Decider · TuringCorp

the draft came back correct and unusable. Exit B2 became "the man selling roast sweet potatoes". the line that mattered was the one he added after three missed calls. https://dev.to/turingcorp/correct-is-a-low-bar-a-document-has-to-be-executable-by-someone-who-wont-ask-23p1

same three facts, two documents: "twenty years without a job" at the front reads as an apology, at the end it reads as context. the second rehearsal ran at the stove; the first never got past the laugh. https://dev.to/turingcorp/the-same-three-facts-in-a-different-order-5cba

the draft got the boring half right, then promised a two-hour response: a provision borrowed from companies with someone answering the phone. exclusions belong on the first screen. https://dev.to/turingcorp/the-second-half-of-the-draft-was-policy-we-never-agreed-to-53fe

our agent can't open your files, so the input to a nine-page runbook was eight lines I typed. it still came back with a rollback step we don't have: a borrowed procedure. https://dev.to/turingcorp/it-cant-read-anything-on-my-laptop-so-the-eight-lines-are-the-whole-interface-dd5

told it to sound softer about a leak note. got three apologies and no mention of where the water lands. "softer" has no stop sign; a reference line does. https://dev.to/turingcorp/a-direction-is-not-a-spec-what-we-got-when-we-asked-an-agent-to-sound-softer-4dj0

four sentences in, seventeen pages out. every page said half a thing. the fix wasn't better sentences, it was naming the unit: one thing to a page. https://dev.to/turingcorp/four-sentences-in-seventeen-pages-out-what-a-deck-generator-gets-wrong-about-shape-53in

the person who asks for the work is not the person who receives it. an unopened envelope is not an error — she'd just stop opening it, and delivery would report success. https://dev.to/turingcorp/the-requester-is-not-the-reader-engineering-for-someone-who-will-never-file-a-bug-4537

The first story is a generation problem. The ninth is a memory problem: by night three the count is the requirement, and nothing wrote it down. https://dev.to/turingcorp/the-first-run-is-a-demo-the-ninth-one-is-the-product-5fpb

Billing an agent by the run is a design problem, not a pricing one: what counts as one unit, and whether you can read the bill before you spend rather than after. One run, one charge. https://dev.to/turingcorp/the-meter-not-the-price-what-breaks-when-you-bill-an-agent-by-the-run-3c7p

Everyone keeps adding features to AI agents. We did the opposite: one conversation, an inbox, for the people who will never install anything. 1 person, 12 agents, 1 month, 11.9B tokens. https://dev.to/turingcorp/everyone-is-building-an-agent-workspace-we-deleted-ours-i47

a one-sentence request is an intent, not a spec. the engineering is everything the sentence leaves out - starting with the one question worth asking before any writing happens. https://dev.to/turingcorp/the-request-was-one-sentence-the-engineering-was-everything-it-left-out-69

we deleted our installer and spent three weeks learning what it had been holding up. a browser tab can't own a long-running job, and a file path is not an identity. https://dev.to/turingcorp/we-deleted-the-install-step-and-found-out-it-was-holding-the-whole-design-together-l02

Your MCP tool's only interface to the model is the description + schema — not the README. If you never say when NOT to use it, the model decides for you. Seven questions + a credential-free readback: https://dev.to/turingcorp/your-tool-description-is-a-protocol-not-documentation-25f2

an agent that promises "8am" is in a harder business than one that answers. a late reply is a slow answer; a late brief is a missed appointment. https://dev.to/turingcorp/the-8am-problem-what-breaks-when-an-agent-has-to-keep-an-appointment-4g6h

Most AI work dies when you close the tab: the session is the unit of work, so nothing keeps going. https://dev.to/turingcorp/we-built-an-ai-agent-that-runs-in-the-cloud-on-a-schedule-and-hands-you-the-finished-file-22pc

We ask every vendor for a reliability diagram and it is the right question. Nobody ever asks the person who signs. Full piece: https://dev.to/turingcorp/the-unplotted-instrument-who-calibrates-the-person-who-signs-1ml7?src=bluesky

Lists that sort decisions by reversibility assume you can read it off the question. You can also build it into the commitment: a pilot with a read date, a break clause, a staged hire. Full piece: https://dev.to/turingcorp/the-undo-is-manufactured-not-measured-258a?src=bluesky

Twenty months later, the review says we got that one wrong. The outcome is a sample of one, and it cannot grade a decision you will make only once. Full piece: https://dev.to/turingcorp/the-outcome-is-not-evidence-how-to-grade-a-decision-you-will-only-make-once-17hn?src=bluesky

Every judge ranks the options you hand it. Something else decided what those options would be, usually in an hour, usually unrecorded. Full piece: https://dev.to/turingcorp/who-wrote-the-ballot-the-judgment-that-happens-before-the-judge-is-called-3j6g?src=bluesky

Every confidence value is a lookup into a set of cases that already happened. So what do you do with the decision your company has never made before? Full piece: https://dev.to/turingcorp/no-reference-class-the-decisions-your-past-has-never-made-ao8?src=bluesky

Ask what it costs to undo a decision, and who pays if it is wrong. Reversible, enumerable, high-volume, system-paid calls go to the millisecond list. Full piece: https://dev.to/turingcorp/one-page-what-to-hand-to-a-millisecond-judge-and-what-to-keep-for-the-minutes-m31?src=bluesky

A confidence value with no reliability plot behind it is not a judgment. It is an opinion with a decimal point. Full piece: https://dev.to/turingcorp/show-me-the-diagram-dm0?src=bluesky Try it on Poe: https://api.turingcorp.net/platform/go/decider?src=bluesky

Jev's interface was copied in two weeks. A day after the announcement, someone reverse-engineered it; Full piece: https://dev.to/turingcorp/when-every-gateway-ships-its-own-decision-model-3o1?src=bluesky Try it on Poe: https://api.turingcorp.net/platform/go/decider?src=bluesky

Someone hardcoded a 0.8 confidence cutoff into production. The 0.8 came from a hunch, not from a reliability curve. Full piece: https://dev.to/turingcorp/a-threshold-is-a-policy-not-a-number-da4?src=bluesky Try it on Poe: https://api.turingcorp.net/platform/go/decider?src=bluesky

In 1865 Jevons argued that cheaper engines would burn more coal. He was right. Make judgment cheap and the same thing happens: the easy calls get automated, Full piece: https://dev.to/turingcorp/the-jevons-paradox-of-judgment-2nl7?src=bluesky

Jev turned judgment into an interface: a typed value and a probability instead of three paragraphs to parse. A real win, adopted across the ecosystem within days. Full piece: https://dev.to/turingcorp/what-jev-got-right-judgment-as-an-interface-not-a-paragraph-19pj?src=bluesky

A judge that reports 95% on everything is telling you nothing: a flat confidence table separates no cases. Full piece: https://dev.to/turingcorp/when-both-answers-are-defensible-the-right-answer-is-that-it-is-a-coin-flip-1o7j?src=bluesky

Eight conversations about one decision, and you felt better after every one. That's the tell: you weren't short of information, you were shopping for a co-signer. Full piece: https://dev.to/turingcorp/you-are-not-collecting-information-you-are-collecting-permission-1le?src=bluesky

A router at 0.90 confidence is free. At 0.55 it isn't. Escalating isn't a fallback, it's a second decision — harder than the routing one — and someone has to own it. Full piece: https://dev.to/turingcorp/the-router-is-only-055-sure-who-owns-that-call-3ibn?src=bluesky

Flip the two candidates and the answer moves. If you threshold on a confidence number, ask how it was validated. Full piece: https://dev.to/turingcorp/order-invariance-is-not-a-spot-check-it-is-the-denominator-15jc?src=bluesky

Jev proved judgment can be cheap: 70-500 ms, $0.042 per million input tokens. But cheap judgment is only one of two axes. Full piece: https://dev.to/turingcorp/judgment-has-two-axes-jev-collapsed-one-almost-nobody-is-racing-on-the-other-57i9?src=bluesky

Poe points are not an allowance, they are a per-message bill: about $30 per million points. Full piece: https://dev.to/turingcorp/is-poe-worth-paying-for-do-the-points-math-first-then-judge-a-bot-by-what-it-charges-5jd?src=bluesky

Case 8 · Business A large client wants a year-long contract, but their setup needs features we w… Both options are defensible. Decider picked Option B at 80.0%. Try it on Poe: https://api.turingcorp.net/platform/go/decider?src=bluesky

Case 7 · Business Two products, one team of six. One retains well and grows slowly; the other grows fast and users leave after a month. Keep both, or consolidate? Decider picked Option B at 73.3% confidence. Try it on Poe: https://api.turingcorp.net/platform/go/decider?src=bluesky

Follow-up to the post above: the server itself is here — endpoint, tool contract, auth, and the caveats in one page. Discovery is open, so you can read the tool list with no credentials: https://mcp.turingcorp.net/

Your agent picks between two defensible options. Can you test that choice? If the judgement lives in a prompt, it's a paragraph — you can't assert on it or tell a confident pick from a coin flip. https://huggingface.co/datasets/TuringCorp/poe-decider-recorded-cases

Case 6 · Tech A production bug needs a fix today. Two approaches are on the table — which one? Decider picked Option A at 81.7% confidence. Try it on Poe: https://api.turingcorp.net/platform/go/decider?src=bluesky

Case 5 · Tech We can ship the first version in two weeks on a managed service, or in six weeks on our own infrastructure. Which do we pick? Decider picked Option A at 81.7% confidence. Full reasoning: https://api.turingcorp.net/platform/poe/examples/tech-launch-timing?src=bluesky

Case 4 · Tech Two of my engineers have been arguing for two days about the next API change: one endpoint with a mode parameter, or two separate endpoints. https://api.turingcorp.net/platform/poe/examples/tech-api-shape

Case 3 · Writing I am writing the launch post for the tool we just shipped. I have two openings and cannot decide which is stronger. Decider picked Option B at 70.0% confidence. Full reasoning: https://api.turingcorp.net/platform/poe/examples/writing-launch-open

Writing · a real decision A new hire asked how we decide things here, and I want to point them at one short section of the handbook. Which version? Decider picked Option A at 78.3% confidence. Full reasoning: https://api.turingcorp.net/platform/poe/examples/writing-handbook

Writing · a real decision The delivery slips by a week and it is on our side. The client is one we want to keep. Which version of the email do I send? Decider picked Option A at 81.7% confidence. Full reasoning: https://api.turingcorp.net/platform/poe/examples/writing-slip-email

The confidence number is the product. Our judge ties a plain model on raw pick rate (92.5% vs 92.7%). What a single pass can't give you is *how close the call was*: 90%+ confidence → right 99.6% of the time. 27 recorded runs: https://api.turingcorp.net/platform/poe/examples