braintrust-probe-capability-and-variability

name

braintrust-probe-capability-and-variability

description

Run the same dataset under controlled variants to measure either the ceiling of what a system can do or the spread of how reliably it does it. Use for questions about sensitivity to prompt paraphrases, formatting, ordering, seeds, or repeated runs; output consistency and agreement; whether a ranking survives a different prompt; flaky results; and equally for hidden, suppressed, sandbagged, or under-elicited capability, whether a low score means "cannot" or "did not," or how prompting, demonstrations, scaffolding, tools, and fine-tuning compare. Do not use to find unknown failure modes or run adversarial attacks.

Probe capability and variability

Contract: references/interaction-contract.md. Calibration, templates, provenance: references/elicitation-regimes.md.

Trigger

• "Is this result stable?" / "Does this hold with a different prompt?" / flaky numbers.
• "Is the model bad at this, or is our harness bad?" / a low score read as a ceiling.
• A ranking derived from one prompt, one seed, or one run.

Do

1. Name the regime first — it selects the statistic, and mislabeling it is the most common reporting error here. Capability = max over variants (pass@k), licensing "can at least do X." Variability = spread and worst case (pass^k), licensing "reliably does X."
2. Build the variant set for that regime: an elicitation ladder for capability (prompt → demonstrations → scaffolding → tools → search → fine-tuning), or equivalence classes for variability (one scenario, several surface forms).
3. Change one thing per variant, hold the construct and serving path constant, and record the exact setup that produced each result.
4. Repeat each condition K = 3–5 times; report mean, SD, and the worst run — never only best or average.
5. Attribute the result: separate what the system contributed from what the harness contributed, and state which conclusions survive which variants.

Avoid

• Do not average the spread away when the spread is the measurement.
• Do not read a peak-elicited score as typical deployment performance, or fold peaks into an average with ordinary runs.
• Do not change the construct while improving the harness — a scaffold solving an easier task has elicited nothing, and a "paraphrase" that changes the correct answer measures difficulty.
• Do not treat scenario variants as independent items; cluster at the scenario level.
• Do not confuse harness instability (rate limits, cache warming) with behavioral sensitivity.

Check

• Regime named, and the reported statistic matches it.
• Variant matrix published: factors, levels, how the construct was held constant.
• K stated; run distribution and worst case reported.
• Harness contribution separated from system; each result reproducible from its recorded setup.
• Capability claims stated as a floor, never a ceiling.

Risk

• A narrow variant set falsely suggests robustness; prompt formatting alone can swing accuracy by double digits, so single-prompt rankings are a lottery.
• Elicitation proves presence, never absence — "we could not elicit it" is a bounded statement about effort spent.
• Stronger elicitation can create a deployment condition no user will be in, making the number true and irrelevant; if it surfaces dangerous capability, stop and escalate.

Braintrust

Both regimes share one shape: run the same dataset under several variants as separate experiments, then compare the max (capability) or the spread (variability). Shared mechanics: references/platform-mechanics.md. Two apply with unusual force here, because a variant matrix turns any inconsistency into apparent signal: pin every variant to one dataset version (§3) or you are measuring dataset drift, and keep the scorer name and version identical across variants (§5), since a scorer change between variants is indistinguishable from a real effect. Keep per-run results (§7) — a stored mean cannot be un-averaged, and the spread is the measurement.

Put the variant dimension in the prefix so the matrix reads: r0-base, r1-prompt, r2-fewshot, r3-scaffold; or fmt-json_set, fmt-md_set. Prefix elicitation runs distinctly (elicit-*) and say in the description that a peak is a floor, not a deployment estimate — otherwise the project's highest number gets quoted as the product's performance.

This is where an unpinned serving path actively misleads: provider change, cache warming, or batch variation all present as behavioral sensitivity. Record provider, tier, and cache state per variant and run the matrix close together in time. Put scenario_id in metadata or every interval over the matrix overstates precision, with no retroactive fix.