lastlight-evals 0.10.0 → 0.12.0

package/dashboard/dist/index.html

@@ -28,8 +28,8 @@
       rel="stylesheet"
     />
     <title>Last Light — Eval Dashboard</title>
-    <script type="module" crossorigin src="/assets/index-B1jNgf2P.js"></script>
-    <link rel="stylesheet" crossorigin href="/assets/index-Sic0jI7y.css">
+    <script type="module" crossorigin src="/assets/index-LFdFc6Z9.js"></script>
+    <link rel="stylesheet" crossorigin href="/assets/index-B4TAaIUR.css">
   </head>
   <body>
     <div id="root"></div>

package/datasets/aacr-bench/README.md

@@ -0,0 +1,146 @@
+# AACR-Bench — the external adjudication instrument
+
+This directory holds **no data**. AACR-Bench is downloaded on first use into
+`.eval-cache/aacr-bench/dataset.json` (gitignored) and pinned by sha256 in every
+report. This file is the record of what the dataset is, what it is good for, and
+the three things about it that will mislead you if nobody writes them down.
+
+- **Source:** [`Alibaba-Aone/aacr-bench`](https://huggingface.co/datasets/Alibaba-Aone/aacr-bench) on Hugging Face
+- **Licence:** Apache-2.0
+- **Published alongside:** [`alibaba/open-code-review`](https://github.com/alibaba/open-code-review) (Apache-2.0, Go)
+- **Single file:** `dataset.json`, ~2.1 MB, a JSON array of **2,145** objects
+- **Fetched from:** `https://huggingface.co/datasets/Alibaba-Aone/aacr-bench/resolve/main/dataset.json`
+- **sha256 (first fetch, 2026-08-22):** `0804505f0a474765ce2840c832cfeaa6c4f0250dd6ccb169fe73c6758b245a86`
+
+## Why it is not committed
+
+2.1 MB of somebody else's corpus does not belong in this tree, and the eval
+package's `files` list would ship it to npm. The reproducibility story is the
+**sha256 stamped in `report.json`** instead, exactly as `datasets/pr-review/instances.json`
+is generated rather than vendored. `.eval-cache/` is already in `apps/evals/.gitignore`.
+
+Override the cache location with `LASTLIGHT_EVALS_CACHE`, or point at a local copy
+with `--dataset <path>`.
+
+## The consumer
+
+`apps/evals/scripts/aacr-adjudicate.ts` — see its header doc-block, which is the
+long-form version of everything below.
+
+```bash
+cd apps/evals
+
+# The two deterministic floors. No model calls, no keys, no spend.
+npx tsx scripts/aacr-adjudicate.ts --arm keep-all --all
+npx tsx scripts/aacr-adjudicate.ts --arm drop-all --all
+
+# The whole pipeline including prompt construction, with a stub decision.
+npx tsx scripts/aacr-adjudicate.ts --arm llm --dry-run --limit 50 --print-prompt
+
+# A real run. Prints its cost first and REFUSES without --yes.
+npx tsx scripts/aacr-adjudicate.ts --arm llm --limit 50 --yes --out
+```
+
+## Row shape
+
+Every key is present on every one of the 2,145 rows (verified).
+
+| Field | Type | Notes |
+|---|---|---|
+| `project_main_language` | str | C++ 508, TypeScript 422, Java 272, Go 245, C 206, Python 151, JavaScript 141, Rust 92, PHP 56, C# 52 |
+| `pr_url` | str | `https://github.com/OWNER/REPO/pull/N` — 200 distinct PRs across 50 repos |
+| `pr_source_commit` | str | **This is `base.sha`, not the head** (verified 12/12) |
+| `pr_target_commit` | str | A historical head sha |
+| `pr_change_line_count` | int | |
+| `pr_category` | str | e.g. `Code Refactoring / Architectural Improvement` |
+| `is_ai_comment` | bool | **1,597 true / 548 false** |
+| `note` | str | **The review comment text — the whole input.** p50 283 chars, p95 867, max 2,223 |
+| `path` | str | File the comment is on |
+| `side` | str | `right` \| `left` |
+| `source_model` | str | Which model wrote it; **empty string on the 548 human rows** |
+| `from_line`, `to_line` | int | |
+| `category` | str | Code Defect 1022, Maintainability and Readability 905, Performance 144, Security Vulnerability 74 |
+| `context` | str | Diff Level 1017, File Level 744, Repo Level 384 |
+| `label` | int | **1 = expert-verified CORRECT (1,505), 0 = INCORRECT (640)** |
+
+Cross-tabs worth having:
+
+| | label=1 | label=0 | base rate |
+|---|---|---|---|
+| AI-authored | 1,114 | 483 | 69.8% |
+| human-authored | 391 | 157 | 71.4% |
+
+`source_model`: GPT-5.2 575, (human) 548, Claude-Code/Claude-4.5-Sonnet 279,
+Qwen-Coder-480B 254, GLM-4.7 229, Deepseek-V3.2 136, Gemini-3-Pro 124.
+
+## No clone. No checkout. That is the point
+
+Adjudication takes the **comment text plus its metadata** and nothing else. No
+repo is cloned, no sha is resolved, `gh` is never invoked, and the two commit
+fields are recorded for provenance only. That is why this instrument could be
+built before any of the review-reconstruction work in
+[`docs/plans/deterministic-pr-levers.md`](../../../../docs/plans/deterministic-pr-levers.md) —
+it has no dependency on WP1–WP5.
+
+## The three things that will mislead you
+
+### 1. This is not review recall
+
+The script measures whether an arm can tell a valid comment from an invalid one
+**when handed the comment**. It says nothing about whether our reviewer would
+have *generated* that comment — which is the actual bottleneck (1 of 25 gold
+findings on `skillspro`). An arm can score perfectly here and move review recall
+by zero.
+
+### 2. It is not the AACR-Bench leaderboard's metric
+
+The published leaderboard scores review **generation** against the 1,505 label=1
+rows as a gold set to be found — Open Code Review v1.3.1 at 20.0% recall
+(301/1505), Claude Code v2.1.169 at 28.9% (435/1505). Different task, different
+denominator. Per the `01b` house rule and
+[WP9 AC9](../../../../docs/plans/deterministic-pr-levers.md#external-validation-wp9),
+our number and theirs are never averaged, pooled, or put in the same column.
+
+The 640 label=0 rows are, as far as the leaderboard is concerned, not part of the
+gold set at all. They exist for exactly the classification task this script runs,
+and WP9's write-up of the dataset ("1,505 annotated ground-truth issues") does not
+mention them.
+
+### 3. 74% of it is machine-authored
+
+1,597 comments were written by GPT-5.2, Claude-4.5-Sonnet, Qwen-Coder-480B,
+GLM-4.7, Deepseek-V3.2 or Gemini-3-Pro and *then* expert-verified. Only 548 are
+human-authored. So the corpus over-represents the findings AI reviewers already
+produce, and an adjudicator tuned on it is tuned to police machine output. The
+script always reports the two halves separately and never pools them.
+
+Contamination applies as it does to the Martian set: these are public
+repositories, and the PRs are historical. Treat it as a *contaminated, large,
+public* set complementary to the *clean, tiny, private* `skillspro` one — never
+pooled with it either.
+
+## What a result here means for WP6
+
+[WP6](../../../../docs/plans/deterministic-pr-levers.md#adjudication-and-the-attention-boundary-wp6)'s
+adjudicator **may re-rank, re-tier and demote, but may not delete a finding
+without a probe transcript refuting it.** Every time a filter has been measured
+against a conservative reviewer it has cost recall — our own candidate v2
+(micro-recall 1/25 → 2/25, F1 halved, reverted), BitsAI-CR's ReviewFilter
+(precision 54.5 → 67.1, recall 45.5 → **39.8**), and Open Code Review's own
+deterministic layer (discards ~5,100 findings and loses 134 real defects doing
+it).
+
+So the headline metric is **retention** (kept ÷ label=1), not interception. The
+deterministic arms pin what those numbers are *without* the intervention, which
+is the only thing that makes the intervention's numbers a guard rather than a
+decoration:
+
+| Arm | retention | interception | precision | F1 |
+|---|---|---|---|---|
+| `keep-all` — **what production does today** | 100.0% (1505/1505) | 0.0% (0/640) | 70.2% (1505/2145) | 0.825 |
+| `drop-all` | 0.0% (0/1505) | 100.0% (640/640) | n/a (0/0) | n/a |
+
+`keep-all`'s precision is the corpus base rate and its F1 of **0.825** is the bar.
+An arm scoring below it has traded recall for nothing — which is the exact shape
+of the reverted candidate v2. A WP6 adjudicator drops in as a fourth arm by
+implementing the `Arm` interface in the script and registering in `ARMS`.

package/datasets/pr-review/README.md

@@ -19,8 +19,15 @@ holds 50 PRs across Sentry / Grafana / Cal.com / Discourse / Keycloak:
 npx tsx scripts/import-martian.ts            # full 50
 npx tsx scripts/import-martian.ts --limit 3  # a quick subset first
 npx tsx scripts/import-martian.ts --dry-run  # preview without writing
+npx tsx scripts/import-martian.ts --tier pr-review-heldout  # import into a sibling tier dir
 ```
 
+`--tier <name>` (default `pr-review`) picks the tier directory the import writes
+(`datasets/<name>/`) and the `name` stamped into a freshly-created `tier.json`,
+so a held-out split can live beside this tier under its own name. The tier's
+`defaultWorkflow` stays `pr-review` either way — every imported case runs the
+pr-review workflow.
+
 Then run the tier (heavy — clones the real repos, calls a judge model):
 
 ```bash
@@ -50,6 +57,40 @@ Each case's shape (`src/schema.ts`):
 > a real issue the annotators missed scores as a false positive. That understates
 > precision, which is why the default is F1 rather than the precision-weighted F0.5.
 
+## `anchors.json` — the deterministic anchor labels
+
+`anchors.json` **is** committed (unlike `instances.json`) and is the frozen input to the
+code-facts *evidence-coverage* metric. Martian's gold set carries only
+`{severity, description}` — no file, no line — so there is no deterministic join from a gold
+finding to a place in the diff. `scripts/facts-anchors.ts` builds one: it pulls code-shaped
+tokens out of the gold prose (tokenizer `v1`, documented rule-by-rule in the script) and marks a
+finding **anchored** when one of them matches, on a word boundary, an added-or-removed line of the
+three-dot (`base...head`, merge-base) diff. No model is involved — an LLM in the denominator would
+make every coverage number downstream unfalsifiable.
+
+```bash
+npx tsx scripts/facts-anchors.ts              # regenerate + print the report
+npx tsx scripts/facts-anchors.ts --dry-run --audit       # the hand-audit sheet
+npx tsx scripts/facts-anchors.ts --dry-run --unanchored  # what could not be anchored
+```
+
+Three things about it are load-bearing:
+
+- **Freeze the labels, not the tokenizer.** The metric's denominator IS the tokenizer's output, so
+  the artifact is committed and stamps `tokenizer: "v1"`. A better tokenizer ships as `v2` in a new
+  file; it never rewrites what past numbers meant.
+- **It carries its own error bar.** The `audit` block records a seeded random sample of 20 anchored
+  findings, each read by hand, with per-finding verdicts. Quote the anchor rate with that rate
+  attached, and don't tune the tokenizer to improve it.
+- **No gold text.** Only derived labels (anchors, `path:line`, severity, Martian's `bug_type` /
+  `requires_context` / `language`). Join back to your local `instances.json` on
+  `instanceId` + `goldIndex` to read a description.
+
+The headline: **99/137 anchored (72.3%)**. That is a property of the *gold text*, not of
+code-facts — it is the ceiling on what any identifier-level evidence layer could ever be scored
+against. The 38 unanchored are mostly prose with nothing code-shaped in it (20 of them name no
+identifier at all), plus a stylesheet/i18n cluster where the finding is a value, not a name.
+
 ---
 
 **Attribution.** Cases derive from Martian's