Observer.
Proof-design for software built by agents. Observer turns an accepted product spec into a reviewed proof design, compiles it into an immutable verification manifest, and runs it through real QA tools — so when the checks go green, the green means what it claims.
Built on AssuranceSpec, a framework-neutral standard for committed verification intent. In development at OpenAgents — designed in the open, dogfooded on our own desktop app before anyone else's code.
THE FAILURE MODE — A PASSING SUITE IS NOT PROOF
It proves the mock behaves like the mock — not that the product behaves like the spec.
Written after the code, it inherits the code's assumptions — including the wrong ones.
Both sides of a client–server or renderer–host boundary are faked, so integration defects pass clean.
Line coverage measures execution, not correctness — activity dressed up as evidence.
Skipped, stale, flaky, and inconclusive results quietly disappear into a passing summary.
Teams cannot reliably answer, criterion by criterion, whether the exact release artifact behaves as designed. Evidence links help you find the test — they cannot tell you it was the right test. A link is not a verdict.
HOW IT WORKS — SPECS BOUND TO REAL SYSTEM BYTES
WHAT KEEPS THE PROOF HONEST
The intended candidate must pass and a known-bad candidate must be refuted. An oracle that accepts both is ruled unsound.
A declared cross-process or client–server boundary drives both real sides, or cites a qualifying end-to-end receipt. Mock-only rows do not satisfy it.
Checks designed before the code can be red without blocking trunk. Once activated, required negative or unknown states block their gate.
Skipped, stale, flaky, unavailable, and inconclusive results stay visible as exactly what they are.
Criterion traceability, execution evidence, and explored frontier are separate ledgers. Line coverage never stands in for behavior.
No green manifest or run grants merge, deploy, spend, or public-claim authority. Release decisions stay separately owned.
DOGFOOD — OUR OWN SYSTEMS FIRST, IN THE OPEN
product_spec: docs/mvp/openagents-codex-workroom-mvp.product-spec.md format: ProductSpec 0.1 revision: 6 sha256: fba7963334eb736582003e7d903d0e57164e7fecb2c158c302af7fb23e3f6ef1 criteria: CW-AC-01 … CW-AC-18 assurance: revision 2 · reviewed and admitted · 18/18 obligations designed evidence: 18 CONFIRMED · 18 falsifiers REFUTED · 18 sensitivity receipts
That binding is real, quoted from the repository — Observer's first subject is the OpenAgents Desktop MVP spec itself. The design documents, format tooling, and generated first proposal are public in docs/assurance.
STATUS — REPORTED THE WAY OBSERVER WOULD REPORT IT
Bounded proposal profile — parser, serializer, structural validator, and CLI — implemented and tested in the open repository.
Revision 2 is reviewed and admitted: all 18 candidate oracles confirmed, all 18 falsifiers refuted, with 18 sensitivity receipts and a green Desktop gate.
A coverage skeleton from the exact spec bytes — no semantic proof inference is claimed.
Review annotations and exact-digest admission are implemented. Rich semantic planning remains reviewable proposal work, never compiler output or self-admitted policy.
Exact-digest bindings and byte-stable compilation are implemented with golden-byte and stale-input regression coverage.
The narrow native Bun-test adapter and normalized eight-axis receipts are implemented. Browser, device, resilience, accessibility, and formal adapters remain planned.
Sharded manifest runs within declared budgets; undistillable exploration stays inconclusive instead of disappearing.
The Desktop MVP now has a reviewed public-safe criterion trace. It reports the current missing proof honestly; multi-project hosting remains planned.
This page holds itself to the standard it describes: planned surfaces stay visibly unbuilt, and nothing here is a public product promise. Public claims live only in the product-promise registry.
WHAT'S COMING
The proof contract, free
- validate product specs, assurance specs, and environment profiles
- compile deterministic assurance manifests
- run obligations through your own native test tools
- local JSON/HTML evidence reports
- bring-your-own model for proposal and exploration
- no OpenAgents account required
Existing tests are imported before anything is generated — Observer rebuilds the harness around a proof-design control file; it does not discard the harnesses that already work.
Managed evidence, when it earns it
- managed browser, OS, and device matrices
- parallel QA Swarm exploration
- retained encrypted evidence and trends
- opt-in shareable evidence pages
The paid service sells managed environments, compute, and evidence retention. It does not hold the basic proof contract hostage.