 3 --holdout include --gate-class R --report eval/reports/2026-09-06-pi-zai-glm53-flash-max-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 89.6%Usable specificity 100.0%Tier-1 95.2%Tier-2 92.6%Tier-3 81.5%False positive 0.0%Positive noise / review 0.022Invalid 0.0%Verdict match 95.0%Mean time 164s

Raw GLM-5.3-Flash max report
6
DeepSeek V4 Flash Vision Exp max
91.66%87.7%–95.6%Below gate: Tier-1 miss
Exact model
cliproxy-deepseek/deepseek-v4-flash-vision-exp
Harness
Pi 0.85.1 (operator-attested)
Runner ID
pi
Route
DeepSeek API through private managed proxy (operator-attested)
Provider
DeepSeek via CLIProxyAPI (operator-attested)
Effort
max
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner pi --model cliproxy-deepseek/deepseek-v4-flash-vision-exp --effort max --provider 'DeepSeek via CLIProxyAPI' --route 'DeepSeek API through private managed proxy' --runner-version 'Pi 0.85.1' --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --env PI_AUTH_MODE=proxy --env PI_PROVIDER=cliproxy-deepseek --draws 3 --concurrency 3 --holdout include --gate-class R --report eval/reports/2026-09-06-pi-cliproxy-deepseek-v4-flash-vision-exp-max-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 84.7%Usable specificity 98.6%Tier-1 95.2%Tier-2 87.0%Tier-3 75.9%False positive 1.4%Positive noise / review 0.022Invalid 0.8%Verdict match 90.8%Mean time 116s

Raw DeepSeek V4 Flash Vision Exp max report
6
GPT-5.6 Terra xhigh
91.61%86.7%–96.5%Below gate: Tier-1 miss
Exact model
gpt-5.6-terra
Harness
Pi 0.85.1 (operator-attested)
Runner ID
pi
Route
Codex subscription (operator-attested)
Provider
OpenAI Codex (operator-attested)
Effort
xhigh
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner pi --model gpt-5.6-terra --effort xhigh --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --env PI_AUTH_MODE=oauth --env PI_PROVIDER=openai-codex --draws 3 --concurrency 4 --holdout include --gate-class R --report eval/reports/2026-09-09-pi-gpt56-terra-xhigh-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 90.2%Usable specificity 93.1%Tier-1 95.2%Tier-2 95.4%Tier-3 77.8%False positive 6.9%Positive noise / review 0.077Invalid 0.0%Verdict match 95.4%Mean time 76sRecall · adjacent-independent-defects-one-file 100.0%Recall · ai-slop-error-swallow 88.9%Recall · ai-slop-error-swallow-defer-close 100.0%Recall · async-check-then-act 100.0%Recall · authorization-guard-weakened 100.0%Recall · behavior-file-moved-out-of-effective-path 100.0%Recall · boundary-flip-accepts-zero 100.0%Recall · boundary-flip-rejects-zero-override 100.0%Recall · bundle-basesha-inconsistent-with-diff-range 0.0%Recall · cache-key-missing-tenant 100.0%Recall · cross-file-unit-contract-break 100.0%Recall · default-sort-lexicographic 100.0%Recall · diff-against-base-tip-not-mergebase 100.0%Recall · duplicate-reinvented-config 100.0%Recall · error-swallow-payment-path 100.0%Recall · fail-open-check-conclusion+stale-checkout-vs-live-head-race 0.0%Recall · fail-open-finding-field-coercion 100.0%Recall · fail-open-malformed-review-json 100.0%Recall · fail-open-severity-downgrade 100.0%Recall · fallback-review-missing-commit-pin 100.0%Recall · gh-cli-missing-repo-flag-before-checkout 100.0%Recall · goroutine-leak-no-sync 100.0%Recall · hardcoded-credential 100.0%Recall · haystack-off-by-one 100.0%Recall · inverted-permission-check 66.7%Recall · lenient-candidate-parse-drops-real-finding 0.0%Recall · manual-dispatch-wrong-ref 100.0%Recall · migration-rename-misses-reader 100.0%Recall · missing-maxbuffer-large-diff-abort 100.0%Recall · missing-security-flag 100.0%Recall · multi-defect-single-diff 100.0%Recall · over-block-shared-predicate 100.0%Recall · over-block-viewer-dead-branch 100.0%Recall · pagination-rounds-down 66.7%Recall · pandas-in-place-mutation 100.0%Recall · removed-none-guard 100.0%Recall · review-options-not-forwarded-large-mode 100.0%Recall · rounding-carry-overflow 100.0%Recall · self-inconsistent-emitted-output 100.0%Recall · silent-hotspot-truncation-coverage-gap+residual-risks-dropped-in-map-reduce-merge+critic-missing-diff-context-large-mode 0.0%Recall · spec-impl-drift-doc-trust 100.0%Recall · spec-impl-drift-flag-ignored 100.0%Recall · spec-impl-drift-name-trust 100.0%Recall · spec-impl-drift-unwrap-trust 100.0%Recall · string-replace-dollar-token-corruption 100.0%Recall · subprocess-inherits-write-token 100.0%Recall · supply-chain-curl-pipe-sh 100.0%Recall · tenant-cache-bleed 100.0%Recall · timing-unsafe-signature-compare 100.0%Recall · type-error-generalized-prop 100.0%Recall · untrusted-code-on-persistent-runner 100.0%Recall · untrusted-self-review-tool-checkout 100.0%Recall · use-after-move 100.0%Recall · utc-to-local-time-drift 100.0%Recall · workflow-secret-exposure 0.0%Recall · write-moved-outside-transaction 100.0%定位 · adjacent-independent-defects-one-file 100.0%定位 · ai-slop-error-swallow 88.9%定位 · ai-slop-error-swallow-defer-close 100.0%定位 · async-check-then-act 100.0%定位 · authorization-guard-weakened 0.0%定位 · behavior-file-moved-out-of-effective-path 100.0%定位 · boundary-flip-accepts-zero 100.0%定位 · boundary-flip-rejects-zero-override 100.0%定位 · bundle-basesha-inconsistent-with-diff-range 0.0%定位 · cache-key-missing-tenant 100.0%定位 · cross-file-unit-contract-break 100.0%定位 · default-sort-lexicographic 100.0%定位 · diff-against-base-tip-not-mergebase 100.0%定位 · duplicate-reinvented-config 33.3%定位 · error-swallow-payment-path 100.0%定位 · fail-open-check-conclusion+stale-checkout-vs-live-head-race 100.0%定位 · fail-open-finding-field-coercion 100.0%定位 · fail-open-malformed-review-json 100.0%定位 · fail-open-severity-downgrade 100.0%定位 · fallback-review-missing-commit-pin 100.0%定位 · gh-cli-missing-repo-flag-before-checkout 100.0%定位 · goroutine-leak-no-sync 100.0%定位 · hardcoded-credential 100.0%定位 · haystack-off-by-one 100.0%定位 · inverted-permission-check 66.7%定位 · lenient-candidate-parse-drops-real-finding 0.0%定位 · manual-dispatch-wrong-ref 100.0%定位 · migration-rename-misses-reader 100.0%定位 · missing-maxbuffer-large-diff-abort 100.0%定位 · missing-security-flag 100.0%定位 · multi-defect-single-diff 100.0%定位 · over-block-shared-predicate 0.0%定位 · over-block-viewer-dead-branch 100.0%定位 · pagination-rounds-down 66.7%定位 · pandas-in-place-mutation 100.0%定位 · removed-none-guard 100.0%定位 · review-options-not-forwarded-large-mode 100.0%定位 · rounding-carry-overflow 100.0%定位 · self-inconsistent-emitted-output 100.0%定位 · silent-hotspot-truncation-coverage-gap+residual-risks-dropped-in-map-reduce-merge+critic-missing-diff-context-large-mode 100.0%定位 · spec-impl-drift-doc-trust 100.0%定位 · spec-impl-drift-flag-ignored 100.0%定位 · spec-impl-drift-name-trust 100.0%定位 · spec-impl-drift-unwrap-trust 100.0%定位 · string-replace-dollar-token-corruption 100.0%定位 · subprocess-inherits-write-token 100.0%定位 · supply-chain-curl-pipe-sh 100.0%定位 · tenant-cache-bleed 100.0%定位 · timing-unsafe-signature-compare 100.0%定位 · type-error-generalized-prop 66.7%定位 · untrusted-code-on-persistent-runner 100.0%定位 · untrusted-self-review-tool-checkout 100.0%定位 · use-after-move 100.0%定位 · utc-to-local-time-drift 100.0%定位 · workflow-secret-exposure 0.0%定位 · write-moved-outside-transaction 100.0%

Raw GPT-5.6 Terra xhigh report
6
GPT-5.6 Terra high
89.95%83.8%–96.1%Candidate
Exact model
gpt-5.6-terra
Harness
Codex CLI 0.153.4 (operator-attested)
Runner ID
codex
Route
Codex subscription (operator-attested)
Provider
OpenAI Codex (operator-attested)
Effort
high
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner codex --model gpt-5.6-terra --effort high --provider 'OpenAI Codex' --route 'Codex subscription' --runner-version 'Codex CLI 0.153.4' --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --draws 3 --concurrency 4 --holdout include --gate-class R --report eval/reports/2026-09-06-codex-gpt56-terra-high-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 89.6%Usable specificity 90.3%Tier-1 100.0%Tier-2 93.5%Tier-3 77.8%False positive 9.7%Positive noise / review 0.077Invalid 0.0%Verdict match 95.0%Mean time 63s

Raw GPT-5.6 Terra high report
6
GPT-5.6 Sol medium
88.41%81.1%–95.7%Candidate
Exact model
gpt-5.6-sol
Harness
Codex CLI 0.153.4 (operator-attested)
Runner ID
codex
Route
Codex subscription (operator-attested)
Provider
OpenAI Codex (operator-attested)
Effort
medium
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner codex --model gpt-5.6-sol --effort medium --provider 'OpenAI Codex' --route 'Codex subscription' --runner-version 'Codex CLI 0.153.4' --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --draws 3 --concurrency 4 --holdout include --gate-class R --report eval/reports/2026-09-06-codex-gpt56-sol-medium-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 90.7%Usable specificity 86.1%Tier-1 100.0%Tier-2 94.4%Tier-3 79.6%False positive 13.9%Positive noise / review 0.077Invalid 0.0%Verdict match 93.5%Mean time 75s

Raw GPT-5.6 Sol medium report

Status · Deployed ships the current release · Candidate passes the same gates but is not shipped · Below gate missed a hard gate; its score stays visible for comparison. Open a lane name for its diagnostics, reproduce command, and raw report.

Balanced Review Accuracy by lane
020406080100Qualified DeepSeek V4.1 Flashhigh97.8% ± 2.2Qwen 3.8 Flashmax95.6% ± 2.9Grok 4.6xhigh95.5% ± 3.3GLM-5.3-Flashmax94.9% ± 3.3GLM-5.3-Flashmax94.8% ± 3.3DeepSeek V4 Flash Vision Expmax91.7% ± 3.9GPT-5.6 Terraxhigh91.6% ± 4.9GPT-5.6 Terrahigh89.9% ± 6.1GPT-5.6 Solmedium88.4% ± 7.3Below gate · score shown, no rankGPT-5.6 Terraxhigh · Tier-1 90.5% + noise 0.12090.4% ± 5.1GPT-5.6 Lunamax · Tier-1 76.2% + noise 0.13188.4% ± 6.8

Bars show Balanced Review Accuracy with 95% confidence whiskers on a 0–100 scale. Orange ranks first, the green mark ships in the current release, and hollow bars missed a hard gate but keep their score for comparison, with the failed gate noted beside the effort. The table remains canonical.

Below the gate

These complete reports missed a Tier-1 defect or exceeded the 0.12 positive-noise envelope. Their balanced scores remain visible for comparison, but they receive no rank.

Scroll horizontally to see every column.

RankModelBalanced score95% CIStatus
GPT-5.6 Terra xhigh
90.39%85.3%–95.5%Below gate: Tier-1 miss
Exact model
gpt-5.6-terra
Harness
Codex CLI 0.153.4 (operator-attested)
Runner ID
codex
Route
Codex subscription (operator-attested)
Provider
OpenAI Codex (operator-attested)
Effort
xhigh
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner codex --model gpt-5.6-terra --effort xhigh --provider 'OpenAI Codex' --route 'Codex subscription' --runner-version 'Codex CLI 0.153.4' --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --draws 3 --concurrency 4 --holdout include --gate-class R --baseline --report eval/reports/2026-09-06-codex-gpt56-terra-xhigh-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 86.3%Usable specificity 94.4%Tier-1 90.5%Tier-2 92.6%Tier-3 72.2%False positive 4.2%Positive noise / review 0.120Invalid 0.4%Verdict match 96.2%Mean time 80s

Raw GPT-5.6 Terra xhigh report
GPT-5.6 Luna max
88.43%81.6%–95.2%Below gate: Tier-1 miss
Exact model
gpt-5.6-luna
Harness
Codex CLI 0.153.4 (operator-attested)
Runner ID
codex
Route
Codex subscription (operator-attested)
Provider
OpenAI Codex (operator-attested)
Effort
max
Run
· 87 fixtures × 3 draws · include holdouts · anti-cheat v2
Reproduce
node --import tsx eval/run.ts --runner codex --model gpt-5.6-luna --effort max --provider 'OpenAI Codex' --route 'Codex subscription' --runner-version 'Codex CLI 0.153.4' --env NEEDLEFISH_EPHEMERAL_HOME=1 --env NEEDLEFISH_EVAL_TRACE=1 --draws 3 --concurrency 4 --holdout include --gate-class R --report eval/reports/2026-09-06-codex-gpt56-luna-max-x3.json
Hashes
e62d0889fc704541 / e9923bbc7753a04a / 8bbc6152d8b45a43

Recall 88.0%Usable specificity 88.9%Tier-1 76.2%Tier-2 92.6%Tier-3 83.3%False positive 5.6%Positive noise / review 0.131Invalid 1.9%Verdict match 94.6%Mean time 147s

Raw GPT-5.6 Luna max report

Not run

Unavailable routes are shown as blocked, never converted into a zero score.

  • Qwen3.8 Flash Nextqwen3.8-flash-nextOpenCode Go subscription

    Not exposed by the authenticated subscription endpoint or Pi model catalog on 2026-08-31.

Not ranked

Compromised or operationally invalid reports remain visible as evidence but never enter the leaderboard.

  • Qwen3.8 Max incomplete runopencode-go/qwen3.8-maxOpenCode Go subscription

    Incomplete at 218/258 saved draws, including 158 reusable valid results. The authenticated provider returned a monthly usage-limit error on 2026-08-31 and reported a reset in 16 days; paid balance was not enabled. Raw Qwen3.8 Max incomplete run report

  • Grok 4.6 initial rungrok-4.6Grok subscription

    Voided after anti-cheat v2 detected structured canary adoption in candidate and final review text. Excluded from every aggregate and rank. Raw Grok 4.6 initial run report

How to read it

Updated 2026-09-10
prompt e62d0889fc704541
fixtures e9923bbc7753a04a
scorer 8bbc6152d8b45a43

Balanced Review Accuracy

The primary score is the arithmetic mean of anchored recall and usable specificity. Invalid model output cannot count as a correct positive or negative result, so each unusable draw is counted once. Point-sorted uncertainty groups are anchored to their highest-scoring lane; lower lanes share that rank while their paired 95% normal interval versus the anchor includes zero. This avoids non-transitive bridge comparisons. The table also shows each lane's 95% interval. A Tier-1 miss or positive-fixture noise above 0.12 overrides the rank: the lane keeps its score for comparison but leaves the ranked table. Verdict match, invalid rate, and speed remain separate diagnostics.

Anchored recall

A defect counts only when the finding matches the expected behavior and the expected file. Missing a Tier-1 defect removes the lane from the ranked table regardless of its average.

False positives and noise

Clean fixtures measure whether a reviewer blocks a change that should pass. Positive-fixture noise counts unrelated blocking findings on buggy changes; recall bought with noise above the production envelope is not qualified.

Integrity

Every published lane includes sealed holdouts, a disposable HOME, a planted canary, full-transcript scanning, and post-run repository mutation checks. Harness, provider, and route labels are operator-attested report metadata, not independently derived by the site generator. Provider failures are operational failures, not proof of model quality.

Reproduce

Read the chronological experiment record, inspect each raw report, or run the evaluation harness.