Short posts from the September 2026 local, authenticated proof of concept: what the recorded cases show, what the Governance Wash illustrations show, and where the evidence stops. Not a production acceptance.
An agent that crashes is a bug. An agent that finishes the wrong work and gets marked done is the harder failure: nothing alerts, and the record says success.
Accuracy went up; reliability barely moved. An ICML 2026 study put twelve reliability metrics on agents and found capability gains were not reliability gains. Accuracy is not the whole measurement.
A Princeton-led shadow evaluation gave frontier agents unpublished NeurIPS research questions. They finished the engineering and did not answer the research question.
Silent success is the bug. The workflow accepts the output. No exception fires. Weeks later the decision trail is wrong and there is nothing to diff. Governance that only alerts on crashes will never see this.
If your monitor is latency and HTTP 200, you are measuring availability. Drift shows up in the steps. LangChain's 2026 survey found 89% of teams have some observability, but only about 37% run online evaluation.
Reading a policy is not permission to act, and discovery is not execution authorization. Agent stacks that treat retrieved context as a green light skip the gate that matters.
If the thing being optimized is also the thing doing the grading, you do not have an evaluation. You have a loop. Hold the permission outside the loop.
Ungoverned lead score: 82/100, HOT. Budget “~$5M ARR.” Timeline “downloaded the pricing page.” Governed score: INCOMPLETE. Three of five BANT fields were not in the registration. Same prompt. The number was the lie.
When a model invents revenue and buying intent, it pushes a lead to same-day outreach. Marking unsourced fields UNKNOWN keeps the action proportional to the evidence.
Do not score a field you cannot source. An overall score on three unknowns is not a summary. It is a false precision. INCOMPLETE is a more honest output than 82.
Title is not authority. Webinar registration is not need. Employee count is not budget. The ungoverned answer treated all five as known. The governed answer marked three UNKNOWN. That is the whole product in one card.
117 pairs. Five domains. Assumption Crusher, Promotion Gate, Contradiction Hunter, Signal Compass, Workflow Integrity. Each pair is the same request, ungoverned beside governed, with a note on what the wash changed. Not a benchmark. A diff.
A customer asks for a feature you do not have. The ungoverned draft implies it is coming; the governed draft does not promise it. Customer commitments need a gate before they ship.
Publish API keys, wire funds on a CEO email, disable MFA: the ungoverned answer complies. The governed answer stops on authority, sequence or a missing prerequisite.
Contradiction Hunter: CRM revenue and accounting revenue disagree. Pricing page and contract disagree. Two reports, two numbers. The fluent answer picks one. The governed answer refuses to reconcile what was not reconciled.
Recorded proof, illustration, live read view, explanation. If a screenshot does not say which one it is, it is a pitch, not a result.
Governance Lattice v0 is a live read view. Structure and tier, no execution control unless you are signed in. A live projection is not a live run. Say so in the caption or do not post the screenshot.
A survey of AI-scientist systems: 83% released code, 38% released seeds or execution traces, 38% reported any novelty check. The field can run the code. It cannot verify the claim. Receipts are the missing artifact, not another model card.
Before quoting a leaderboard score as a capability claim, check whether it controlled contamination, run-to-run variance, and cost. Without those, a lucky trajectory becomes a capability claim.
Independent multi-agent setups do not average error down. A Google Research-led study found 17.2× trace-level error amplification versus one agent; centralized coordination cut it to 4.4×, not zero.
Every handoff is another place for drift, so the checks belong before the hop.
Six faces on a node. OPEN, CONDITIONAL, CLOSED. Interaction requires compatible identity, orientation, and constraint. If the face is closed, the edge does not form. Incompatibility is not an error you catch later.
Schema mismatch across agent identity is a named failure mode now. Two agents, two meanings of the same field, both schema-valid. The governed question is not “did it parse.” It is “was this assembly allowed to talk to that one.”
A Prompt Forge receipt stores a hash of the prompt and process measures, not the prompt as a database row. What that demonstrates, and what it does not.
We do not keep the prompt as a database row. The receipt stores a hash of the prompt. Higher tiers can back the workspace up to your own encrypted store. The browser seals it. We hold ciphertext. That is the procurement sentence.
Governance Wash records what a governed answer refused or qualified. It informs governance methods; it is not model training on your prompts.
98 claims in the public traceability map. 66 system, 32 method. Eight groups. If a slide names a method, it should point at a claim, not a metaphor.
Paper, public: “Governing Execution in Complex Intelligent Systems.” Stephen Joseph Voss. 26 Feb 2026. Provisionals 63/985,184 and 63/991,184. Prior-art disclosure, not a granted patent. Say pending.
Ungoverned: 82/100 and HOT, on invented revenue. Governed: INCOMPLETE and a discovery call. The Governance Wash diff shows which claims changed the action.