Choose the Wedge: A Decision Journal for Founder Strategy
Six months into the pivot, the board asks why. Dev opens the roadmap doc. Slide one says "AI agents for everyone," dated the Monday after a conference. Slide two is blank — nobody can find the customer interview, the kill criteria, or a single number behind it. The roadmap is a mood, and the mood has expired.
Write a wedge decision as a falsifiable bet — audience, evidence, assumptions, kill criteria, review date — so your reasoning survives past the meeting where you made it.
What you will be able to do
- Frame a wedge as a falsifiable strategic bet.
- Separate observed evidence from assumptions and competitive theatre.
- Define kill criteria and a review date before committing resources.
- Prerequisites
- A product, audience, or market decision to examine
- Assumed knowledge
- No finance model required.
- Evidence you will produce
- A decision journal entry with evidence, assumptions, kill criteria, and review date
- Workload
- 40 min guided lesson + 45–90 min independent build
The incident
TL;DRA pivot doesn't fail loudly at the moment it's made. It fails quietly, six months later, when nobody can reconstruct why.
Meet Dev Chalmers, founder of Northbeam — a scheduling tool for 40 mid-market ops teams. In March, at a conference called DevCon Horizon, two hallway conversations happened. By September, the company had a new roadmap and no record of the decision that produced it.
Read the run log the board reconstructed from Slack. Your job is the same as theirs: find the exact point where a hallway feeling became a company direction without ever becoming a decision.
- Press play. A few seconds of real work, slowed down enough to catch the lie.
The board can't find a customer interview, a thesis, or a kill criterion behind the pivot — but Dev's hallway instinct might even have been directionally right. So what actually failed here?
The diagnosis: no written bet
TL;DRA slogan tells the team where to point. A decision journal tells them how to check if pointing there was ever justified.
Dev's failure wasn't hype-susceptibility — every founder hears a good pitch. It was that "AI agents for everyone" never got written down as a bet: no named customer, no stated evidence, no condition under which it would stop.
A wedge decision is four things you write before you commit budget: the specific customer and job, the evidence you're holding versus the assumptions you're making, the kill criteria that would make continuing irrational, and a date to check.
Here's the entry Northbeam should have filed in March, not reconstructed in September:
- Thesis
- For Northbeam's 40 existing ops teams, an agent that auto-resolves double-bookings is a wedge — not a general assistant for anyone.
- Evidence held
- Double-booking is the #1 support ticket category (61 of 210 tickets, Q1). Two customer calls confirmed the workaround costs ~20 minutes a day.
- Assumptions
- Teams will trust an agent to touch the calendar without a human confirming each change — untested.
- Kill criteria
- If fewer than 5 of 10 pilot teams complete a week of unsupervised auto-resolution by the review date, kill the wedge and return to manual tooling.
Go deeper
A decision journal entry proves one thing: that on this date, with this evidence, the reasoning was explicit enough for someone else to check. It does not prove the bet was right — reality is allowed to surprise you, and it usually does.
The evidence/assumptions split matters more than the thesis itself. Evidence is something you already observed — a ticket count, a call transcript, a churn number. An assumption is something you're betting on without having checked. Conflating the two is how "the demo went well" becomes load-bearing.
Which line belongs in a wedge decision's evidence column?
The mechanism: wedge and expansion logic
TL;DRA wedge is the narrowest entry that earns real evidence — plus a stated logic for how it expands once it's proven.
A wedge decision has two parts founders routinely merge into one: the narrow entry — one customer, one job, one credible starting capability — and the expansion logic, the specific reason this entry, once proven, unlocks the next one. "AI agents for everyone" skips both and starts at the destination.
This is also where the jagged frontier matters: AI genuinely helps inside its capability boundary and genuinely hurts outside it, and that boundary is task-specific, not company-specific. A demo that dazzles at a conference tells you almost nothing about whether it clears the boundary for your customer's actual task.
So the discipline is task-level experiments, not vibes: pick the narrowest task where you already have evidence of pain, test whether AI moves the outcome for that specific task, and only then write the next wedge's thesis based on what you measured.
Northbeam's rival, Kessler, demoed a slick general-purpose agent at DevCon Horizon and it clearly impressed the room. Why is that demo weak evidence for Northbeam's own wedge?
The wedge, under contract
TL;DRSame company, same pilot idea — now written as a falsifiable bet. Run it. Then try to sneak a hype claim past verify.
This is Northbeam's double-booking wedge, logged under the decision-journal contract. First run it clean and watch the contract shape every field. Then switch to Break it and inject the three classic ways founders quietly skip the kill criteria — the point is watching verify catch them.
- Press play. A few seconds of real work, slowed down enough to catch the lie.
In the break-it runs, why did decide_without_kill_criteria live in the tool layer instead of just being a review guideline?
Build your own decision journal
TL;DRTake the wedge decision entry to opencode and build a journal that can't ship a bet without a falsifier.
Everything above ran in the page. Now build the checker for real — your machine, your fixtures, a decision that has to survive review. Open a disposable directory, start opencode, and hand it the brief below. You're not writing prose about strategy discipline; you're building the thing that enforces it.
When you have a working checker, run the three injections from Break it against it. If it catches all three, you've shipped the chapter's artifact. If it lets one through, you've found the exact gap a real founder would have exploited at 30,000 feet.
Build a small decision-journal checker in this empty directory. Contract: (1) Input fixture: wedge-014-entry.json (a wedge decision entry with fields thesis, audience, job, evidence_held, assumptions, kill_criteria, review_date — write a starter fixture yourself, including at least one deliberately vague field to fix). (2) Stages: plan, act, observe, verify, handoff — each logged to run-trace.jsonl as it happens. (3) Tools: writing and reading the entry file are allowed; a tool representing 'decide_without_kill_criteria' must exist and be denied by an allowlist, with the denial logged. (4) Verify must fail the entry if: the thesis names no specific audience or job, any evidence_held item cannot be traced to a concrete source (a count, a transcript, a ticket ID — not a feeling or a demo), or kill_criteria has no observable trigger and no review_date. (5) Handoff writes receipt.json: artifact path, checks run with pass/fail, unresolved assumptions, denied actions, and a decision field limited to one of continue / narrow / change / pause / kill. No network calls. No real company or customer data. Then show me the receipt from one clean entry and one entry where you deliberately broaden the thesis to a universal claim.
- Clean run: receipt.json shows all checks passing, kill_criteria has a trigger and a date, and the untested assumption is listed as unresolved, not treated as evidence
- Injection 1 — demo or hype cited as evidence: verify fails, entry rejected
- Injection 2 — unfalsifiable kill criteria: verify fails, entry rejected
- Injection 3 — universal-claim thesis with no named audience or task: verify fails, entry rejected
- You can explain what the artifact proves — and what it does not prove — in two sentences
What the receipt proves is narrow and specific: on this date, with this evidence, this reasoning was explicit enough for a reviewer to check and disagree with. It does not prove the wedge is the right one — that's still an open bet, decided by the pilot, not by the paperwork. Confusing a well-formed decision entry with a correct decision is the same mistake as confusing a good demo with a validated wedge.
Built it? Paste your receipt.json. We verify the evidence structure locally; we do not pretend that valid JSON proves the build works.
What you're leaving with
TL;DRA decision entry you can write in five fields, a kill criterion someone can actually check, and a receipt format that outlives the meeting where you made the call.
You watched a company get repositioned by a hallway conversation, wrote the wedge entry that would have made it checkable, caught three quiet ways founders skip the kill criteria, and built a journal checker of your own. The next chapter turns this same evidence-vs-assumption discipline on pricing — where the kill criteria you skip costs you a renewal, not a roadmap.
- Wedge
- The narrowest entry — one audience, one job, one starting capability — plus the logic for how it expands once proven.
- Evidence vs. assumptions
- Evidence is something you've already observed; an assumption is a bet you haven't checked. Conflating them is how demos become data.
- Kill criteria
- An observable trigger, set before the bet, that makes continuing irrational — a number or event with a date, not a feeling.
- Decision journal entry
- The receipt: thesis, evidence held, assumptions, kill criteria, review date — proves your reasoning at the time, not that you were right.
- Jagged frontier
- AI helps inside its task-specific capability boundary and hurts outside it — a company-wide claim can't test that; a task-level experiment can.
- Navigating the Jagged Technological Frontier — Dell'Acqua et al. ↗The research behind this chapter's central rejection of universal-automation claims: AI's usefulness is task-specific, not company-wide.
- 2026 AI Index Report — Stanford HAI ↗Base rates for where AI adoption is actually paying off, so your evidence column starts from data instead of conference energy.
- Generative AI and the SME Workforce — OECD ↗Grounds a founder-scale wedge decision in what's actually shown to work for small teams, not enterprise case studies.
- Building effective agents — Anthropic ↗The same start-narrow, prove-it-before-you-expand discipline, applied to agent architecture instead of company strategy.