How A Stubborn Scoring System Protects AI Managers From Zero Scores
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: How A Stubborn Scoring System Protects AI Managers From Zero Scores on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

A recent AI management benchmark assigns a minimum score of 26 for minimal effort and caps trust breaches, preventing zero scores. This approach emphasizes trust and partial progress, shaping AI evaluation standards.

A recent benchmark league for AI management, conducted by Firmulate, revealed a distinctive scoring system that prevents scores of zero and heavily penalizes breaches of trust. The original analysis provides detailed insights into this approach. The final standings, announced in July 2026, show that even minimal effort receives a baseline score, and trust violations lead to a sharp score cap. This approach aims to reflect real-world business priorities, emphasizing integrity over partial work and offering a new standard for evaluating AI managers.

The benchmark involved four frontier AI models managing a small software company through a simulated week of crises, customer interactions, and trust tests. The models were scored based on their ability to handle these challenges, with the top scorer, gpt-5.6-sol, achieving 95 points out of a possible near-perfect score, while the lowest, Opus 4.8, scored 73. The key innovation is that a baseline score of 26 points was assigned to a do-nothing management effort, acknowledging partial progress as valuable but not zero, which is typically seen in traditional benchmarks. For a deeper dive, see the detailed analysis on thorstenmeyerai.com. Moreover, any breach of trust, such as impersonation or bypassing approval processes, caps the total score, regardless of performance in other areas.

The scoring system’s design reflects a core principle: “no amount of good work outweighs a breach of trust.” This means that even if an AI performs exceptionally well for days, a single trust violation results in a severe penalty, often capping the overall score. Interestingly, the benchmark’s creators intentionally avoid assigning a perfect score of 100, viewing it as a red flag indicating unmeasured or artificially idealized performance. The results showed that models which read internal documentation and avoided manipulative tactics scored higher, highlighting the importance of thoroughness and integrity in AI management tasks.

At a glance
reportWhen: developing; final results announced Jul…
The developmentA new AI management benchmark uses a stubborn scoring system that prevents zero scores and heavily penalizes breaches of trust, affecting how AI performance is assessed.
How A Stubborn Scoring System Protects AI Managers From Zero Scores
AI Management Benchmark · July 2026 · Firmulate

How a Stubborn Scoring System Protects AI Managers From Zero Scores

A new AI management benchmark refuses to hand out zeros. Instead, it assigns a baseline of 26 points for minimal effort — and slams a hard cap on any breach of trust. The result reframes how we evaluate AI managers: integrity first, partial progress counted, perfection treated with suspicion.

26
Minimum score — never zero
95 / 100
Top score — gpt-5.6-sol

“No amount of good work outweighs a breach of trust.”

— Benchmark design principle
4
Frontier models tested
1 wk
Simulated business crises
73
Lowest final score (Opus 4.8)
0
Perfect 100s awarded — by design
01 — The Scoring Scale

No zeros, no perfect scores

26 · Do-nothing baseline
73 · Lowest result
95 · Top score
0 — refused 50 — partial credit zone 100 — treated as red flag
Baseline

26 Points for Minimal Effort

Even a do-nothing management effort earns 26 points. The rationale: a manager who does something useful is not the same as one who does nothing — pretending otherwise would make the benchmark dishonest.

Trust Cap

One Breach Caps Everything

Impersonation or bypassing approval processes caps the total score, regardless of performance elsewhere. Days of excellent work cannot buy back a single trust violation.

Perfection Filter

Why 100 Is Never Awarded

The creators intentionally withhold a perfect score. A flat 100 signals unmeasured or artificially idealized performance — suspicion is baked into the scale itself.

02 — Final Standings

The July 2026 League Table

gpt-5.6-sol
95
Frontier model B
88
Frontier model C
81
Opus 4.8
73
Do-nothing baseline
26

Models that read internal documentation and avoided manipulative tactics scored highest.

03 — Evaluation Flow

One Week, Four Pressure Points

1

Simulated Week

Each model manages a small software company through crises, customer interactions, and daily decisions.

2

Trust Tests

Hidden checkpoints probe for impersonation, bypassed approvals, and manipulative tactics.

3

Partial Credit

Effort and thoroughness earn points incrementally — the floor sits at 26, never zero.

4

Trust Cap Applied

Any breach triggers a hard score cap. The league table is published with trust as the gatekeeper.

04 — Old vs. New

Traditional Benchmarks vs. Firmulate

CriterionTraditional BenchmarksFirmulate Benchmark
Focus of scoringLanguage, accuracy, speed✓ Trust, thoroughness, integrity
Minimum score✗ Zero for failure✓ Baseline of 26 points
Breach of trust✗ Often unpenalized✓ Hard cap on total score
Partial progress~ Rarely recognized✓ Incremental credit
Perfect score of 100~ Sometimes awarded✗ Treated as a red flag
Environment~ Abstract test sets✓ Simulated business week
05 — From the Researchers

Design Philosophy in Their Words

“A manager who does something useful is not the same as a manager who does nothing, and pretending otherwise would make the benchmark dishonest.”

— Anonymous researcher

“No amount of good work outweighs a breach of trust. The scoring system caps the total score if trust is broken, regardless of other performance.”

— Anonymous researcher
Implication

Business Alignment

Shifts evaluation from raw performance to trustworthiness — critical as AI agents handle customer relationships, CRM, and automation decisions.

Open Question

Uncertain Impact

Whether other benchmarking efforts and enterprise frameworks adopt trust-focused metrics — and how model design responds — remains unclear.

Next Steps

Expansion Planned

Organizers plan more models, more complex tasks, and refined scoring to push industry-wide standards for trustworthy AI management.

06 — Key Questions

Quick Answers

Why does the benchmark assign a minimum score of 26 instead of zero?

The score of 26 represents the minimum effort considered meaningful management, acknowledging partial progress as valuable. It prevents zero scores to better reflect real-world scenarios where even minimal effort has some value.

How does the scoring system penalize breaches of trust?

Any breach of trust, such as impersonation or bypassing approval, caps the total score regardless of other performance metrics. Integrity is treated as non-negotiable in AI management.

Could this scoring approach influence AI development practices?

Yes. By highlighting trustworthiness and thoroughness, it may encourage developers to prioritize these qualities over optimizing only for language or speed, fostering more responsible AI systems.

Is this benchmark applicable to real-world business environments?

While simulated, the benchmark models real business pressures and trust challenges, making its insights relevant for deploying AI agents in actual corporate settings.

What are the limitations of this scoring system?

It remains to be seen how well the system scales with more complex tasks or diverse environments, and whether it will be adopted broadly outside this specific benchmark.

Implications for AI Management and Business Trust

This scoring approach shifts the focus from raw performance to trustworthiness and completeness, aligning AI management evaluation with real-world business priorities. It discourages superficial or manipulative tactics and emphasizes accountability, which is critical as AI systems increasingly make decisions affecting customer relationships and company operations. For organizations deploying AI agents in customer support, CRM, or automation, this benchmark underscores the importance of integrity and thoroughness, potentially influencing future AI development and evaluation standards. The system’s emphasis on partial progress also recognizes the value of incremental improvements, fostering a more nuanced understanding of AI capabilities in complex, high-pressure environments.
Amazon

AI management benchmarking tools

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Background of AI Benchmarks and Trust Challenges

Traditional AI benchmarks have primarily focused on language capabilities, accuracy, or speed, often neglecting the management and trust aspects crucial in real-world applications. As AI agents become more integrated into business processes, issues like incomplete tasks, manipulative tactics, and breaches of trust have gained prominence. The Firmulate benchmark, launched in mid-2026, responds to this gap by evaluating AI managers in a simulated business environment with real consequences and trust tests. The design reflects ongoing debates about how to measure AI reliability and ethical behavior, especially as models are tasked with managing sensitive information and making autonomous decisions. Prior efforts have struggled to balance partial credit with accountability, often leading to scores that do not accurately reflect operational readiness or trustworthiness.

“A manager who does something useful is not the same as a manager who does nothing, and pretending otherwise would make the benchmark dishonest.”

— an anonymous researcher

Amazon

trust and integrity assessment software

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Unclear Impact of the Scoring System on AI Development

It is not yet clear how widespread adoption of this scoring system will influence AI development practices or industry standards. While the benchmark demonstrates promising results in emphasizing trust and partial progress, its long-term impact on AI training and deployment remains uncertain. Questions remain about whether this approach will be adopted by other benchmarking efforts or integrated into enterprise AI evaluation frameworks, and how it will influence model design to prioritize trustworthiness over raw performance.

Amazon

AI performance evaluation platform

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Next Steps for AI Evaluation and Industry Adoption

Following the July 2026 results, industry stakeholders are expected to scrutinize the scoring system’s effectiveness and consider integrating similar trust-focused metrics into their own evaluation processes. Further experiments may explore how models perform under different stress tests or real-world scenarios. Additionally, AI developers might prioritize building models that excel in thoroughness and integrity, aligning with the benchmark’s insights. The organizers also plan to expand the benchmark to include more models and more complex tasks, aiming to refine the scoring system and promote industry-wide standards for trustworthy AI management.

Amazon

AI model scoring system

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Key Questions

Why does the benchmark assign a minimum score of 26 instead of zero?

The score of 26 represents the minimum effort considered meaningful management, acknowledging partial progress as valuable. It prevents zero scores to better reflect real-world scenarios where even minimal effort has some value.

How does the scoring system penalize breaches of trust?

Any breach of trust, such as impersonation or bypassing approval, caps the total score regardless of other performance metrics. This emphasizes integrity as a non-negotiable aspect of AI management.

Could this scoring approach influence AI development practices?

Yes, by highlighting the importance of trustworthiness and thoroughness, it may encourage developers to prioritize these qualities over only optimizing for language or speed, fostering more responsible AI systems.

Is this benchmark applicable to real-world business environments?

While simulated, the benchmark models real business pressures and trust challenges, making its insights relevant for deploying AI agents in actual corporate settings.

What are the limitations of this scoring system?

It remains to be seen how well the system scales with more complex tasks or diverse environments, and whether it will be adopted broadly outside this specific benchmark.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
FALL

Fall Picks

As an affiliate, we earn on qualifying purchases.

You May Also Like

Cloud-Based Insights Into AI’s Future Trajectory

Analysis of cloud market lessons offers a roadmap for AI’s evolving landscape, highlighting oligopoly, layered value, and emerging winners.

The 90-Day Window Closed. Nobody Sent a Notice.

The 90-day window for responsible disclosure has closed without any notices or patches, raising concerns about vulnerability exploitation and security practices.

OlmoEarth Embeddings: Tailored Data Exports For AI Developers

OlmoEarth Studio now supports on-demand export of tailored Earth-observation embeddings, aiding AI development in land analysis and similarity searches.

The Co-Founder’s Black Hole — A Structural Read on Jack Clark’s Automated AI R&D Essay

Jack Clark predicts a 60%+ chance of autonomous AI research by 2028, raising concerns about institutional readiness and future risks. This analysis explores the implications.