Evaltudepreview

Part 10 — Integrated Case Studies

Guardrail experiments

Diagnosis and correction·Evaluation·7 min read

The case study's most instructive result: a correction policy that added ~25 points to the 3B configuration but netted to zero on 7B while changing four trials. This is the two-sided guardrail lesson, measured — aggregate effect and trial-level movement on real systems.

The diagnosis named a precise, concentrated failure — ranking predicted where breakdown was expected — and a precise failure invites a precise fix. So the project built a correction policy targeting exactly that confusion, and the result is the most instructive in the whole case study: it added ~25 points to the configuration it was built for, then, transferred unchanged to a stronger configuration, netted to zero while changing four trials. This is the two-sided guardrail lesson — fixes, regressions, and changed incorrect outputs — measured on real systems. It is why the book insists on comparing pre-policy and post-policy outcomes and inspecting the changed trials.

What you will understand by the end

  • The correction policy the case study built, and what it targeted.
  • Why it added ~25 points to the 3B model.
  • Why the same policy had zero net accuracy effect on 7B while still changing behavior.
  • Why this proves a guardrail is a two-sided trade tuned to one system.

The policy: target the diagnosed failure

The correction policy is a deterministic rule aimed at the exact failure the diagnosis found: when the model predicts top_n_by_metric (a ranking) but the question contains no ranking language ("top", "highest", "best"...), downgrade the pattern to metric_by_dimension_with_filter (a breakdown). It's a rule keyed to the ranking-vs-breakdown boundary — precise because the diagnosis was precise.

   IF model predicts top_n_by_metric  AND  question has no ranking cue
   THEN downgrade pattern → metric_by_dimension_with_filter

+25 points on the 3B configuration

On the Qwen 3B configuration it was built for, the policy worked exactly as intended: the model made the ranking-for-breakdown mistake often, and the rule caught and corrected a large fraction of those, lifting post-policy accuracy by roughly 25 points over the validated pre-policy output. A single hand-written rule, targeting one diagnosed confusion, moved the number more than a lot of model tuning would — because it addressed the dominant failure class.

Key idea

A guardrail keyed to a precisely-diagnosed, concentrated failure is high-leverage: the case study's rule added ~25 points because it targeted the dominant failure type. This is the payoff of good failure analysis — a precise diagnosis enables a precise, powerful fix.

The same rule netted to zero on 7B

Then the revealing part. Transferred unchanged to the stronger Qwen 7B model, the same rule touched four trials: it produced one fix, one regression, and two wrong-to-different-wrong changes. The fix and regression cancelled in aggregate accuracy, so the net effect was zero. But the flat score did not mean "no effect": the policy changed four behaviors and newly broke one task. The rule fires on a condition (a top_n prediction with no ranking cue), and that condition only approximates the failure it targets.

Watch out

The guardrail was tuned to the 3B model's failure mode, so it encodes assumptions about how that model fails. On a model that fails differently, the same firing condition can produce a different mixture of fixes, regressions, and changed incorrect outputs — even when the headline accuracy stays flat. A correction policy is part of a specific system, not a portable upgrade; it must be re-measured on every model it's applied to.

What it proves

This single experiment is the guardrail chapter made concrete on real systems. It proves a guardrail is a two-sided trade: its value is the net of fixes and regressions, not the count of catches; and it's tuned to one system, so it doesn't transfer for free. Comparing the validated pre-policy output and post-policy output (complete system) reveals the aggregate effect, while keeping every intervened trial openable reveals movement that the gap between the two dials can hide. On 7B, that gap is zero even though four trials changed.

Observed evidence

On the 3B configuration, the policy touched 28 trials: 23 fixes, one regression, and four wrong-to-different-wrong changes, for a +25-point net accuracy effect. On the 7B configuration, it touched four: one fix, one regression, and two wrong-to-different-wrong changes, for a zero-point net effect. The flat 7B score concealed meaningful behavioral changes, including one newly broken task. See every case the policy changed →

Mental model

The case study's correction policy targeted the diagnosed ranking-for-breakdown confusion and added ~25 points to the 3B configuration it was built for. Transferred unchanged to 7B, it netted to zero while changing four trials, including one new regression. A guardrail is a two-sided trade tuned to one system. Measure both its net accuracy effect and its trial-level movement.

Common mistakes

  • Reading a zero delta as no effect. On 7B, one fix cancelled one regression while two other incorrect outputs also changed.
  • Transferring the guardrail unchanged. Tuned to the 3B failure mode, it changed a model that failed differently and introduced a new failure despite a flat aggregate score.
  • Reporting only post-policy accuracy. Without the validated pre-policy result and the changed trials, you cannot see the guardrail's net effect or the behavior underneath it.
  • Treating a guardrail as a free add-on. It's part of a specific system and must be re-measured per model.

Practical guidance

  • Build guardrails from a precise diagnosis — the case study's rule was powerful because it targeted the dominant, well-characterised failure.
  • Always report validated pre-policy and post-policy accuracy; the gap is the guardrail's net accuracy effect, but it does not describe every changed trial.
  • Inspect every intervened case for fixes and regressions, and re-measure per model — never assume a rule transfers.
  • Treat a correction policy as versioned with the system it was tuned to, not a portable upgrade.

Summary

  • The case study's correction policy targeted the diagnosed ranking-for-breakdown confusion and added ~25 points to the 3B model.
  • Transferred unchanged to the 7B model, it touched four trials — one fix, one regression, and two wrong-to-different-wrong changes — so it netted to zero while still changing behavior.
  • This proves a guardrail is a two-sided trade (net of fixes and regressions) tuned to one system — the guardrail-analysis lesson on real data.
  • The aggregate effect is visible in pre-policy versus post-policy accuracy; the hidden movement is visible only in the changed trials.

Knowledge check

The correction policy added 25 points to the 3B configuration. What happened when the same rule was applied unchanged to 7B?

Because the rule fires on a condition (a top_n_by_metric prediction with no ranking cue in the question) that only approximates the actual failure, and it was tuned to the 3B model's failure mode. The 3B model made the ranking-for-breakdown mistake often, so the rule caught and corrected many genuine errors. On 7B, it touched four trials: one fix, one regression, and two wrong-to-different-wrong changes. The fix and regression cancelled, producing zero net accuracy change, but the trial-level movement included one newly broken task. A guardrail encodes assumptions about how one model fails, which is why it must be re-measured per model rather than assumed to transfer.

How was the project able to see both the +25 on 3B and the hidden movement on 7B, and why does that matter?

Because it compares validated pre-policy and post-policy outputs and keeps every intervened trial openable. The aggregate comparison shows a +25-point effect on 3B and a zero-point effect on 7B; the changed trials show that the 7B policy still fixed one case, broke one, and changed two incorrect outputs. Both views matter: the aggregate quantifies net accuracy effect, while the trial accounting exposes behavior that cancellation hides.

Related chapters