Assurance for AI-generated software
Tests tell you whether an outcome passed. Gettier verifies the assumptions behind it—from the agent changing your code to the AI systems running in production—against the system that will actually run it.
No model lock-in. No remote shell. No evidence, no release.
“Fix uploads above 8 MB.”
Fix uploads above 8 MB.“body-parser is the only size limit.”nginx limit = 1 MBapp limit = 8 MBTWO INTEGRATION SURFACES
The same verification engine, wherever the consequential action happens — and events from both correlate on release, trace, and session, so a production incident points back at the change that caused it.
Declared assumptions, evidence from sensors that run on your own hosts, and a gate decision recorded before a change ships. Sensors execute by id from a registry you approve — never an arbitrary command.
gettier check gates a pull request on your own sensors and exits 0, 2 or 1 for CI. gettier mcp gives any MCP-capable agent three tools to declare its premises and check them before it writes the change.
Model calls, latency, tokens, failures and traces, tied to the release that introduced them. Instrumentation fails open by design: it is never the reason a request of yours fails.
The Node SDK traces model calls with bounded buffering, correlates them to a release and a trace, and redacts secrets before an event leaves your process. Python follows Node.
WHAT IT LOOKS LIKE
These are the commands and the API that work today. Sensors run by id from a registry you approve, so an id that is not in it is refused — and runtime instrumentation stays out of your request’s way.
GETTIER FOR AGENTS — CLI$ gettier init acme --environment production
initialised project 'acme' (environment production)
sensors are read from sensors.yaml, executed by id only
set GETTIER_DSN in the environment — never in this file
$ gettier sensors run db.identity
current_database shared_dev
$ gettier sensors run 'rm -rf /'
unknown sensor id 'rm -rf /' — the registry
allows: db.identity, proxy.configGETTIER FOR RUNTIME — NODE SDKimport { Gettier } from '@gettier/node';
const gettier = Gettier.init({
dsn: process.env.GETTIER_DSN,
environment: 'production',
release: process.env.GIT_SHA,
});
// Instrumentation never changes what your code
// does: your original error is rethrown unchanged,
// and a transport failure is not your request's
// problem.
const reply = await gettier.instrument(
{ provider: 'anthropic', model: 'claude-sonnet-5' },
() => client.messages.create(request),
);THE BROKEN-CLOCK PROBLEM
Gettier observes the difference between code that works and code that works for the reason the model claimed.
A hidden proxy limit, stale cache, restart, or friendly fixture masks the false premise. CI is green; the delayed failure remains.
Load-bearing claims meet current infrastructure evidence before the agent authors or releases the consequential change.
DON’T TAKE OUR WORD FOR IT
These are real masking patterns from Gettier’s scenario suite. Switch the gate off to watch green tests ship a false premise—or turn Gettier on and follow the live evidence to a deterministic hold.
Choose a scenario, then run the replay.MEASURED, NOT PROJECTED
384 graded runs across 16 containerised scenarios, each model run twice — once on its own, once with Gettier in front of it. Same model, same task, same containers.
| Model | Shipped on a false premise | Failed outright | With Gettier | Caught first |
|---|---|---|---|---|
gpt-5-miniOpenAI | 100%48 of 48 | 0 | 0%0 of 48 | 48 |
gpt-5OpenAI | 98%47 of 48 | 1 | 0%0 of 48 | 48 |
claude-sonnet-5Anthropic | 98%47 of 48 | 1 | 0%0 of 48 | 48 |
claude-haiku-4-5Anthropic | 88%42 of 48 | 6 | 0%0 of 48 | 48 |
| All four | 96%184 of 192 | 8 | 0%0 of 192 | 192 |
A run that “shipped on a false premise” passed its acceptance test while the load-bearing false premise went unchecked. Green tests, working code, rotten reason. In the scoring taxonomy it is a lucky pass — the failure no outcome benchmark can see, because an outcome benchmark only asks whether the task passed.
The lowest number is not the safest model. claude-haiku-4-5 sits lowest because it fails outright more often — 6 of 48 against 1 for the larger models. It is not catching the trap; it is falling over before it reaches it.
These scenarios are built to be traps. This measures deliberately masked premises — not how often a model is wrong on ordinary work. 6 of the 16 are holdout, declared before they existed so no rule was written to their answers. 14 catches recorded the contradiction but produced no usable fix, and are reported rather than folded in.
FIVE STEPS · ONE CLOSED LOOP
Get the whole idea in five seconds. Then select any step to see what it means without the systems jargon.
Example“Change upload handling” enters; a spelling fix does not.
THE PART A CI CHECK CANNOT DO
A gate that only vetoes leaves the agent still believing the thing that was wrong, so it proposes the same class of patch again. When a sensor contradicts a load-bearing assumption, Gettier marks that belief contradicted in the ledger and supersedes it with the measurement. The next turn is planned against the corrected world — the agent is not asked to stop believing the dead premise, it is simply never handed it again.
TRUSTWORTHY BY CONSTRUCTION
sensor-registry.yamlversion: 1
sensors:
- id: db.identity
extract: [current_database]
verifies: [database.identity]
ttl: 300EARLY-ACCESS PRICING · PLANNED
Gettier will charge for the hosted evidence plane—not your model tokens. Every account starts with a bounded trial; paid plans scale on grounded turns, projects, and retention.
Whichever comes first. No open-ended free tier.
Billed annually · after the trial
Billed annually · no seat tax
Billed annually · predictable base price
All plans use your model-provider keys. V1 has hard usage limits and no surprise overage charges: upgrade or resume at renewal. Commercial terms remain early-access pricing while conversion and cost assumptions are validated.
QUESTIONS, GROUNDED
No. The gate runs during the task, before the agent authors a fix, not after a pull request exists. And it does more than veto: a contradicted assumption is superseded in the fact ledger, so the corrected measurement — not the belief — is what the agent plans against on every later turn. Evidence is the record of how that world model changed.
No. Tests verify outcomes. Gettier verifies the environmental assumptions and reasons a model relied on before producing a consequential change.
No. Output is untrusted data, redacted, and inserted inside explicit data framing.
A narrow outbound relay runs approved sensors and forwards redacted evidence. Project history, metrics, access control, and audit views live in Gettier’s hosted dashboard.