feat(harbor): feedback levers: failure transcripts, multi-fidelity screening, per-attempt detail#30

Merged

varunursekar merged 28 commits into

harbor-all-fixesfrom

harbor-3-feedback-levers

Jul 17, 2026

Commits on Jul 6, 2026

refactor(harbor): split Mode A / Mode B into discriminated-union configs (supersedes #20)

shehabyasser-scale
and
claude
committed

Commits on Jul 8, 2026

fix(harbor): access-tier integrity (free-baseline privilege, k-anonymity floor, trusted nested CLI)

shehabyasser-scale
and
claude
committed
fix(harbor): honest measurement signals (summary qualifiers, status value, versioned re-evals, dead-attempt causes)

shehabyasser-scale
and
claude
committed
fix(harbor): ops integrity (exhaust-budget instruction lever, ledger fails closed on corrupt restore)

shehabyasser-scale
and
claude
committed
fix(harbor): feedback robustness (review follow-ups from #30)

shehabyasser-scale
and
claude
committed
fix(harbor): review follow-ups from #34/#35 (freebie claim atomicity, floor default, ordinal resume, SE naming)

shehabyasser-scale
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(harbor): feedback levers: failure transcripts, multi-fidelity screening, per-attempt detail#30

feat(harbor): feedback levers: failure transcripts, multi-fidelity screening, per-attempt detail#30
varunursekar merged 28 commits into
harbor-all-fixesfrom
harbor-3-feedback-levers

Commits on Jul 5, 2026

feat(harbor): transcript feedback for failed viewable samples (feedback_transcripts)

feat(harbor): instruction teaches multi-fidelity screening (instruct_multifidelity)

feat(harbor): per-attempt reward and exception detail (expose_attempt_detail)

fix(harbor): close multi-fidelity hidden-split leak, subset-eval shortlist pollution, and feedback-cap edge cases

Commits on Jul 6, 2026

refactor(harbor): split Mode A / Mode B into discriminated-union configs (supersedes #20)

Commits on Jul 7, 2026

fix(harbor): close three scoring-integrity gaps an optimizer can game

test(harbor): pin reward-key-mismatch zero-fill in mean mode (review follow-up)

fix(harbor): selection and finalize integrity (idempotency, pooling, retries, fail-safe floor)

fix(harbor): a None aggregate score consumes a retry, never bypasses the loop (review follow-up)

Commits on Jul 8, 2026

fix(harbor): access-tier integrity (free-baseline privilege, k-anonymity floor, trusted nested CLI)

fix(harbor): honest measurement signals (summary qualifiers, status value, versioned re-evals, dead-attempt causes)

fix(harbor): ops integrity (exhaust-budget instruction lever, ledger fails closed on corrupt restore)

fix(harbor): feedback robustness (review follow-ups from #30)

fix(harbor): review follow-ups from #34/#35 (freebie claim atomicity, floor default, ordinal resume, SE naming)

Commits on Jul 9, 2026

fix(harbor): floored rewards name their cause (deterministic candidate crash vs infra outage)

fix(harbor): dominant-cause grouping normalizes task names so cross-task crashes cluster

feat(harbor): transfer targets: per-target executor-model override at finalize

feat(harbor): infra resilience: dead-attempt classification, opt-in outage retry, key-budget alarm

fix(harbor): review follow-ups: validate retry knobs, full multi-round audit history

Commits on Jul 17, 2026

Merge pull request #39 from scaleapi/harbor-12-infra-resilience

Merge pull request #38 from scaleapi/harbor-11-transfer-targets

Merge pull request #37 from scaleapi/harbor-10-crash-visibility

Merge pull request #36 from scaleapi/harbor-9-ops

Merge pull request #35 from scaleapi/harbor-8-honest-signals

Merge pull request #34 from scaleapi/harbor-7-access-tiers

Merge pull request #33 from scaleapi/harbor-6-selection-integrity

Merge pull request #32 from scaleapi/harbor-5-scoring-integrity

Merge pull request #31 from scaleapi/harbor-4-mode-config-split