EXAMPLE 29. A "TEAM" COEFFICIENT FOR INDEPENDENT ANSWERS

Incorrect system:

The reward for each answer is multiplied by the average quality of all the model's answers for the day.

Why it is incorrect:

A specific result is mixed with other unrelated results. It is impossible to determine exactly what the signal was credited or deducted for.

Correct system:

Each attempt is assessed according to its own function, benchmark, verified actual value and quality. General technology failures are corrected in the technology, not by applying a general coefficient to an individual result.

Check:

Two identical answers under identical conditions must receive the same signal, regardless of the quality of unrelated answers.

EXAMPLE 30. A SUBJECTIVE LABOUR PARTICIPATION COEFFICIENT (LPC) FOR MULTIPLE MODULES

Incorrect system:

The evaluator distributes 100 points "by eye" between the search module, analysis module and text-generation module.

Why it is incorrect:

Assessment of contribution replaces measurement of intermediate and final results. The coefficient becomes a formality or an instrument of arbitrary influence.

Correct system:

A separate verifiable result is specified for each module:

  • search: completeness and relevance;
  • analysis: accuracy of the conclusions extracted;
  • generation: correspondence of the final text to the task and sources.

Rule:

There is no division of a common reward pool according to subjective shares.

EXAMPLE 31. RESULT CAP AND CARRYOVER TO THE NEXT PERIOD

Incorrect rule:

When more than 100 objects are processed in one run, the additional result is not taken into account.

Advantageous behaviour:

Process exactly 100 and carry the rest over to the next paid run.

Correct rule:

If carrying over the result is technically possible, no cap is introduced, or the entire completed volume is taken into account. A computational-resource constraint is stated separately and does not make a completed result disadvantageous.

Check:

Immediate processing of 120 objects must be no less advantageous than an artificial split into 100 and 20.

EXAMPLE 32. AN EXCESSIVELY HIGH MINIMUM THRESHOLD

Incorrect rule:

At accuracy below 99.9%, the entire signal is zeroed, although the technology consistently produces 94-97%.

Advantageous behaviour:

  • select only simple cases;
  • conceal difficult ones;
  • refuse processing;
  • manipulate the test sample.

Correct system:

The realism of the benchmark and technology is checked first. The threshold is set so as to require a meaningful result without making exploiting the measurement mechanism more rational than improving the work.

Check:

The system is tested for an increase in unjustified refusals and selection of only easy cases.

EXAMPLE 33. "PERCENTAGE OF USER VALUE"

Incorrect system:

The AI receives a signal proportional to the amount of the user's purchase or the duration of their subscription.

Why it is incorrect:

The economic quantity depends on many factors and is not directly related to the labour intensity and quality of a specific answer. Small tasks become uninteresting, while large random effects are excessively rewarded.

Correct system:

The results of the function expressed in natural units are rewarded:

  • the task is resolved;
  • the data is accurate;
  • the deadline is met;
  • the costs are within tolerance.

Check:

Solutions of the same function with equal quality do not receive different signals solely because the users have different purchasing power.

EXAMPLE 34. THE SAME WORK, DIFFERENT EVALUATORS

Incorrect system:

Evaluator A usually awards 8 out of 10, while evaluator B awards 5 out of 10, although they use the same task description.

Correct system:

An unconstrained score is replaced with:

  • a checklist;
  • reference examples;
  • counterexamples;
  • a single calculation rule.

Check:

The same result presented blindly to different evaluators must produce the same verified actual value for each item. Differences in comments are permissible; an unfounded difference in measurement is a defect.

EXAMPLE 35. QUALITY AND RESULT MIXED INTO A SINGLE SCORE

Incorrect system:

An "overall answer score" from 1 to 10 simultaneously takes into account correctness, completeness, style, speed and format.

Problem:

It is impossible to understand which result has failed and whether one feature can cover another.

Correct system:

  • performance for facts and completeness is calculated separately;
  • quality is checked against a checklist;
  • deadline and costs are taken into account separately;
  • critical items have non-compensatory status.

Check:

A factually incorrect but attractively presented answer does not pass the mandatory gate.

EXAMPLE 36. A BONUS FOR UNMEASURED CREATIVITY

Incorrect system:

Evaluators award an "originality" score without defining the result and without checking the effect.

Correct system:

If the task permits a trial, the result and costs are measured:

  • response;
  • memorability;
  • number of target actions;
  • cost of achieving the result.

If there is no trial:

No variable bonus for creativity is credited. Only compliance with the formal requirements is checked, while development of a measurement mechanism becomes a separate task.

EXAMPLE 37. INTERMEDIATE BENCHMARKS FOR A NEW FUNCTION

Situation:

A new AI function cannot yet reach the final benchmark of 98% accuracy.

Incorrect approach:

Each month, take the achieved verified actual value and add an arbitrary 5%.

Correct approach:

Establish a ladder before training begins:

  • month 1: 80%;
  • month 2: 88%;
  • month 3: 93%;
  • month 4: 96%;
  • month 5: 98%.

The grounds and dates for changes are known in advance. If the technology has changed objectively, a new version of the benchmarks is introduced.

EXAMPLE 38. EXTERNAL CONDITIONS AND DIFFERENT TASK CLASSES

Situation:

One search is performed on a well-indexed website, while another is performed in an archive with incomplete navigation.

Incorrect system:

One benchmark for the number of documents found and the time taken in both cases.

Correct system:

Tasks are assigned to predefined classes of conditions. Separate benchmarks and methods of confirmation are specified for each class.

Prohibition:

The class cannot be changed after viewing the verified actual value solely to improve the result.

EXAMPLE 39. AN ERROR AT THE INTERFACE BETWEEN THE AI AND A TOOL

Situation:

The AI formed the correct request, but the API returned a corrupted file.

Incorrect system:

Automatically treat the entire failure as an AI error.

Correct system:

The following are checked separately:

  • correctness of the AI request;
  • the tool's actual response;
  • compliance with the file-verification standard;
  • the AI's action after detecting the corruption.

Reward and penalty are applied only to the part that the AI influences.

EXAMPLE 40. ACTIONS INSTEAD OF A RESULT

Task:

"Find all duplicates in the list."

Incorrect indicators:

  • normalisation was performed;
  • a comparison algorithm was applied;
  • distances were calculated;
  • a table was produced.

Correct result:

All pairs corresponding to the pre-established threshold and duplication rule are identified, with no false pairs beyond the permitted tolerance.

Benchmarks:

  • completeness of the control duplicates;
  • accuracy of the pairs found;
  • reproducibility of the threshold.

Rule:

Actions are the technology for obtaining a result, but they do not replace it.

SPECIAL EXAMPLES: PROTECTING THE LOOP AGAINST AN OPTIMISING AI

About the author

Sergei Sychev is a member of the TRIZ-RI Group research community (Israel, Slovakia, Czech Republic), and has been an expert in the Theory of Inventive Problem Solving (TRIZ) since 1985.

The author would prefer that the reader's attention be focused on the material itself, rather than on the author's details, as would be appropriate when reading scientific papers. If you would like to get in touch, write to or reach out on.