Technology
Replay
A system that cannot explain its important decisions becomes increasingly difficult to trust. Replay is PING's answer: the ability to reconstruct what happened, why it happened, and prove that the reconstruction matches the original.
In developmentWhat replay must answer
For important behavior, PING aims to answer, after the fact:
- What happened?
- Why did it happen, and what information caused it?
- Which authority made the decision?
- Which version of the system executed it?
- Can we reproduce the result, and prove the replay is equivalent?
Replay is therefore not debugging infrastructure. It is institutional memory: the difference between a system that learns and a system that merely runs.
Determinism as a product feature
Where PING claims deterministic behavior, determinism is a guarantee, not an implementation detail. Same causal history should produce the same replay ordering, the same execution, the same canonical state, the same canonical bytes, the same hash. Constitutional outputs must not depend accidentally on machine, OS, locale, database ordering, timestamps, or random IDs.
Where nondeterminism is intentional (sampling, timing, external data), it is made explicit and kept outside deterministic constitutional outputs. Accidental nondeterminism is a bug; declared nondeterminism is a design choice.
The witness concept
A witness is a compact, checkable attestation that a particular execution produced a particular canonical result. Witnesses let two parties agree on what happened without re-running the world: compare the witness, compare the hash, and you know whether you are looking at the same history. This is the direction; the full witness protocol is still being designed.
Status
Replay semantics are designed and partially implemented: event ordering, canonical state derivation, and hash-chained attestation exist in prototype form. The end-to-end “replay this decision and prove equivalence” loop is in development. Until it is complete, this page documents the design, not a finished capability.
From the studio notebook
- The Retry That Charged Twice
A lost response, a second click, two charges: why retries without an identity duplicate work, and the small index-card fix that makes asking twice safe.
- What a Hash Taught Us About Trust
Our studio learned to run on hashes the way you learn to run on anything: by being burned in its absence. The hash ends the argument before it starts.
- The Night GitLab Deleted Its Own Database
One tired engineer, one terminal on the wrong server, 300 gigabytes of production data gone: a backup is not a file, it is a restore you have performed.
Talk to us about your business.
PING Social helps real businesses put AI to practical use: fewer missed opportunities, less busywork, systems that remember.
