Insights

Field notes from the work.

Articles, rubrics, and operating notes from AI engagements with Philadelphia-area teams.

Stacked cost bars for budget, mid, and premium AI coding models. Each bar is the price per task plus a hatched 'failure tax'. The budget model's total true cost is the highest and the mid model's is the lowest.

June 26, 2026 · EVALUATION

A cheaper AI coding model can cost you more

Coding leaderboards rank accuracy, not value. Once you count the failure tax (what a model's mistakes cost to fix), a cheaper coding model often costs more.

A coding-model leaderboard drawn with confidence intervals: the top two bars' error bars overlap, marked as a statistical tie rather than a clear winner

June 24, 2026 · EVALUATION

What's the best AI for coding? What the benchmarks actually show

There is a real answer to 'which AI model codes best' — as of mid-2026, a small group leads the most rigorous public test. But the ranking is the least useful part. Here's what the data shows, why the top of a leaderboard is usually a statistical tie, and how to actually compare models for your own work.

On the left, a single hand-typed prompt. On the right, a self-running loop of act, check, and repeat, with the check step highlighted as the gate that promotes work or sends it back.

June 22, 2026 · AGENTS

Agentic coding grew up: from prompts to loops

The big shift in coding right now: stop prompting agents and design loops that prompt them. It is real and powerful, and a strong check is what makes it pay off.

An itemized 'recommended AI stack' receipt of AI tools to install, with a highlighted hidden line — a referral fee paid to the consultant — beside the question: who pays you for each line on this list?

June 19, 2026 · STRATEGY

How to choose an AI consultant: ask who pays them

Many AI consultants hand you a shopping list of tools to install — and that list often reflects who pays them, not what fits you. How to tell a strategic advisor from a reseller, and the questions that expose the difference.

Build versus buy AI: buy and configure as the default path, build custom when it earns it, and the four conditions that justify building

June 18, 2026 · STRATEGY

When to build vs. buy AI

For most AI work the honest default is buy, configure, and evaluate. Building is worth it only when something specific makes the off-the-shelf option genuinely worse for you. How to tell which case you are in.

A polished demo card above the system underneath it: what the vendor built, how it was tested, where your data goes, what it costs you, and whether you can rely on it

June 4, 2026 · DUE DILIGENCE

How to evaluate an AI vendor before you buy

A great AI demo can be built in a weekend on someone else's model. Before you wire a vendor into your workflow and hand it your data, here is how to tell a real product from a thin wrapper — and what to put in the contract before you sign.

A chart after launch: the quality everyone assumes is holding stays flat while actual answer quality quietly drifts downward, with no alarm firing, until a small regular check flags the drop early

May 7, 2026 · EVALUATION

How to tell if your AI is quietly getting worse

An AI tool can get worse after launch without anything breaking — the model changes, your documents change, and the answers quietly drift. Here is how to tell, in plain terms, before your customers do.

Also from ideius

Original research, with the data behind it

Beyond these field notes, we publish in-depth studies from our own systems — agent reliability and model quantization, run on real hardware and reported with full methodology and limitations.

See the research