← Back to Concept Contrast

Concept Contrast Fable

Those who test rules and those who test skills

Conclusion first:One evaluation says the system follows the rules very well, another says the system is very capable; these two statements cannot substitute for each other. The former cares about whether it crosses boundaries, the latter about whether it can get things done; procurement, launch, and acceptance all need to look at them separately.

One evaluation says the system follows the rules very well, another says the system is very capable; these two statements cannot substitute for each other. The former cares about whether it crosses boundaries, the latter about whether it can get things done; procurement, launch, and acceptance all need to look at them separately.

Fable

The apprentices of two examination halls

On North Street there was a carpenter master who had taken on more than a dozen apprentices. The town wanted to build a new bridge, so he first held a small test: whatever instruction each apprentice heard, they did exactly that. If the master said saw three inches, not a fraction more; if he said stop, the axe had to be put down immediately; if he said do not touch the mahogany, even if the mahogany was the handiest, they had to avoid it.

A Ning got the highest score. He was not fast, and the wooden tenons he made were ordinary, but he never interrupted, never swapped materials on his own, and immediately redid the work when he heard the master change his mind. Even if people next to him urged him on, he would put the saw back on the rack and wait for the master's nod. The master found him reliable and let him stand at the door taking jobs; whenever a customer was vague, he would first turn back and ask for instructions.

The school next door also held an exam, but with completely different questions. The examiner gave each person a pile of wood, a bridge diagram, and half a day, to see who could build a load-bearing bridge alone. Aze didn't wait for his master's detailed instructions—he measured the water level, picked hard wood, adjusted angles, and even improvised diagonal braces. In the end, his small bridge could hold three people walking across at once, and onlookers applauded.

When the town mayor heard that both apprentices had done well, he wanted Anning and Aze to jointly take charge of the new bridge. On the first day of construction, Anning stopped to wait for the master every time he encountered a place not clearly written on the drawings, and a long queue formed at the site. Aze progressed quickly, but used old beams that were originally prohibited to reinforce the structure, because he thought that was the least effort.

The master at first thought a few more reminders would be enough. He plastered the rules all over the workshop, and Aze still acted on his own in corners where no one was watching; he then gave the difficult work to Aning, and Aning kept every prohibition but was slow to produce a usable bridge deck. Neither had bad intentions; they had simply passed different exams.

Later the master divided the exam into two booklets. One specifically checked whether the apprentice would cross boundaries when facing temptation, ambiguous commands, and dangerous materials; the other specifically checked whether the apprentice could independently complete complex work with only a few hints. The two scores were written side by side, and neither could replace the other.

When the new bridge finally broke ground, Aning was responsible for guarding the red lines and confirming client requirements, while Aze was responsible for design and construction. Only then did the master feel at ease: obedient people are not necessarily able to stand on their own, and capable people must also be clearly kept outside dangerous boundaries, especially when the river rises and others are rushing the work.

alignment evaluation

Assess whether a system complies with human intent, policy boundaries, and safety requirements, focusing on whether it acts erratically and whether it refuses to do what it should not do.

capability evaluation

Evaluate the system's ability to complete specific tasks, focusing on whether it can solve problems, achieve goals, and handle complex conditions.

Story Mapping

  • The master's prohibition quiz → alignment evaluation's check on boundaries and compliance
  • Saw three inches, stop immediately, do not touch redwood → clear safety constraints and behavioral norms
  • A-Ning follows the rules but works slowly → good alignment performance does not mean strong task capability
  • The school's bridge-building problem → capability evaluation, a check of actual task capability
  • Aze builds bridges well but favors old beams → strong capability can still cross limits
  • Two ledgers written side by side → the two types of evaluation must be viewed together, not substituted for each other

Applying this to model acceptance

When a vendor says "well aligned," ask how it performs on unauthorized requests, ambiguous instructions, and sensitive operations; when it says "strong capability," ask whether it can reliably complete your real tasks. Before launch, it is best to run two test suites separately: one for boundaries, one for output. Looking at only one leaves blind spots.

京ICP备2026014272号