Nightly · Evidence-gated · Human-merged
A config-driven engine that runs one research-and-evolution cycle
against a repository each night. In an isolated cloud session it forms
one falsifiable hypothesis, tests it against the
repo’s real benchmarks, and records the result as a gist, an issue,
and a draft PR. Every run ends in ACCEPT,
REJECT, or INCONCLUSIVE. It never merges — a
human reviews and decides.
Freeze the model. Evolve the harness. Evaluation is not promotion — the machine never merges; a human does.
What it is
Two nightly cycles run today in a cloud scheduler — one against
Ruflo, one against MetaHarness — each driven by its own ~800-line
prompt. They run the identical 26-step pipeline and differ only in a
small, well-defined set of per-repo details. Two instances is exactly
enough to see the seam: the shared spine becomes an engine, and the
differences become a dream.config.
A dream.config declares the per-repo delta — rotation
slots, cron, evaluator entrypoints, ADR shape — and compiles to the
full routine prompt. The scheduler copy and the in-repo mirror
become build outputs of one source, so the drift between them stops
being a convention and becomes structurally impossible.
The expensive stages delegate to the packages that already own them: flywheel for the promotion gate, darwin for bounded evolution, redblue for adversarial critique. They are optional backends selected by config — a night without any of them is a degraded night, not a failed one.
The session never merges, never self-promotes flywheel state, never weakens a test or a threshold, never edits gold answers, never force-pushes, never publishes packages. Its only publication surfaces are a gist, an issue, and an always-draft pull request.
The nightly pipeline
Scroll to walk the cycle. Each stage narrows what tomorrow needs to search — the system optimizes for shrinking the search space, not for producing pull requests.
Nightly cycle
The night begins by reading every previous night. The ten-column
LEDGER.md is the only durable cross-night memory,
and it feeds deterministic learning signals: suppress a finding
already made, bias away from a surface that has gone fourteen
nights without a merge, throttle a low-scoring streak.
One deep surface and a couple of scan targets come off the
rotation declared in dream.config. The session reads
the repository and the field around it — competitor frameworks,
published results, the repo's own ADRs — before it is allowed to
have an opinion.
Exactly one falsifiable claim is written down and frozen before anything is measured. Freezing is what makes the night honest: a hypothesis that can be edited after seeing the numbers is not a hypothesis, it is a summary.
The frozen claim becomes an actual change — a diff you could read, not a description of a direction. Vagueness here is what lets a loop declare success without ever having tested anything.
The unmodified harness is measured first, in this session, on this checkout. A baseline inherited from a previous night is a baseline measured under conditions nobody re-verified.
The candidate runs against the same tasks, the same sample size, the same frozen model. Only the harness is allowed to vary — that is the whole point of freezing the model, and it is what makes the difference attributable.
A red/blue harness attacks the result: what would have to be true for this measurement to be wrong? Then the reward-hack scan asks the harder question — did the candidate improve the work, or just improve the score?
@metaharness/redblueIf the candidate survives critique, a fenced evolution stage explores variations around it under a fixed generation and population budget. Bounded is the operative word: the search is allowed to be creative, not allowed to be open-ended.
@metaharness/darwinEvery measurement is captured as a receipt that can be replayed later, and the promotion gate decides — on the retained evidence, not on the narrative — whether the candidate cleared the bar. The session may read the gate's answer. It may never move the bar.
@metaharness/flywheel
The night's published gist and the session's commit are hashed
together —
sha256(sha256(gist) + SESSION_COMMIT) — into a
single witness. Anyone can re-derive it; a single changed byte in
either input breaks it. Provenance stops being a claim and starts
being checkable.
An issue goes up carrying the hypothesis, the measurement, the critique, and the witness. It goes up whether the answer was yes or no, because the point of the night is the finding, not the outcome.
If there is a change worth proposing, it opens as a draft — every night, without exception. This is the invariant everything else protects: the machine recommends, and a human decides. Promotion is a human act.
The night closes by appending one row: date, surface, hypothesis, finding, verdict, evidence, artifacts, witness. The row is written even when the run halts on budget — a night with no row is a night that never happened, and tomorrow would have to rediscover it.
@dream-machine/ledgerThe three verdicts
The verdict vocabulary is deliberately small and deliberately unflattering. Two of the three outcomes produce no pull request at all, and both of them are good nights.
The frozen hypothesis survived measurement, critique, and the reward-hack scan, and cleared the promotion gate on retained evidence. The result is a draft pull request and an issue — a recommendation for a human, not a merge.
The hypothesis was measured cleanly and did not hold. This is a successful night: a question that was open at dusk is closed by morning, and tomorrow's search space is smaller for it. Most nights should land here.
The measurement could not be taken — blocked credentials, a budget halt, a sample too small to separate the arms. Stated with its reason, this is also a success, because the one thing it refuses to do is guess. A ledger row is written either way.
What it composes
The Dream Machine builds on the ruvnet stack as config-selected evaluation backends — never hard dependencies. It follows the source routines' own rule: use the best available implementation, and do not introduce a dependency merely to satisfy a prompt.
| Capability | Package | Used for |
|---|---|---|
| Promotion gate, receipts, replay |
@metaharness/flywheel
|
Evidence retention and the promotion gate |
| Bounded evolution |
@metaharness/darwin
|
The fenced Darwin stage |
| Adversarial red/blue |
@metaharness/redblue
|
Adversarial critic and reward-hack scan |
| Security scan, genome, audit |
metaharness
CLI
|
Security review and control-plane discovery |
| Memory, hooks, orchestration | ruflo (@claude-flow/cli) |
The cloud session's task and memory substrate |
| Vector memory over prior nights |
ruvector
/
agentdb
|
Roadmap — semantic search across the ledger |
Quickstart
The CLI surface is small on purpose: scaffold a config, compile it, and then work with the two durable artifacts a night produces — the ledger row and the witness.
npx dream-machine init # scaffold a dream.config for this repo
npx dream-machine compile # dream.config → the routine prompt
npx dream-machine ledger # parse / append / verify LEDGER.md
npx dream-machine witness # stamp and verify a night's provenance
npx dream-machine schedule # prepare the nightly routine
npx dream-machine tui # browse past nights in the terminalWalk each of these end to end in the tutorials, or see what a week of nights looks like on the dashboard.
{
"repo": "ruvnet/metaharness",
"cron": "0 8 * * *",
"slots": [
{ "deep": "generator-genome",
"scan": ["router", "turn-credit"] }
],
"adrTemplate": "metaharness-7-section"
}The config declares only the per-repo delta. Onboarding a third repository is writing one of these — not forking and hand-editing an 800-line prompt.
Tutorials & walkthrough
Get the CLI, scaffold a dream.config, and understand
every field it asks you for.
Turn the config into the routine prompt, and see why the scheduler copy and the in-repo mirror can no longer drift apart.
Re-derive sha256(sha256(gist) + SESSION_COMMIT) by
hand and confirm a night's provenance yourself.
The ten columns, what the learning signals do with them, and how a REJECT row makes tomorrow cheaper.
Put the compiled routine on a cron and keep the in-repo mirror in lockstep with it.
The management view: verdict distribution, findings, witnesses, and links out to every issue and draft PR.