Reviews

2 open branches

Merged: 0 24h 16 total

Open

Open mrmrs-1b7902 Correct benchmark outcome attribution, workflow parity and measurement integrity

mrmrs 3d ago

Open oakmark-throughput-benchmark-transplant Add oakmark: per-repo push/clone throughput benchmark for oak (ForgeMark analogue)

mrmrs 2mo ago

Merged

Merged e9b767a5 Preserve measured agent failures and validate campaign copies before admission

mrmrs 9d ago

Merged b297e8a2 Version saved baseline comparisons and enforce claim evidence eligibility; retain explicit diagnostic scoring and outcome denominators

mrmrs 11d ago

Merged f86d1109 Review benchmark validity and add upstream-inspired VCS conformance

mrmrs 11d ago

Merged a9e8fa39 Add a controlled actual-Serve benchmark lane for the legacy selected-branch clone path. Pin fixture/server/client provenance; compare head, manifest, history, worktree bytes/modes, local artifacts, receipt and normalized network work; expose clone plus first status/no-op commit costs and stat-cache validity. Require explicit control identity and harness-controls markers, isolate credentials, bound processes, and state that A/A plus quiet-host evidence remain necessary before any timing claim.

mrmrs 13d ago

Merged 13de5e54 Measure bounded end-to-end agent task costs without production mutations.

mrmrs 13d ago

Merged 88e86b3f Rebuild benchmarks as a clean single-root repository from oak/benchmarks main@009dead264ca; source tree verified by SHA-256 manifest 1cc497d666f1a1c8e27b040edfab9e7730dd48c3b47567e616943e471570a5c8.

mrmrs 2mo ago

Closed

closed without merging

Closed mrmrs-d7b59e Add reproducible end-to-end agent campaigns

mrmrs 12d ago

Closed mrmrs-f0d825 Add frozen local agent campaign plans and resumable durable journals. Validate exact manifests, safe names and paths, deterministic counterbalanced schedules, pinned configuration/harness/binaries, mandatory measurement locking, explicit real-run attestation, bounded admission and agent timeout, started/finished fsync checkpoints, artifact hashes and terminal-row identity. Never silently rerun interrupted episodes. Rebuild derived rows on no-op resume after a journal-to-rows crash. Fourteen contract tests pass; earlier full985 sandbox run had only18 known loopback permission errors and3 skips; combined unrestricted QA runs in integration worker. Carries reviewed foundation; report and subprocess modules integrate from sibling branches. No paid runs.

mrmrs 12d ago

Closed mrmrs-999547 Bound real-agent subprocess execution and capture; isolate per-trial HOME, sanitize Git plumbing, and mark incomplete usage honestly. Includes reviewed benchmark quality foundation.

mrmrs 12d ago

Closed mrmrs-1ed996 Add diagnostic agent campaign accounting, paired outcome and efficiency reports, and staged qualification plan

mrmrs 12d ago

Closed mrmrs-benchmark-gaps-20260910 Close benchmark correctness and saved-comparison gaps

mrmrs 12d ago

Closed mrmrs-473c2f Require exact output path identities and verify full patches against pristine trees; record oracle versions and fail closed on lost application proof

mrmrs 12d ago

Closed mrmrs-8b790b Close config and asset task false passes with shared trusted external oracles. Verify exact 600 config outputs, four binary contents and manifest identities, fixture validator preservation, and complete Git committed task trees. Run real-agent validation with isolated Python outside editable fixtures; version correctness semantics. Negative partial-commit and bypass cases plus Git/Oak workflow and mock-agent positives pass. Carries reviewed benchmark foundation; full sandbox suite930 has only18 known loopback permission errors and3 skips pending unrestricted integration.

mrmrs 12d ago

Closed mrmrs-9d1d51 Harden agent correctness observations: fail closed on failed or unsupported status, parse current Oak structured history and status, require new snapshots for mutating tasks, and detect mode/type/symlink changes. Version oracle semantics, retain null dirtiness, add negative tests and a source-based coverage audit. Targeted 23 tests and actual Oak 0.103.0 tiny commit/status probe pass; full sandbox run has socket permission errors pending unrestricted integration validation.

mrmrs 12d ago

Closed mrmrs-9df99c Reject incompatible measurement populations in statistics and scorecards; document methodology audit and acceptance criteria

mrmrs 12d ago

Closed mrmrs-dfa321 Add exact VCS preservation diagnostics and pinned upstream test review

mrmrs 12d ago

docs/adr/0004-efficiency-gates-oak-baseline-only.md 29 lines · 1.2 KB

ADR-0004: Efficiency gates apply to the Oak-baseline comparison only

Status: accepted
Date: 2026-06-10

Context

regression_report.py gates on exact agent-efficiency metrics (tool calls, output bytes). Oak currently emits more text than Git by design in several operations; gating Oak-vs-Git on output bytes would permanently fail and teach people to ignore the gate. The release-blocking question is "did this changeset make Oak worse than released Oak", not "does Oak emit more than Git" — the latter is tracked as an optimization metric, not a gate.

Decision

Efficiency thresholds (--tool-calls-threshold-pct, --output-bytes-threshold-pct, --tokens-threshold-pct) apply only to target-vs-Oak-baseline comparisons. The Git comparison keeps latency guardrails only. Token-estimate gates default off because char/4 estimates are advisory (see token calibration in docs/statistical-methodology.md).

Consequences

CI can run --fail-on-regression without permanent Oak-vs-Git failures.
Oak-vs-Git token/output deltas remain visible in every report as optimization signal.
If Oak-vs-Git output parity becomes a goal, add a separate explicit gate rather than widening this one.

# ADR-0004: Efficiency gates apply to the Oak-baseline comparison only

- Status: accepted
- Date: 2026-06-10

## Context

`regression_report.py` gates on exact agent-efficiency metrics (tool calls,
output bytes). Oak currently emits more text than Git by design in several
operations; gating Oak-vs-Git on output bytes would permanently fail and teach
people to ignore the gate. The release-blocking question is "did this
changeset make Oak worse than released Oak", not "does Oak emit more than
Git" — the latter is tracked as an optimization metric, not a gate.

## Decision

Efficiency thresholds (`--tool-calls-threshold-pct`, `--output-bytes-threshold-pct`,
`--tokens-threshold-pct`) apply only to target-vs-Oak-baseline comparisons.
The Git comparison keeps latency guardrails only. Token-estimate gates default
off because char/4 estimates are advisory (see token calibration in
`docs/statistical-methodology.md`).

## Consequences

- CI can run `--fail-on-regression` without permanent Oak-vs-Git failures.
- Oak-vs-Git token/output deltas remain visible in every report as
  optimization signal.
- If Oak-vs-Git output parity becomes a goal, add a separate explicit gate
  rather than widening this one.

1	`# ADR-0004: Efficiency gates apply to the Oak-baseline comparison only`
2
3	`- Status: accepted`
4	`- Date: 2026-06-10`
5
6	`## Context`
7
8	`regression_report.py` gates on exact agent-efficiency metrics (tool calls,
9	`output bytes). Oak currently emits more text than Git by design in several`
10	`operations; gating Oak-vs-Git on output bytes would permanently fail and teach`
11	`people to ignore the gate. The release-blocking question is "did this`
12	`changeset make Oak worse than released Oak", not "does Oak emit more than`
13	`Git" — the latter is tracked as an optimization metric, not a gate.`
14
15	`## Decision`
16
17	Efficiency thresholds (`--tool-calls-threshold-pct`, `--output-bytes-threshold-pct`,
18	`--tokens-threshold-pct`) apply only to target-vs-Oak-baseline comparisons.
19	`The Git comparison keeps latency guardrails only. Token-estimate gates default`
20	`off because char/4 estimates are advisory (see token calibration in`
21	`docs/statistical-methodology.md`).
22
23	`## Consequences`
24
25	- CI can run `--fail-on-regression` without permanent Oak-vs-Git failures.
26	`- Oak-vs-Git token/output deltas remain visible in every report as`
27	`optimization signal.`
28	`- If Oak-vs-Git output parity becomes a goal, add a separate explicit gate`
29	`rather than widening this one.`