dna

_{ONE GENOME, REVEALED ONCE}
a genome, unfurling itself exactly once. base by base.

live at https://dna.maurice-frank.com, running since 2026-06-09. at human scale:

rate	whole genome in
30 b/s (live)	~3.3 years
~101 b/s	~1 year
~1235 b/s	~30 days

build

requires samtools (>= 1.16 for samtools consensus) and python3.

scaffold/build.sh \
  --cram sample.cram \
  --ref  reference.fa \
  --out  artifacts \
  --start-epoch 2026-07-01T00:00:00Z \
  --runtime 31536000          # ~1 year; or use --rate 101 (bases/sec)

# add the optional pileup (large on disk):
scaffold/build.sh ... --pileup --max-depth 64

inputs: cram, .crai (else auto-created), and the reference fasta (else auto-created). the reference is mandatory; a cram cannot be decoded without it.

run

the runtime is a cloudflare worker over an r2 bucket. no server, no state.

npx wrangler dev --remote      # runs against the real bucket
npx wrangler deploy
node worker/gate.test.mjs      # asserts nothing past f(now) is reachable

the page talks to three endpoints:

endpoint	what it gives
/meta	n, rate, startEpoch, contigs. cacheable, no sequence in it
/head?from=	the slice revealed since from, ending exactly at f(now)
/pileup?pos=	one pile-up column, refused if pos is past f(now)

everything else is served straight from web/ as a static asset.

touristic attractions are fetched on the worker from ensembl and ucsc, keyed to the 200 kb window around the playhead and memoised in the edge cache, so a window costs one round of upstream calls per hour no matter how many people are watching. with nobody watching there are no requests and so no fetches at all. which attraction is showing is derived from the clock rather than pushed, so every viewer sees the same one at the same moment without the worker holding any state.

pacing lives in wrangler.toml (START_EPOCH, RATE). changing either moves the playhead, so treat them as fixed once the piece is running.

the gate

the genome is revealed exactly once, in step with wall-clock time, and nothing ahead of the playhead is retrievable — that is the whole point of the piece and it is enforced in one place.

the r2 bucket is private: no public access, no r2.dev domain. the only path from those objects to a response body is readBases() in worker/src/index.js, and every caller clamps its end to currentIndex() first. an endpoint that reads the consensus goes through that function or it does not read the consensus. worker/gate.test.mjs pins the behaviour.

deploy

pushes to main deploy via .github/workflows/deploy.yml, which needs CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID in repo secrets.

the artifacts are uploaded once, by hand — wrangler caps a single object at 315 mb, so the 3.1 gb consensus needs a multipart client against the s3 endpoint:

wrangler r2 bucket create dna-artifacts -J eu
rclone copy artifacts/ r2:dna-artifacts/ --include consensus.bin --include meta.json \
  --s3-chunk-size 100M --s3-upload-concurrency 4

the bucket is eu-jurisdiction, so the s3 endpoint is https://<account>.eu.r2.cloudflarestorage.com — the .eu. matters, a default-jurisdiction bucket is simply not visible there and vice versa. rclone also wants region = auto. the worker reaches the bucket through its binding rather than the endpoint, but jurisdiction = "eu" still has to be on the [[r2_buckets]] block or the binding resolves to the wrong namespace.

a token scoped to one bucket cannot ListBuckets, so rclone lsd r2: returning 403 is expected — list inside the bucket instead.

the bucket must stay private. do not enable the r2.dev public url.