Accelerated dogfooding#

We do not substitute a busy afternoon for 30 elapsed days. We do compress the feedback loop by running the entire cold-consumer journey repeatedly, in new temporary repositories, and retaining every outcome in a commit-bound record.

npm run dogfood:soak -- --trials 20 --out evidence/dogfood/accelerated-soak.json
npm run dogfood:soak:check

Each trial installs the packed candidate, exercises the demo, authors and onboards a policy, observes RED, verifies advisory behavior, fixes the repository to GREEN, verifies the evidence bundle, calls all MCP tools, and installs the same candidate through an immutable Git snapshot. The runner does not stop on the first failure. It records the exit code, duration, output digests, success markers, and a sanitized diagnostic tail for failed attempts.

The evidence class is deliberately author-operated-accelerated-soak. It supports a claim about repeated journey reliability on the recorded platform and commit. It does not support claims about 30-day stability, independent adoption, production use, or private-estate tenant behavior.

Three-ring vision#

Honest Mirror: this public repository performs and publishes the complete lifecycle it asks adopters to use. The accelerated soak is the first machine-derived reliability record.
Continuous Proof Crucible: PR, nightly, and release runs retain failures, exercise policy changes and recovery, and measure scoped outcomes rather than merely counting gate executions.
Private consuming estate: private repositories consume the exact public artifact. Only sanitized aggregate outcomes cross the boundary, and they count as public BCE dogfood only when bound to the public package version, package integrity, and source commit. Private doctrine, topology, tenant identifiers, and raw telemetry never cross it.

An independent operator remains a separate evidence class. Repetition by the authors cannot turn itself into an independent witness.