diff --git a/docs/decisions/0027-adopt-program-driven-language-development.md b/docs/decisions/0027-adopt-program-driven-language-development.md
new file mode 100644
index 0000000..7df916d
--- /dev/null
+++ b/docs/decisions/0027-adopt-program-driven-language-development.md
@@ -0,0 +1,52 @@
+# 0027: Adopt Program-Driven Language Development
+
+Date: 2026-09-23
+Status: accepted 2026-09-23 (BDFL ruling on review at 6b36621).
+
+## Context
+
+Hum's north star (0024) is a claim about humans and agents reading, writing,
+and verifying real programs. Prose cannot carry that claim; programs must.
+This record adopts the loop Hum now follows: real programs drive the
+language, friction is recorded per program, and additions clear a
+general-fix test before they land.
+
+## The loop
+
+1. **Real programs drive the language.** wordfreq is the first real program
+   (WO27). Next: the config-file parser and the state machine with ownership
+   transfer chosen by 0026 to stress the closed-world accountability
+   objective — not to demo syntax. Programs 2 and 3 are the current order;
+   wordfreq's ledger may reorder them — update this record rather than
+   treating the order as fixed.
+2. **Friction goes into per-program friction ledgers.** Every friction point
+   is classified: program gap, tooling gap, docs gap, or language gap. The
+   ledger is a deliverable alongside the program.
+3. **Missing surface is a STOP; gaps become decision records before
+   implementation.** Worked examples: 0021 recorded the missing newline-literal
+   mechanism as an open dependency blocking wordfreq's Part 2 instead of
+   inventing one; 0022 resolved it as a decision record first.
+4. **The general-vs-specific test.** An addition must be the general fix, not
+   a one-program special case. Program-specific helpers are rejected by
+   default. Precedent: 0021's "no general string library" ban — `text_split`
+   is the single tokenization primitive; anything further needs its own
+   decision record.
+5. **Finished programs join the bake-off corpus.** The scorecard is re-run
+   against each finished program, so the 0014 ownership bet keeps being
+   tested (backlog item 2). Corpus provenance is recorded per program.
+6. **"Best for humans and agents" is measured four ways:** the friction
+   ledgers (exist today); the agent repair benchmark — agents fixing broken
+   programs, success rate and attempts (to be built); the human review study
+   (backlog item 3; to be built); and CI minutes per PR (exist today).
+
+## Consequences
+
+- Every language addition cites its motivating evidence — a friction-ledger
+  entry, a benchmark or measurement result, or a study — and the decision
+  record that specified it. Some legitimate additions (safety profiles,
+  perf-debt fixes, 0026 study results) will not come from program friction.
+- The general-vs-specific test is a review gate: a proposal that serves only
+  its motivating program is rejected by default.
+- The bake-off scorecard and the four measures are re-run as programs finish.
+  A finished program that moves none of the measures is reported: a program
+  that teaches nothing is itself a signal.