The decision record renaissance

In March 2026, Martin Fowler published a fresh bliki entry on Architecture Decision Records — a two-page format for capturing a single decision, its context, and its consequences, riffing on an idea Michael Nygard wrote down back in 2011. Fowler doesn't write entries about dead practices. A year earlier, in March 2025, AWS published its own best-practices guide for running ADR reviews, complete with meeting formats and approval workflows. When both the methodology canon and the world's largest cloud vendor are re-litigating a fifteen-year-old documentation format, something is up.

What's up, I'd argue, is that decision records are the rare document class whose value went up when AI made writing nearly free. Everything else — status updates, summaries, specs nobody reads — got cheaper to produce and cheaper to ignore. A record of what you knew, what you weighed, and what would have changed your mind is the one thing a model cannot generate after the fact, because after the fact it no longer exists anywhere. Not even in your head. Especially not in your head.

Your memory is a hostile witness

The case for writing decisions down starts with an unflattering fact: you cannot be trusted to remember your own reasoning. Psychologists have documented hindsight bias — the "knew-it-all-along" effect, studied since Baruch Fischhoff's experiments in 1975 — for half a century. Once you know how something turned out, your memory quietly rewrites the deliberation so the outcome looks like it was obvious. The uncertainty you actually felt, the option you almost picked, the assumption that later proved wrong: gone, replaced by a tidy story in which you basically saw it coming.

This is why the classic remedy is a decision journal. Farnam Street's long-running template has you record, at decision time, the situation, the alternatives you rejected and why, the range of outcomes you expect with rough probabilities — even how you feel physically, since "decided while exhausted and furious" is context your future self will absolutely delete. The point isn't the paperwork. The point is that a prediction written down before the outcome is the only version of your judgment that can't be doctored by the outcome.

The same logic runs the other direction, too. Outcome bias means a good result launders a reckless process, and a bad result convicts a sound one. Without a record, your team's post-mortems grade decisions by how the dice landed. With one, you can occasionally discover the more useful truth: we decided well and got unlucky — or the reverse, which is the one nobody volunteers.

Text got cheap, judgment didn't

Then the ground shifted under the whole document economy. Paul Graham's October 2024 essay "Writes and Write-Nots" predicted a world split between people who can write — and therefore think — and people who let the machine do both. In June 2025, Nature Reviews Bioengineering ran an editorial titled "Writing is thinking", fretting that if AI drafts the paper, we're reading the model's thoughts rather than the researchers'. By November 2025, practitioners were publishing setups for fully AI-generated ADRs — agents that watch a repository and file decision records automatically, removing the discipline problem that killed most ADR habits.

Read those three together and the shape of the moment appears. Producing a plausible decision document now costs nothing. A model can emit an articulate rationale for any choice you've already made, in the house style, with a confident "Consequences" section. Which means the document itself stopped being evidence of anything. The scarce artifact — the thing that still can't be faked cheaply — is judgment captured at decision time: what you actually knew then, what you actually weighed, and the tripwire that would have made you choose differently. An AI can transcribe that for you, and honestly should. It cannot invent it later without lying.

That's an interpretation, not a measured fact, so hold it loosely. But it explains why the renaissance is happening now rather than in 2015: when text was expensive, a written rationale signaled effort. Now it signals nothing, and the market has repriced accordingly. Provenance is the new prose.

How working teams actually run this

The mechanics are boring on purpose. Fowler's March 2026 write-up describes the settled practice: records live in the repository under doc/adr, numbered monotonically, in markdown, one decision each. A record is never edited into agreement with the present — it gets marked superseded and a new one links back, so the log stays an honest history instead of a wiki page that always agreed with you. AWS's March 2025 guidance adds the social protocol: readout-style reviews where people read silently and comment in writing, fewer than ten participants, one to three sessions and then you decide — a decision process that itself has a stop condition.

The format has been leaking out of engineering, plausibly because every function suffers the same amnesia. Marketing re-argues the pricing page; ops re-argues the vendor choice; the reasoning lives in a departed employee's Slack DMs. Async-first teams have converged on decision memos — context, options considered (including "do nothing"), the call, the owner, the date — precisely because a distributed team can't rely on hallway memory. And Amazon's contribution predates all of it: the framing Jeff Bezos spelled out in Amazon's 2015 shareholder letter, published April 2016 — Type 1 decisions (one-way doors, decide slowly and deliberately) versus Type 2 (two-way doors, decide fast and cheaply reversible). Classifying the door is itself the first line of a good decision record — it tells you how much record the decision deserves.

Six fields, written before you decide

The recommendation, distilled from all of the above into something you can use this week. Before any decision that clears the one-way-door bar, write down:

1. Context — the situation and constraints as known right now, five lines max.
2. Options considered — every live candidate, including "do nothing," each with the reason it might win.
3. Criteria and weights — what actually matters, and how much, decided before you look at how the options score. Weighting after scoring is how pros-and-cons lists launder foregone conclusions.
4. Expected outcome — what you predict happens, with a rough probability attached.
5. The flip condition — the smallest change in facts or weights that would reverse the call. If you can't name one, you haven't examined the decision; you've decorated it.
6. Review date — when you'll reread this against reality. A record nobody revisits is a diary, not an instrument.

Fields three and five are where the weighted-matrix method earns its keep: forcing explicit weights makes the tradeoffs legible, and hunting for the flip point tells you whether the verdict is sturdy or one nudged weight away from flipping. That pair — plus a shareable receipt of the whole thing — is the entire premise of Flipweight, but the fields work fine in a text file. The tool matters less than the timestamp.

The decision rule to take away: if a decision is worth a meeting, it's worth six fields written before the meeting ends. Your future self will read them with gratitude and mild embarrassment — which is exactly the feeling of learning something.