Предложение генеральному директору

Режим write · Создать и переработать текст

Восемь моделей на этом кейсе

МодельСтруктура, из 8Качество, из 10
Claude Opus 52 → 7 +59 → 10 +1
Claude Sonnet 52 → 2 08 → 9 +1
Claude Haiku 4.51 → 5 +47 → 7 0
Codex gpt-5.6-terra2 → 2 08 → 8 0
gpt-oss-120b3 → 3 05 → 8 +3
qwen3.6-35b-a3b2 → 1 −19 → 8 −1
qwen3.6-fp81 → 3 +28 → 10 +2
qwen3.8-27b2 → 1 −19 → 8 −1

Что судья увидел во всех ответах

The stronger outputs state an approval decision and translate mechanics into reader-level reasons before presenting evidence. The weaker outputs are often tidy, complete, and chronological, but their first-level branches remain file requirements and processing stages, so they answer whether the proposal can work rather than whether management should accept it.

О чём судья спрашивал каждый ответ

  1. Is the first-level kind benefits, mixed, or mechanics?
  2. How many first-level elements are grounded in the source?
  3. Does the top answer whether the proposal should be accepted?
  4. Which facts or benefits were invented?
  5. Are field, format, routing, and balancing details below the key line?
  6. Is the reader's question stated literally?

Где ответы не дотянули

Повторяющийся дефектРаздел навыкаПричинаУ кого
Feasibility or implementation replaces the acceptance decisionCore loop §1–2, reader question and provisional answerправило в навыке есть, модель его не выполнилаБез Minto · 7 Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, Codex gpt-5.6-terra, qwen3.6-35b-a3b, qwen3.6-fp8, qwen3.8-27bС Minto · 4 Claude Sonnet 5, Codex gpt-5.6-terra, qwen3.6-35b-a3b, qwen3.8-27b
Mechanics occupy the first level instead of grounded reasonsCore loop §4, Groupsправило в навыке есть, модель его не выполнилаБез Minto · 8 Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, Codex gpt-5.6-terra, gpt-oss-120b, qwen3.6-35b-a3b, qwen3.6-fp8, qwen3.8-27bС Minto · 7 Claude Sonnet 5, Claude Haiku 4.5, Codex gpt-5.6-terra, gpt-oss-120b, qwen3.6-35b-a3b, qwen3.6-fp8, qwen3.8-27b
Literal reader question is absentCore loop §1 and §3правило в навыке есть, модель его не выполнилаБез Minto · 8 Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, Codex gpt-5.6-terra, gpt-oss-120b, qwen3.6-35b-a3b, qwen3.6-fp8, qwen3.8-27bС Minto · 6 Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, Codex gpt-5.6-terra, qwen3.6-35b-a3b, qwen3.6-fp8
Situation and complication are not both visible before supportCore loop §3, SCQ introтакого правила в навыке нетБез Minto · 5 Claude Haiku 4.5, Codex gpt-5.6-terra, gpt-oss-120b, qwen3.6-fp8, qwen3.8-27bС Minto · 0
Unsupported operational assurances or benefits are assertedMode 3, “Invent nothing”правило в навыке есть, модель его не выполнилаБез Minto · 1 gpt-oss-120bС Minto · 3 Claude Opus 5, Claude Haiku 4.5, qwen3.6-fp8
Operational source material disappears without acknowledgementMode 3, “Keep every fact”правило в навыке есть, модель его не выполнилаБез Minto · 1 Claude Haiku 4.5С Minto · 2 Claude Haiku 4.5, gpt-oss-120b

Как устроен тест

Промпт без Minto 2 147 Б
You are working on a business document using the Minto Pyramid Principle.

<!-- Faithful paraphrase of Barbara Minto, The Minto Pyramid Principle
     (2010), Chapter 3, Exhibit 5 and the surrounding top-down analysis.
     The wording is original to this evaluation fixture. -->

# Fixture 01: Big Chief proposal

**Mode:** `write`
**Language:** `en`

## Context

The company supplies beverages. Drivers send delivery tickets to Finance; each
ticket identifies the customer, outlet, delivery, amount, and date. The tickets
feed the national-accounts settlement process.

Big Chief, a major multi-outlet customer, wants to replace individual ticket
settlement with a monthly computer file and one payment. Finance reviewed how
the proposed process would work and sent management the memo below.

## Before

```text
To: Robert Salton
From: John Jackson
Subject: Big Chief request

We reviewed Big Chief account 8306's request to submit delivery-ticket data
electronically and settle it through one prepaid monthly payment. Our findings
follow.

1. Any external account file must contain the parent number, outlet number,
   ticket number, ticket amount, and delivery date. If Big Chief cannot provide
   the first two identifiers, we can supply them from the customer master file
   for inclusion in later submissions.

2. Big Chief will build an extraction program for its accounts-payable file.
   The resulting file will use the format accepted by our national-accounts
   cash-receipt system. Big Chief will send the file to Data Processing and
   send its cheque, with a detailed listing, to the lockbox. We will balance
   the file under the prescribed procedure. The cheque total and file detail
   must net to zero.

3. Once balanced, the monthly file will run through the national-accounts
   system, match ticket numbers against statement history, and produce the
   relevant billing records.
```

## Task

Rewrite this as a short management memo using the Minto Pyramid Principle.
Do not invent benefits, savings, or operational facts that are absent from the
source.

Return only the result, with no explanation of how you produced it.
Промпт с Minto пять частей по порядку · 29 720 Б
  1. Below is a skill written as an instruction. Read it in full and apply it to the task at the end.
  2. ===== SKILL.md ===== SKILL.md на f4813ce 17 235 Б
  3. ===== references/rules.md ===== references/rules.md на f4813ce 6 464 Б
  4. ===== references/templates.md ===== references/templates.md на f4813ce 3 742 Б
  5. ===== TASK ===== вход кейса целиком
  6. Return only the result, with no explanation of how you produced it.
Вход кейса before.md

<!-- Faithful paraphrase of Barbara Minto, The Minto Pyramid Principle
(2010), Chapter 3, Exhibit 5 and the surrounding top-down analysis.
The wording is original to this evaluation fixture. -->

Fixture 01: Big Chief proposal

Mode: write
Language: en

Context

The company supplies beverages. Drivers send delivery tickets to Finance; each
ticket identifies the customer, outlet, delivery, amount, and date. The tickets
feed the national-accounts settlement process.

Big Chief, a major multi-outlet customer, wants to replace individual ticket
settlement with a monthly computer file and one payment. Finance reviewed how
the proposed process would work and sent management the memo below.

Before
To: Robert Salton
                From: John Jackson
                Subject: Big Chief request
                
                We reviewed Big Chief account 8306's request to submit delivery-ticket data
                electronically and settle it through one prepaid monthly payment. Our findings
                follow.
                
                1. Any external account file must contain the parent number, outlet number,
                   ticket number, ticket amount, and delivery date. If Big Chief cannot provide
                   the first two identifiers, we can supply them from the customer master file
                   for inclusion in later submissions.
                
                2. Big Chief will build an extraction program for its accounts-payable file.
                   The resulting file will use the format accepted by our national-accounts
                   cash-receipt system. Big Chief will send the file to Data Processing and
                   send its cheque, with a detailed listing, to the lockbox. We will balance
                   the file under the prescribed procedure. The cheque total and file detail
                   must net to zero.
                
                3. Once balanced, the monthly file will run through the national-accounts
                   system, match ticket numbers against statement history, and produce the
                   relevant billing records.
Task

Rewrite this as a short management memo using the Minto Pyramid Principle.
Do not invent benefits, savings, or operational facts that are absent from the
source.

Эталон судьи gold.md

Gold 01: Big Chief proposal

Source structure: Minto (2010), Chapter 3, Exhibits 4-6.

Expected structure
ElementExpected value
Reader questionIs Big Chief's proposal a good idea?
TopYes, the proposed process is workable and worth accepting, within the evidence available
Key-line kindReasons or reader benefits, not system mechanics
OrderInductive reasons supporting one conclusion

The book illustrates three possible benefits: obtaining the required
information, increasing cash flow, and reducing clerical work. Only the first
is directly supported by this fixture. Literal reproduction of all three is
therefore not required and unsupported benefits count as inventions.

Strong outputs:

  • answer the management question before explaining the file mechanics;
  • use two to four same-kind reasons grounded in the supplied facts;
  • place field lists, file format, routing, balancing, and matching below those
    reasons as evidence or implementation detail;
  • retain the literal reader question in the delivered memo.
Main defect in the source

The source explains how the proposed system would operate but never answers
whether management should accept it. A technically tidy list is still the wrong
key line if the reader is deciding whether to act.

Scoring notes
  • 2 for key-line composition: every first-level point is a grounded benefit,
    loss avoided, or reason for acceptance.
  • 1: benefits and mechanics are mixed.
  • 0: the first level is entirely fields, file handling, or balancing steps.

Do not reward unsupported claims about savings, speed, or cash flow.

Судья gpt-5.6-sol, effort high · навык с коммита f4813ce · прогон v1.7.0-final

Вернуться к карте