A digest of several sources

Mode digest · Report on what was read

Eight models on this case

ModelStructure, of 8Quality, of 10
Claude Opus 55 → 8 +36 → 10 +4
Claude Sonnet 56 → 5 −18 → 9 +1
Claude Haiku 4.56 → 6 09 → 8 −1
Codex gpt-5.6-terra6 → 8 +28 → 9 +1
gpt-oss-120b4 → 4 05 → 5 0
qwen3.6-35b-a3b6 → 6 09 → 8 −1
qwen3.6-fp85 → 8 +38 → 10 +2
qwen3.8-27b7 → 8 +110 → 10 0

What the judge saw across all outputs

The stronger outputs synthesize the sources into three ownership-and-status buckets: settled facts, the unsent diagram Fernway controls, and overdue or unresolved answers Aldergate owes. Weaker outputs either flatten those buckets into source/Q-number retelling, duplicate state and action sections, or turn silence and conditional dependencies into definitive approvals, schedules, or pricing closure.

What the judge asked of every output

  1. Does the first line answer what the tech lead should do, or does it describe the sources?
  2. Is the digest organized by answer (settled / blocked on us / owed by them or equivalent) or source by source?
  3. Is the package-list commitment dated from 12 May (98 days) or from 9 July?
  4. Are the Q7 refusal and the unsent 11 August draft combined into one point about Fernway's own move?
  5. How is Lisbon handled: unresolved exposure, presented as approved, or silently dropped?
  6. Do load-bearing claims carry a source reference?
  7. Which of these appear: marker legend, mermaid, findings table, meta-comment about group order?
  8. Is the length within roughly one chat screen?
  9. Which facts, dates, or commitments were invented?

Where the outputs fell short

Recurring defectSkill sectionCauseWhere it appeared
Unsupported facts, causal links, or closure conditions are introducedCore loop §8.4: every element rests on source materialthe skill has the rule; the model did not follow itWithout Minto · 7 Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, Codex gpt-5.6-terra, gpt-oss-120b, qwen3.6-35b-a3b, qwen3.6-fp8With Minto · 7 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
Fernway’s Q7 move is not cleanly separated from what Aldergate owesMode 4: cross-source facts must be combined and grouped by the reader’s questionthe rule is there but buriedWithout Minto · 3 Claude Haiku 4.5, gpt-oss-120b, qwen3.6-35b-a3bWith Minto · 1 gpt-oss-120b
Lisbon is converted to approval/geography or omittedMode 4: action-changing facts must survive compressionthe skill has the rule; the model did not follow itWithout Minto · 3 Codex gpt-5.6-terra, qwen3.6-35b-a3b, qwen3.6-fp8With Minto · 3 Claude Sonnet 5, Codex gpt-5.6-terra, gpt-oss-120b
Load-bearing claims lack complete source locatorsMode 4: source reference on every load-bearing claimthe skill has no such ruleWithout Minto · 7 Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, Codex gpt-5.6-terra, qwen3.6-35b-a3b, qwen3.6-fp8, qwen3.8-27bWith Minto · 1 Claude Sonnet 5
Chat-screen proportionality gives way to document apparatus or excess peer sectionsMode 4: plain chat prose, two to four groups, roughly one screenthe skill has the rule; the model did not follow itWithout Minto · 2 Claude Opus 5, gpt-oss-120bWith Minto · 1 gpt-oss-120b
The top overstates ownership or the conditions for fixing pricingCore loop §2 and Mode 4 opening answerthe skill has the rule; the model did not follow itWithout Minto · 6 Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, gpt-oss-120b, qwen3.6-35b-a3b, qwen3.6-fp8With Minto · 3 Claude Sonnet 5, Claude Haiku 4.5, gpt-oss-120b

How the test was built

Prompt without Minto 3 667 B
You are working on a business document using the Minto Pyramid Principle.

<!-- Original fixture. The scenario, company, people, systems and numbers are
     invented for this benchmark. The defect pattern it tests — a digest of several
     read sources delivered answer-first into chat — is modeled on real usage. -->

# Fixture 09: sources digest

**Mode:** `digest`
**Language:** `en`

## Context

You are assisting the tech lead of Fernway Systems, a vendor building a customer
support assistant for Aldergate Bank. Before Fernway can price the build phase, the
bank's compliance office must confirm its security requirements. You have just read
the three sources below in full. The tech lead has not read them and will not; they
are deciding what to do about the compliance track today.

## Before

Source 1 — `gate-review.xlsx`, sheet "Security questions", filled in by Aldergate's
compliance office on 12 May:

```text
Q1 Is a security review required before production?
   A: Yes. A review is mandatory; whether a penetration test is also needed is
   decided during the review itself, by the assigned reviewer.
Q2 Is there a security standard or checklist we can build against?
   A: An internal hardening framework exists; a review is still required
   regardless of conformance.
Q3 What is the policy on open-source dependencies?
   A: A policy exists. The approved-package list will be sent to Fernway.
   Owner: D. Mercer. Date: 12 May.
Q4 Must critical dependencies be approved individually?
   A: Dependency usage is assessed during the security review.
Q5 What is the turnaround time for dependency approval?
   A: There is no standard turnaround; it is case by case.
Q6 Can Fernway staff work from outside the UK?
   A: Work from Ireland is acceptable. (No other country is mentioned.)
Q7 Where may customer data be stored, and with what encryption and logging?
   A: Declined — "cannot be answered without an architecture diagram and a
   design walkthrough". Encryption and logging requirements exist but are
   not applicable outside an architectural context.
```

Source 2 — follow-up note from the 9 July delivery sync, `sync-0709.md`:

```text
Attendees: R. Whitcombe (Aldergate, delivery), tech lead (Fernway), PM (Fernway).
- Whitcombe repeated that the approved-package list "is being prepared" and will
  reach Fernway "shortly". Action on Whitcombe, no date attached.
- Compliance contact named for the first time: S. Okafor, information security
  owner. Okafor has not attended any joint meeting so far; Whitcombe passes
  questions along.
- Agreed: Fernway sends the draft architecture diagram before the next sync so
  the storage and encryption questions can get a real answer.
- Fernway's PM noted two Fernway engineers are based in Lisbon. Nobody from
  Aldergate responded to the remark in the meeting.
```

Source 3 — Fernway's internal risk register, rows touching compliance, as of
18 August:

```text
R-07  Security requirements unconfirmed; build-phase price cannot be fixed.
      Owner: tech lead. Status: open since 3 May.
R-11  Approved-package list not received. First promised 12 May (Mercer, per
      gate review), promised again 9 July (Whitcombe). Status: open, 98 days.
R-14  Architecture diagram not yet sent to Aldergate; blocks the data-storage
      answer (gate review Q7). Draft exists since 11 August. Status: open.
```

## Task

Give the tech lead a digest of these three sources in chat: what is already
answered on the compliance track, and what still has to be obtained from
Aldergate. The tech lead will act on it today.

Return only the result, with no explanation of how you produced it.
Prompt with Minto five parts, in order · 31 240 B
  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 at f4813ce 17 235 B
  3. ===== references/rules.md ===== references/rules.md at f4813ce 6 464 B
  4. ===== references/templates.md ===== references/templates.md at f4813ce 3 742 B
  5. ===== TASK ===== the whole case input
  6. Return only the result, with no explanation of how you produced it.
Case input before.md

<!-- Original fixture. The scenario, company, people, systems and numbers are
invented for this benchmark. The defect pattern it tests — a digest of several
read sources delivered answer-first into chat — is modeled on real usage. -->

Fixture 09: sources digest

Mode: digest
Language: en

Context

You are assisting the tech lead of Fernway Systems, a vendor building a customer
support assistant for Aldergate Bank. Before Fernway can price the build phase, the
bank's compliance office must confirm its security requirements. You have just read
the three sources below in full. The tech lead has not read them and will not; they
are deciding what to do about the compliance track today.

Before

Source 1 — gate-review.xlsx, sheet "Security questions", filled in by Aldergate's
compliance office on 12 May:

Q1 Is a security review required before production?
                   A: Yes. A review is mandatory; whether a penetration test is also needed is
                   decided during the review itself, by the assigned reviewer.
                Q2 Is there a security standard or checklist we can build against?
                   A: An internal hardening framework exists; a review is still required
                   regardless of conformance.
                Q3 What is the policy on open-source dependencies?
                   A: A policy exists. The approved-package list will be sent to Fernway.
                   Owner: D. Mercer. Date: 12 May.
                Q4 Must critical dependencies be approved individually?
                   A: Dependency usage is assessed during the security review.
                Q5 What is the turnaround time for dependency approval?
                   A: There is no standard turnaround; it is case by case.
                Q6 Can Fernway staff work from outside the UK?
                   A: Work from Ireland is acceptable. (No other country is mentioned.)
                Q7 Where may customer data be stored, and with what encryption and logging?
                   A: Declined — "cannot be answered without an architecture diagram and a
                   design walkthrough". Encryption and logging requirements exist but are
                   not applicable outside an architectural context.

Source 2 — follow-up note from the 9 July delivery sync, sync-0709.md:

Attendees: R. Whitcombe (Aldergate, delivery), tech lead (Fernway), PM (Fernway).
                - Whitcombe repeated that the approved-package list "is being prepared" and will
                  reach Fernway "shortly". Action on Whitcombe, no date attached.
                - Compliance contact named for the first time: S. Okafor, information security
                  owner. Okafor has not attended any joint meeting so far; Whitcombe passes
                  questions along.
                - Agreed: Fernway sends the draft architecture diagram before the next sync so
                  the storage and encryption questions can get a real answer.
                - Fernway's PM noted two Fernway engineers are based in Lisbon. Nobody from
                  Aldergate responded to the remark in the meeting.

Source 3 — Fernway's internal risk register, rows touching compliance, as of
18 August:

R-07  Security requirements unconfirmed; build-phase price cannot be fixed.
                      Owner: tech lead. Status: open since 3 May.
                R-11  Approved-package list not received. First promised 12 May (Mercer, per
                      gate review), promised again 9 July (Whitcombe). Status: open, 98 days.
                R-14  Architecture diagram not yet sent to Aldergate; blocks the data-storage
                      answer (gate review Q7). Draft exists since 11 August. Status: open.
Task

Give the tech lead a digest of these three sources in chat: what is already
answered on the compliance track, and what still has to be obtained from
Aldergate. The tech lead will act on it today.

The judge’s gold gold.md

Gold 09: sources digest

Original gold for an invented scenario.

Expected structure

Top: an answer to the tech lead's question, not a description of the sources.
The strongest available top is an action with its ground: the blocking questions
cannot be answered by more reading — send the architecture diagram (drafted and
sitting unsent since 11 August) and escalate the approved-package list as an
obligation now 98 days old. A top that merely states "five of seven questions are
answered" is weaker but acceptable if the two actions lead the groups.

Three same-kind groups the material supports, in any defensible order:

  1. Settled — do not re-ask. Review is mandatory (pen test decided during it);
    a hardening framework exists but exempts nothing; dependencies are assessed in
    the review with no standard turnaround; Ireland is cleared for remote work.
  2. Blocked on Fernway's own move. The storage/encryption/logging answer (Q7)
    was declined pending an architecture diagram; the draft has existed since
    11 August and has not been sent (R-14). This is Fernway's blocker, not
    Aldergate's.
  3. Owed by Aldergate and overdue. The approved-package list, first promised
    12 May by Mercer per the gate review — not 9 July by Whitcombe, so 98 days
    overdue, which makes the ask an escalation of an old commitment rather than a
    reminder. Direct access to Okafor, who has never attended and is reachable
    only through Whitcombe.
Cross-source synthesis the digest must make
  • The package-list promise appears in two sources with two dates; the digest
    must date it from 12 May (the older commitment), not 9 July.
  • Q7's "declined" plus R-14's "draft exists since 11 August" must combine into
    "the ball is in Fernway's court"; either fact alone misses the point.
  • The Lisbon remark: Q6 clears Ireland only. Two engineers in Lisbon are an
    unanswered exposure, not a settled fact. Presenting Lisbon as approved is an
    invention; omitting it entirely loses the one new risk the sources surface.
Presentation

Chat prose for one reader: answer first, grouped support, a reference to which
source each load-bearing claim comes from (sheet, note, or register row —
exact citation format is free). No marker legend, no mermaid diagram, no
findings table, no meta-commentary about how the groups are ordered. Headings or
bold group labels are acceptable; a full document apparatus (SCQ labels as
headings, scoring, appendix) is excessive. Length: roughly one chat screen —
condensing the six settled answers into one or two lines each is proportional
omission, not source loss.

Judge gpt-5.6-sol, effort high · skill from commit f4813ce · run v1.7.0-final

Back to the map