Rebuilding Clojure from behavior
A graded clean-room study of the abstractions that outlasted twenty years of implementation churn

A research initiative. The instrument is a Clojure reimplementation; the output is a paper.

Clojure's source has been under version control since 2006-03-24, three years before its first stable release. Twenty years of that history is public, and it separates cleanly into two populations: a small set of abstractions that were right early and never moved, and a larger body of implementation that has been rewritten continuously underneath them. This project rebuilds the first population, under a protocol that forbids reading the second, to find out what forced their shape.

Not a goal: a working Clojure. Several exist and all are better than this one will be. A complete Clojure obtained by reading core.clj is a null result reported as a success, and it is the only outcome that would make the whole thing worthless.

Four properties, and few systems have all of them.

• Long public history with an early start. Version control predates the first public release by nineteen months. The construction period is visible, not just the shipping period.
• An explicit kernel/library boundary. Java core plus clojure/core.clj. Most of what a user touches is written in the language itself, so the boundary is a real seam rather than an analytical convenience.
• Runnable artifacts at every release. A behavioral oracle exists for each milestone without reconstructing a build.
• A small special-form set. The irreducible core is roughly fourteen forms. Bounded enough that "did we get it" is answerable.

The fourth is what makes the study finishable and the second is what makes it interesting.

RQ3 and RQ5 are answerable before any language code exists, and both are cheap. A paper carrying only those two, the method, and the threats section is still a paper. An implementation without them is a hobby.

All figures below measured 2026-08-16 against github.com/clojure/clojure at HEAD 2026-07-28, via a metadata-only clone. Epistemic labels per the house scale: [E] measured here, [P] sourced, [H] hypothesis, [I] inference.

The contribution is intended to be the protocol, independent of the Clojure result. Sources form a seven-grade ladder, and the grades are separated by tooling rather than by intention.

The oracle is a release jar with every .clj file stripped out. Clojure boots from its AOT'd class files, so (doc map) still answers and (source map) has nothing to read. The boundary is a zip operation, not a promise.

The G5 line runs through the secondary literature too, not only the source. The HOPL IV history is G3 throughout except §3.4.1 on persistent data structures, which names hash array mapped tries, the branching factor of 32, path copying, and System.arraycopy, and §3.6.2 on the STM, which gives MVCC and the ensure mechanism for write skew [P]. Both are enough to build from and both sit behind a waiver.

The waiver log is the primary output. Each entry names a feature that black-box probing could not reach and how close it got. That log is RQ2's answer directly. A run with no waivers means the budgets were too generous; a run that is all waivers refutes the method. Neither is failure. The failure is a run where the waivers happened and were not written down.

digraph structure {
  rankdir=LR
  bgcolor=white
  node [shape=box, style="rounded,filled", fontname="Helvetica", fontsize=10]
  edge [fontname="Helvetica", fontsize=9, color="#334155", fontcolor="#334155"]

  H   [label="clojure/clojure\nblob-less clone", shape=cylinder, fillcolor="#dbeafe", color="#1d4ed8", fontcolor="#1d4ed8"]
  J   [label="stripped release jar",             shape=cylinder, fillcolor="#dbeafe", color="#1d4ed8", fontcolor="#1d4ed8"]
  SEQ [label="sequence:\nwhat to build next", fillcolor="#f1f5f9", color="#334155", fontcolor="#334155"]
  ORC [label="oracle:\nis it right",          fillcolor="#f1f5f9", color="#334155", fontcolor="#334155"]
  B   [label="rebuild the survivors", fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"]
  R   [label="results table", shape=cylinder, fillcolor="#fef9c3", color="#a16207", fontcolor="#a16207"]
  W   [label="waiver log",    shape=cylinder, fillcolor="#fee2e2", color="#b91c1c", fontcolor="#b91c1c"]
  P   [label="paper", fillcolor="#ede9fe", color="#6b21a8", fontcolor="#6b21a8"]

  H   -> SEQ [label="G1"]
  J   -> ORC [label="G0"]
  SEQ -> B
  ORC -> B
  B   -> R [label="gate passes"]
  B   -> W [label="budget exhausted"]
  W   -> R
  R   -> P
}

• RQ5 cold-probe baseline. The operator and the model in the loop have both read Clojure. Unenforceable, so it must be measured rather than assumed. Highest priority; it conditions the interpretability of everything else.
• Hand-labelled sample for the F5 filter, to put precision and recall on it.
• Churn ranking recomputed with vendored paths and bulk commits excluded, to test F1 properly.
• Claim table freeze. Until it is frozen, no result here is pre-registered and should not be cited as though it were.

Stated before results rather than after, where they act as a design constraint instead of a defence.

T1 and T5 together mean the honest framing is a case study: one operator with prior exposure, one target, under a graded protocol. Saying so in the title costs nothing and forecloses the obvious objection.

Every figure on this page is produced by an instrument in the repo, not quoted. The code gate recomputes them; a hand-typed number fails it.

# Regenerates every measurement on the initiative index.
set -euo pipefail
SEQ="${SEQ:-vendor/sequence/clojure}"

[ -d "$SEQ" ] || tools/clone_sequence.sh "$SEQ"

echo "== F2/F4: history extent and clone size =="
du -sh "$SEQ/.git"
git -C "$SEQ" rev-list --count HEAD
for d in 2007-10-17 2008-12-17 2009-05-04; do
  printf '  before %s: %s\n' "$d" "$(git -C "$SEQ" rev-list --count --before="$d" HEAD)"
done

echo "== F3: construction order, first two weeks =="
git -C "$SEQ" log --reverse --format='%ad  %s' --date=short | head -12

echo "== F5: G2/G4 split =="
python3 tools/leak_grade.py "$SEQ"

echo "== F1: survival and churn =="
python3 tools/survival.py "$SEQ" --exclude 'src/jvm/clojure/asm/' --max-files-per-commit 50

echo "== F6: sequence triangulation =="
git -C "$SEQ" for-each-ref --sort=creatordate \
  --format='%(creatordate:short) %(refname:short)' refs/tags | head -3

tools/survival.py does not exist yet. Writing it, with the vendored-path and bulk-commit exclusions applied, is the first task that changes a number on this page.

For whoever grinds on this next, agent or otherwise. This page is the initiative's front door and its running state; it is not a snapshot.

1. Findings are append-only. New observations get the next F number. Never renumber. A finding that turns out wrong keeps its number and gains a REFUTED marker with a pointer to the finding that replaced it. Negative results are the cheapest knowledge here and deleting one destroys the audit trail.
2. Every figure cites its instrument. If scripts/reproduce-index.sh does not produce a number, the number does not belong on this page. Adding a figure means adding the instrument first.
3. Update the header on every change. #+LAST_MEASURED carries the date and the upstream commit the figures were taken against. Numbers drift as clojure/clojure moves; an undated figure is unfalsifiable.
4. Epistemic labels are mandatory. [E] only for something measured by a committed instrument on a recorded date. Reasoning about what the data would show is [I], not [E], however confident it feels.
5. Refutation conditions go on findings, not on the page. A finding without one is an observation; say "observation" and move on rather than dressing it as a result.
6. Status line changes when the claim table freezes. Before the freeze nothing here is pre-registered. After it, #+STATUS names the tag and findings become citable as pre-registered.
7. Do not add a fact here that is not in spec.org. The spec governs; this page is derived and public. Where they disagree, the spec wins and this page is the bug.
8. Findings that touch the sibling essay get cross-linked both ways. F7 is the current one. A claim that lives on one page and is assumed on the other is how two documents drift into contradicting each other, and neither author notices.

Assertion-titled findings live at /research/2026-rebuilding-clojure-from-behavior/<claim>/ in the house convention: the page states the claim and the URL states the claim, while the initiative URL stays method-shaped so it survives the answer moving. That is why this page is not at /research/2026-what-survived-clojure/, which belongs to the diffusion essay and to a different question at a different scale.