Evaluate with evidence

Use these measures: Evaluate downstream task quality, perplexity where appropriate, long-context retrieval, latency, memory, throughput, and performance by language or tokenization pattern. Mechanistic probes are evidence about behavior, not proof of human-like concepts.

An evaluation set should represent the actual decision, including easy cases, common cases, rare costly cases, and adversarial or malformed inputs. Freeze a test set before tuning. If examples repeatedly influence prompt, threshold, or architecture choices, move them into a development set and obtain a fresh test set. Report sample count and uncertainty; a 95% score on 20 examples means only one observed miss and says little about rare failures.

Pair offline quality with online operations. A component can score well offline and fail under concurrency, stale data, changed users, or dependency outages. Slice results by relevant dimensions rather than trusting one average. Always compare with a simple baseline: deterministic rules, keyword search, a smaller model, or the current human workflow.

In “The trophy did not fit in the suitcase because it was too large,” the representation for “it” can weigh “trophy” more than “suitcase.” Multiple layers refine that relationship before the next-token prediction.

The worked number is an attention score is QKᵀ/√d; a length n sequence creates an n×n score matrix, so doubling n from 4,000 to 8,000 creates about four times as many pairwise scores. State the unit and denominator whenever you report it. A percentage without a denominator can conceal a tiny sample; a latency without a percentile can conceal slow users; a similarity score without a labeled task can conceal irrelevant neighbors. Compare the observed value with a threshold chosen before seeing the final test result.

Now test the tempting shortcut. Suppose the team optimizes only the most visible metric. The result may look better while the system becomes less trustworthy. The reason is concrete: Attention connects distant tokens and parallelizes training, but standard attention cost grows roughly with the square of sequence length. More parameters increase capacity and compute. Tokenization handles open vocabulary efficiently but splits words unevenly across languages and domains. This is why the decision record must include both the intended gain and the tolerated regression. If the tolerated regression is unknown, the change is not ready for a consequential workflow.

• Prefer a measured baseline over a persuasive demo.
• Keep versions, inputs, and thresholds reproducible.
• Separate syntactic success from semantic correctness and authorization.
• Escalate or abstain when evidence falls outside the contract.
• Re-evaluate when data, traffic, models, providers, or user goals change.

These rules turn the topic into an engineering decision rather than a slogan. They also make disagreement productive: another person can challenge the assumptions, rerun the evaluation, and reach a documented conclusion.

For every percentage, report numerator, denominator, and slice. For every latency, report workload and percentile. For every human rating, define the rubric and check agreement on a shared subset. Compare paired outputs on the same examples when possible; this reduces noise from case difficulty. Investigate regressions, not only the aggregate win. Finally, reserve a fresh set for confirmation after tuning. If the candidate misses a hard safety, authorization, or correctness threshold, a higher average score elsewhere does not compensate—the candidate fails the gate.