ce-doc-review confidence scoring: anchored rubric over continuous floats

Problem

Persona-based document review originally used a continuous confidence field (0.0 to 1.0) that synthesis compared against per-severity numeric gates (0.50 / 0.60 / 0.65 / 0.75) and a 0.40 FYI floor. In practice the continuous scale invited false precision: personas clustered on round values (0.60, 0.65, 0.72, 0.80, 0.85), and gate boundaries created coin-flip bands where trivial score shifts moved findings in and out of the actionable tier. The personas were not genuinely differentiating 0.65 from 0.72; the model cannot calibrate self-reported confidence at that granularity.

Symptoms surfaced in review output:

• Single personas filing 3+ findings all rated 0.68-0.72, all variants of the same root premise
• Findings at 0.65 admitted into the actionable tier on noise, not signal
• Residual concerns and deferred questions near-duplicated findings already surfaced, indicating the persona's own ordering did not distinguish "raise this" from "note this"

Reference pattern: Anthropic's anchored rubric

Anthropic's official code-review plugin (anthropics/claude-plugins-official/plugins/code-review/commands/code-review.md) solves the calibration problem with 5 discrete anchors (0, 25, 50, 75, 100) each tied to a behavioral criterion the model can honestly self-apply:

• 0 — false positive or pre-existing issue
• 25 — might be real but couldn't verify; stylistic-not-in-CLAUDE.md
• 50 — verified real but nitpick / not very important
• 75 — double-checked, will hit in practice, directly impacts functionality
• 100 — confirmed, evidence directly confirms, will happen frequently

The rubric is passed verbatim to a separate scoring agent. Filter threshold: >= 80.

Solution adopted for ce-doc-review

Port the structural techniques — anchored rubric, verbatim persona-facing text, explicit false-positive catalog — and tune the filter threshold for document-review economics. The doc-review threshold is >= 50, not Anthropic's >= 80.

Anchor-to-route mapping

Cross-persona corroboration promotes one anchor step (50 → 75, 75 → 100, 100 → 100). This replaces the prior +0.10 numeric boost.

Within-severity sort: anchor descending, then document order as the deterministic final tiebreak.

Files

• plugins/compound-engineering/skills/ce-doc-review/references/findings-schema.json — confidence is an integer enum [0, 25, 50, 75, 100] with behavioral definitions embedded in the description field
• plugins/compound-engineering/skills/ce-doc-review/references/subagent-template.md — the rubric section personas see verbatim, plus the consolidated false-positive catalog
• plugins/compound-engineering/skills/ce-doc-review/references/synthesis-and-presentation.md — anchor-based gate in 3.2, anchor-step promotion in 3.4, anchor-sorted ordering in 3.8, anchor+autofix routing in 3.7
• plugins/compound-engineering/agents/document-review/*.agent.md — each of the 7 personas carries a persona-specific calibration section that maps domain criteria to the shared anchors
• tests/pipeline-review-contract.test.ts — contract tests that assert the schema enforces discrete anchors and the template embeds the rubric

Why the threshold diverges from Anthropic

Code review and document review have different economics. Anthropic's >= 80 filter is load-bearing for code review because of three constraints that do not apply to doc review:

1. Code review has a linter backstop. CI runs linters, typecheckers, and tests. The LLM reviewer is a second layer on top of automated tooling, and a second layer only adds value by being more selective. If automation already catches the 50-75 tier, the LLM surfacing it again is noise.
2. Code review is high-frequency and publicly visible. Every surfaced finding becomes a PR comment. A reviewer who cries wolf 5 times gets muted. Precision dominates recall.
3. Code claims are ground-truth verifiable. "The code does X" can be proven or refuted by reading it. A 75 in code review often means "I couldn't verify" — which means waiting for someone who can.

Document review inverts all three:

1. Doc review IS the backstop. There is no linter that catches a plan's premise gaps or scope drift. A missed finding in the plan derails implementation weeks later.
2. Doc review is low-frequency and private. One review per plan, not per PR. Surfaced findings are dismissed with a keystroke via the routing menu; they are not public commentary.
3. Premise claims have a natural confidence ceiling. "Is the motivation valid?" and "does this scope match the goal?" cannot be verified against ground truth. Personas working in strategy, premise, and adversarial domains (product-lens, adversarial) legitimately cap at anchors 50-75 because full verification is not possible from document text alone. A >= 80 filter would silence those personas.

Filter at >= 50 for doc review; let the routing menu handle volume. Dismissing a surfaced finding is cheap; missing a real concern is expensive.

When to port this pattern

• Other persona-based review skills with similar economics (no linter backstop, one-shot consumption, dismissal cheap via routing). Default threshold for such skills: >= 50.
• Any scoring workflow where the model is asked to self-report confidence on a continuous scale and clustering on round numbers is observed.

When NOT to port directly

• Code review workflows have linter backstops and public-comment costs. Port the rubric structure, but tune the threshold higher (>= 75). See the "ce-code-review migration" section below for the completed port.
• High-throughput pipelines where the 25 anchor ("couldn't verify") represents most findings. Dropping everything below 50 may be too aggressive; consider surfacing 25 as "needs human triage" instead.

Migration history

Landed in a single atomic change because the schema, template, synthesis, rendering, personas, and tests are coupled — a partial migration would have failed validation at every boundary. The schema change is the load-bearing commit; the persona updates and test updates consume it.

Evaluation

After the migration, an A/B evaluation compared baseline (continuous float) against treatment (anchored integer rubric) across four documents spanning size and type: a 7KB in-repo plan, a 63KB in-repo plan, a 27KB external-repo plan, and a 10KB in-repo brainstorm. Both versions were executed by orchestrator subagents reading their matching skill snapshot as prompt material, dispatching all 7 personas, and emitting the Phase 4 headless envelope. The workspace, per-run envelopes, and timing data live under .context/compound-engineering/ce-doc-review-eval/ during the evaluation.

Confirmed effects

• Score dispersion collapsed. Baseline produced 7-12 distinct float values per document (typical: 0.45, 0.50, 0.55, 0.65, 0.72, 0.80, 0.85) — the exact false-precision clustering the migration targeted. Treatment concentrated on 2-3 anchors per document. Anchors 0 and 25 were never emitted by any persona, which matches the template's "suppress silently" instruction for those tiers.
• Cross-persona +1 anchor promotion fires as specified. Observed on cli-printing-press plan (security-lens + feasibility promoting an IP-range-check finding to anchor 100) and interactive-judgment plan (product-lens + adversarial promoting a premise finding to anchor 100).
• Chain linking, safe_auto silent-apply, FYI routing, and per-persona redundancy collapse all exercised correctly on at least one run.
• The >= 50 threshold is load-bearing on large plans. On cli-printing-press, baseline's graduated per-severity gates admitted 13 Decisions; treatment admitted 21. Inspection of the delta confirmed the new findings were genuine concerns the old gates' coin-flip behavior at boundaries was suppressing — not noise. The migration doc's prediction that "missing a real concern is expensive" held in practice.

Anchor-75 calibration boundary discovered

The evaluation surfaced a boundary issue: on large plans, personas emitted anchor 75 for premise-strength concerns ("motivation is thin," "premise is unconvincing") whose "will be hit in practice" claim was the reviewer's reading, not a concrete downstream outcome. This inflated the actionable tier with strength-of-argument critique that was more appropriately observational.

The subagent template's anchor 75 bullet was refined with a calibration paragraph:

Anchor 75 requires naming a concrete downstream consequence someone will hit — a wrong deploy order, an unimplementable step, a contract mismatch, missing evidence that blocks a decision. Strength-of-argument concerns ("motivation is thin," "premise is unconvincing," "a different reader might disagree") do not meet this bar on their own — they are advisory observations and land at anchor 50 unless they also name the specific downstream outcome the reader hits.

The test the template adds: "will a competent implementer or reader concretely encounter this, or is this my opinion about the document's strength?" The former is 75; the latter is 50.

Re-evaluation with the tightened criterion shifted cli-printing-press from 21 Decisions/4 FYI to 10 Decisions/23 FYI — premise-strength concerns moved to observational routing. The change was not a blanket suppression of premise findings: on interactive-judgment plan, the premise challenge survived the tightening and got cross-persona-promoted to anchor 100, because its concrete consequence was explicit ("8-unit redesign creates maintenance debt across three reference files if the premise is wrong"). The refinement distinguishes grounded premise challenges from hand-wavy framing critique — which is the exact precision the rubric was meant to have from the start.

Limitations

• Small corpus. Four documents is enough to confirm macro patterns (clustering, severity inflation, feature coverage) but not to tune threshold values or anchor boundaries at finer granularity.
• Harness drift between iterations. Iteration-1 orchestrators dispatched parallel persona subagents; iteration-2 orchestrators executed personas inline (nested Agent tool unavailable in that session). This affected side metrics (proposed-fix count on cli-printing-press iteration-2 dropped 15 → 4, likely harness-driven rather than tweak-driven) but did not obscure the tweak's core effect, which was large-magnitude.
• No absolute-calibration ground truth. The evaluation measured the migration's stated failure modes disappearing. Whether an anchor-75 finding literally hits 75% of the time remains unmeasured; no labeled doc-review corpus exists.

ce-code-review migration (2026-04-21)

Ported the same anchored-rubric structure into ce-code-review and bundled it with three additional code-review-specific precision controls. The two skills now share calibration discipline but diverge on threshold and on how independent verification is implemented.

Threshold: >= 75 (not >= 50 like ce-doc-review, not >= 80 like Anthropic)

ce-code-review uses anchor 75 as the gate. P0 findings escape at anchor 50.

>= 75 matches the ce-doc-review choice of using the anchor itself as the threshold (no awkward middle-bucket gap). At >= 75, anchors 75 ("real, will hit in practice") and 100 ("verifiable from code alone") survive; anchors 0/25/50 are dropped. Anthropic's >= 80 under a discrete {0,25,50,75,100} scale would collapse to "anchor 100 only," which is too narrow — it would silence findings where personas can construct the trace but cannot literally read the bug off the code.

The threshold divergence from ce-doc-review (>= 50) is correct for the same reasons documented in the "Why the threshold diverges from Anthropic" section above, applied in reverse: code review HAS a linter backstop, IS publicly visible, and code claims ARE ground-truth verifiable. Code review wants narrow precision; doc review wants broad surfacing.

Validation pass (Stage 5b): the deferred follow-up, now landed

The ce-doc-review plan deferred a "neutral-scorer second pass" to a follow-up plan. ce-code-review implements it as Stage 5b: an independent validator sub-agent per surviving finding, mode-conditional dispatch, and a 15-finding budget cap.

• Why now for code review, not doc review: code review has externalizing modes (autofix applies fixes, headless returns findings to programmatic callers) where false positives have real cost — wrong fixes get committed, downstream automation acts on bad signal. Doc review's worst case is a noisy report a user dismisses with a keystroke; code review's worst case is a wrong-fix PR getting merged.
• Mode-conditional dispatch: validation runs in headless, autofix, and the interactive LFG/File-tickets routing paths. It is skipped in interactive walk-through (the human is the per-finding validator) and report-only (nothing is being externalized). This scopes cost to the cases where false positives have real cost.
• Per-finding parallel dispatch, not batched: independence is the design point. A single batched validator looking at all findings together pattern-matches across them and recreates the persona-bias problem we are escaping. Per-file batching is left as a future optimization for reviews with many findings clustered in few files.
• No validated field on findings: an early plan added a validated: boolean field; it was removed during planning. Surviving findings post-validation are validated by definition (rejected ones are dropped); in modes where validation does not run, the run's mode tells consumers everything they need. A field constant within any mode does no work.
• Conservative failure mode: validator timeout, malformed output, or dispatch error → drop the finding. Unverified findings should not externalize.

The validator's protocol is { "validated": true | false, "reason": "<one sentence>" } answering three questions: is the issue real, is it introduced by THIS diff, and is it not handled elsewhere. Template: references/validator-template.md.

Mode-aware false-positive demotion

ce-code-review's broader persona surface (17 reviewers vs ce-doc-review's 7) means more weak general-quality signal. Stricter precision in externalizing modes was already accom