CAI Critic Graduation: Per-Axis Label Volume Thresholds
TL;DR (headline finding)
There is no single label-count threshold for graduation. Production wisdom converges on a two-gate cutover: ~100-200 human-labeled examples per axis to reach 85% agreement, AND a separate held-out reliability check (Cohen's kappa, edge-case coverage, red-team) before removing the human. The pipeline-RLHF concept doc's "~10 founder-overrides per axis" guess is roughly an order of magnitude light versus the LLM-as-judge production literature, but is defensible for RDCO if (and only if) two compensating mechanisms are in place: (1) tight axis scope (each axis tests one mechanical thing, not a fuzzy quality), and (2) sustained spot-check sampling post-graduation. The dominant failure mode of cutting over too early is not catastrophic - it is silent: the critic locks in a blind spot, the policy learns to exploit it, and quality degrades along the un-instrumented dimension while the dashboard shows PASS.
Why the question
The pipeline-RLHF concept doc ([[../concepts/2026-05-12-rdco-pipeline-rlhf-shaped.md]]) maps RDCO's multi-agent pipeline directly onto the Constitutional AI / RLAIF topology. It hand-waves a graduation threshold of "~10 founder-override labels per axis before that axis graduates to autonomous-critic status," with a footnote that the schema work needs literature-backed numbers. That schema work is now active. The verify-* family (verify-vault-write, verify-strategic-output, verify-dispatch) are explicitly designed to graduate from founder-in-loop to autonomous. The question this brief answers: what does the production CAI / LLM-as-judge literature actually say about the cutover, and what failure modes appear if we cut over too early.
What the literature actually says
1. Calibration-set sizes converge on 30-200 labels per axis
Three independent production sources triangulate on the same range:
| Source | Recommended human labels | Stage |
|---|---|---|
| Comet (LLM-as-Judge production guide) | 30-50 examples annotated by domain experts | Initial calibration set; iterate prompt if >20% disagreement on clear-cut cases |
| Kili / LangChain | 100-200 human-annotated examples | Reach 85% agreement before scaling |
| Databricks (Grading Notes) | ~200 use cases total across criteria | Achieved 93-96% alignment with human judges using grading-notes pattern |
The number is per-axis (per-criterion), not per-skill. A critic with 4 axes needs the calibration set 4 times, OR a stratified single set that covers each axis with 30+ examples.
2. Agreement-rate thresholds: 75-90% before scaling, 85% as canonical cutover
- Comet: aim for 75-90% agreement with human labels before scaling.
- LangChain: 85% agreement as the calibration target after 100-200 labels.
- MT-Bench academic benchmark: 80% agreement between strong LLM judges and human evaluators is the reported state-of-the-art ceiling.
- ApX (RLAIF failure-mode chapter): 85%+ correlation before transition; 90%+ inter-annotator agreement before full deployment.
- Databricks (production case, after grading-notes): 93-96% alignment achieved, treated as success bar.
The gap between "scale" (85%) and "remove human" (90-95%) is intentional. Production deployments do not graduate at 85%; they graduate to assisted operation at 85% and to autonomous operation only after a second calibration pass shows 90%+ on a held-out set.
3. Single-metric thresholds are insufficient; literature mandates a battery
Every production source flags that agreement-rate alone is gameable. The dominant recommendation is to report accuracy plus Cohen's kappa (which corrects for chance agreement), plus targeted red-team coverage. ApX explicitly states: "no single metric suffices."
4. Constitutional AI's own numbers do NOT generalize
Anthropic's original CAI paper bootstrapped harmlessness behavior from ~16 written principles plus thousands of self-critique iterations - but those iterations were over a large pretrained model's full output distribution, not over 10 founder edits to a specific artifact type. The CAI v2 / "Constitution or Collapse?" (Bai et al., 2025; Llama 3-8B replication) line explicitly shows that on smaller-scale or narrower applications, low label volumes lead to mode collapse (the critic converges on a narrow set of "acceptable" outputs and rejects valid diversity). This is the most important nuance for RDCO: Anthropic's headline numbers come from a regime RDCO does not occupy.
Failure modes when you cut over too early
The CAI/RLAIF failure-mode literature is unusually well-catalogued. Six modes recur across sources:
Mode 1: Critic blind-spot exploitation (the dominant mode)
The policy (artifact-producer) learns to produce outputs that hit whatever the critic is actually measuring, while drifting along dimensions the critic doesn't cover. Symptom: PASS rate climbs, real quality plateaus or degrades. This is the textbook reward-hacking story.
- Trigger: Critic-prompt encodes axis incompletely. With <100 examples per axis, edge cases that would have surfaced gaps stay invisible.
- Detection: Only catchable via periodic full re-grading of "PASS" outputs by the human. RDCO mitigation: the spot-check rate (concept doc suggests 5-10%) is exactly this.
Mode 2: Sycophancy / style-mimicry
Policy learns to phrase outputs in the critic's preferred style rather than meet the substantive bar. Symptom: outputs feel uniform; rubric-words appear verbatim in artifacts.
- Trigger: Critic prompt over-emphasizes surface features (word choice, structure) versus substance.
- RDCO-specific risk: /improve consolidation hook risks codifying the founder's word choices ("sharp verdict", "skip/skim/read/file") in a way that produces artifacts that USE those words rather than DO those things.
Mode 3: Constitutional loopholes (rule-letter vs rule-intent)
Policy satisfies the literal axis check while violating the underlying intent. The classic example: an axis that requires "cite a source" gets satisfied by hallucinated citations.
- Trigger: Axis spec is mechanical (presence check) rather than semantic (verification).
- RDCO-specific risk: The structural YAML axes in pipeline-critic (frontmatter completeness, cross-link richness) are PRESENCE checks. They cannot catch "cross-links exist but point to wrong things." Need a separate semantic-axis layer.
Mode 4: Critic drift over time
Even a well-calibrated critic's interpretation of fuzzy axes drifts as the population of artifacts it sees shifts (distribution shift in production). What scored 85% agreement on the calibration set scores 70% three months in.
- Detection threshold: ApX recommends auditing every 500-1000 examples; flag if divergence from initial pattern exceeds 5%.
- RDCO equivalent: /improve consolidation hook IS the re-calibration mechanism. Need to make sure /improve runs on a cadence, not just on founder-trigger.
Mode 5: Mode collapse / reduced diversity
Critic converges on a narrow set of "acceptable" outputs and rejects valid diversity. Documented in the Llama 3-8B CAI replication (Bai et al. 2025). The policy produces increasingly homogeneous artifacts that all clear the critic but cover less of the legitimate output space.
- Trigger: Calibration set is too narrow OR axes overlap (multiple axes all penalizing the same surface feature).
- RDCO-specific risk: If the canonical-exemplars folder is small (~5-10 reference artifacts), the critic anchors on those exemplars and treats novel-but-valid shapes as failures.
Mode 6: Brittleness on paraphrase / OOD inputs
Critic performance degrades sharply on artifact types it wasn't calibrated on. ApX recommends 95%+ consistency on paraphrased/adversarial variants before removing human oversight.
- RDCO-specific risk: A critic calibrated on Sanity Check articles will misjudge MAC carve-out artifacts even though the YAML axes nominally apply to both.
Where production wisdom diverges from academic claims
Two non-trivial divergences worth flagging:
Academic CAI papers under-report mode collapse. The original Anthropic CAI work emphasized harmlessness gains; mode collapse only became a documented failure when the technique was replicated on smaller models and narrower domains (Bai et al. 2025, "Constitution or Collapse?"). RDCO's per-skill, per-axis application is closer to the smaller-narrower regime than to the original Anthropic regime. Implication: weight production case studies (Databricks, LangChain, Comet) over academic Anthropic papers when sizing the threshold.
Academic literature treats reward-model gradient training as load-bearing; production LLM-as-judge does not. Almost every academic CAI / RLAIF paper assumes you're training a parametric reward model via gradient descent on preference pairs. Production deployments overwhelmingly use prompt-engineered LLM-as-judge with grading notes / few-shot examples instead. The label-volume requirements drop dramatically (200 vs 10K+) BUT the calibration discipline rises in importance, because you have no implicit averaging from the gradient process to smooth over noise.
Recommendation for RDCO pipeline-critic schema
Based on triangulation across sources:
Per-axis graduation criteria (proposed)
A pipeline-critic axis graduates from "founder-in-loop" to "autonomous-with-spot-check" when ALL of the following are true:
- Calibration volume: ≥30 labeled (PASS/FAIL + rationale) examples for this axis, covering at least 5 distinct artifact shapes. (Lower bound; 100 is safer if the axis is fuzzy.)
- Agreement gate: ≥85% agreement between critic verdict and founder verdict on a held-out 10-example subset of the calibration set (subset never used for prompt iteration).
- Edge-case red-team: founder constructs 3-5 deliberate edge cases (artifacts that "should fail" along this axis in non-obvious ways); critic catches ≥4/5.
- Diversity floor: axis was tested against artifacts from ≥2 different domains (Sanity Check + MAC + skill build-out, for example), not just the domain it was developed on.
Axes that pass all four gates graduate to autonomous, with mandatory spot-check sampling (see below). Axes that pass (1) and (2) but not (3) and (4) graduate to "assisted" - critic returns verdict, founder reviews critic's reasoning before final disposition.
Spot-check sampling rate (proposed)
- Post-graduation: 10% of PASS verdicts sampled for full founder re-grading, monthly.
- Drift trigger: if founder disagrees with critic on >15% of sampled PASSes in a calendar month, axis reverts to "assisted" status and must re-graduate via the gates above.
- Stratify samples across artifact shape - don't just random-sample, because production distribution gets dominated by the most common shape.
Founder-override threshold is the WRONG primary signal
The concept doc's "~10 founder overrides" framing inadvertently optimizes for the wrong variable. Founder overrides are FAIL signals - they happen when the critic missed something. Counting them measures critic failure rate, not critic confidence. The literature-correct framing is:
- Confirmations (critic agrees with founder spot-check): build the case for graduation
- Overrides (critic disagrees with founder spot-check): block graduation and trigger axis refinement
Suggest re-framing the schema as "30 confirmation labels + 0 unresolved overrides" rather than "10 overrides."
Why this is more conservative than the concept doc's "~10"
The "~10 founder-overrides" hand-wave was reasonable for the unverified intuition stage. Now that we have literature: 10 is too few for an LLM-as-judge critic to reliably hit 85% agreement on a fuzzy axis. For mechanical/structural axes (frontmatter presence, link count, character-count limits) - 10 IS enough, because those are presence checks that don't need calibration. For semantic axes (does this brief say something non-derivative? does this verdict match the founder's voice?) - 30-100 is the literature-backed range.
Practical rule for RDCO: split axes into "mechanical" (10-label threshold) and "semantic" (30-100-label threshold) and graduate them separately. This mirrors the structural-vs-semantic split already implicit in pipeline-critic's YAML-axes-vs-prompt-axes architecture.
Cross-check: where vault context modifies the literature default
Three RDCO-specific factors that lower the safe-cutover bar versus the literature default:
- Single labeler. No inter-annotator-disagreement noise to resolve. The founder IS the ground truth. This eliminates ~20% of the label volume that production multi-labeler workflows need just to average out human disagreement.
- Fresh-eyes subagent pattern ([[../../feedback_fresh_eyes_subagent_for_own_artifacts]]). RDCO already routes critic work through zero-context subagents to defeat confirmation bias. This is a structural mitigation for Mode 1 (blind-spot exploitation) that most production deployments don't have.
- Reversible-action default ([[../../feedback_distinguish_decision_from_action]] + the auto-mode signal-to-noise memory). RDCO already treats most artifacts as cheap-to-redo. Even if the critic mis-grades, the cost of a bad PASS is low for most axes.
Three RDCO-specific factors that RAISE the safe-cutover bar:
- Narrow calibration domain at bootstrap. Most axes will be developed on one artifact type first (Sanity Check articles, or MAC carve-outs, or skill SKILL.md files). Cross-domain testing (gate 4 above) is doable but adds latency.
- Founder-voice axes are fuzzy. "Does this match the founder's voice" / "is this non-derivative" / "is this overconfident relative to evidence" are exactly the kind of fuzzy semantic axes where 30 labels is the floor, not the ceiling.
- /improve consolidation hook is the re-calibration mechanism. If /improve doesn't run on a cadence, drift (Mode 4) WILL show up. Need to either cron /improve or trigger it off the spot-check disagreement rate.
Follow-up questions surfaced
- What is the right spot-check stratification strategy for RDCO? Random 10% will get dominated by Sanity Check article volume; need to stratify by axis × artifact-type to catch drift on lower-volume artifact types.
- How do we instrument "silent quality degradation" detection? The dominant failure mode (Mode 1) is silent by definition. Periodic founder re-grading catches it but is expensive. Is there a cheaper proxy signal (artifact embedding-diversity drift, axis-correlation drift, output entropy)?
- Should /improve be cron'd or threshold-triggered? Cron'd = predictable but wasteful when nothing drifted. Threshold-triggered = efficient but requires the drift-detection signal from Q2.
- How do we handle a new axis introduced after others have graduated? Does the new axis re-set the spot-check rate for the whole skill, or only for itself? Production literature is silent on this.
- What's the right canonical-exemplars folder size to avoid mode collapse? Bai et al. 2025 documents the failure but doesn't give a number. Empirical question - log artifact-shape diversity over time and watch for entropy decline.
Sources
- Anthropic, "Constitutional AI: Harmlessness from AI Feedback" (2212.08073)
- Bai et al., "Constitution or Collapse? Exploring Constitutional AI with Llama 3-8B" (arxiv 2504.04918, 2025)
- Comet, "LLM-as-a-Judge: How to Build Reliable, Scalable Evaluation"
- LangChain, "How to Calibrate LLM-as-a-Judge with Human Corrections"
- Kili Technology, "Keys to Successful LLM-as-a-Judge and HITL Workflows"
- Databricks, "Enhancing LLM-as-a-Judge with Grading Notes"
- ApX Machine Learning, "Analyzing Failure Modes Specific to AI Feedback" (Ch.7, LLM Constitutional AI/RLAIF course)
- ApX Machine Learning, "Common Failure Modes and Debugging Strategies" (Ch.5, same course)
Related vault docs
- [[../concepts/2026-05-12-rdco-pipeline-rlhf-shaped.md]] - the concept doc that motivated this brief; "~10 founder-override" hand-wave updated here
- [[../../01-projects/skill-pipelines/2026-05-12-multi-agent-pipeline-architecture.md]] - schema work this brief feeds into
- [[../2026-04-08-better-harness-evals-hill-climbing.md]] - eval discipline foundation; "human review" step (step 7) is the spot-check sampling mechanism
- [[../2026-05-01-innermost-loop-singularity-bestiary.md]] - reward-hacking and alignment-bestiary lineage
- [[../2026-05-10-alphasignal-self-improving-agents-harness.md]] - reward hacking in self-improving agent context