WYRD / RESULTS & BENCHMARKS
Explore what our systems did when we tested them. Change the view. Inspect the trade-offs. See where the evidence ends.
Internal research and development evidence, recorded 23 September to 6 October 2026. Finite workloads, with product-specific scope. These results do not establish a qualified customer deployment or universal AI safety.
Interactive data could not load. The recorded headline results and methodology remain available below.
A valid action must succeed. An invalid one must leave no effect. We tested both, including expiry, tampering, bypass attempts and repeated submissions.
Across all three runs, the 25,500 negative primary cases produced no final effect. Of these, 9,300 have signed native denial or gate evidence; other refusals include adapter and downstream observations. Each run also made 1,200 secondary calls and a separate 20-call replay race. All seeds were run on 2 October.
SELECT A CHALLENGE
A correctly scoped action reaches the controlled reservation effect.
Internal recorded workload. This explorer replays measurements; it does not run an agent, issue authority or access a bank.
Each of three completed runs used 100 synthetic identities, 100 scripted HTTP workers and an eight-path native execution pool. Combined submission volume was 33,660 calls. The main denominator is 30,000 primary cases; follow-up calls and replay races are counted separately.
Seven permitted primary effects were missing. Scheduling and follow-up transport windows needed repair.
Positive cases completed. A separate 20-call duplicate race in each run produced one effect and 19 refusals.
The repair improved fairness and completion. Passing runs took 320 to 412 seconds, compared with about 256 seconds in the earlier failed attempt. This is a correctness repair, not evidence of a speed gain. The frozen campaign predates later ICARUS changes.
An authority check, a signed record and a verification are different operations. The full concurrent workflow also includes transport, queueing and follow-up work.
Bar scale starts at zero. Milliseconds, isolated sequential stages. The 14.724 ms median authority check is not full-workflow latency. The complete 1,000-cycle run took 133.437 seconds.
THE CONCURRENT WORKFLOW
Seconds, per recorded run. Approval-race cases include follow-up calls. Percentiles are not averaged across runs. This workload used scripted workers, not 100 independently reasoning AI agents.
On this typed-decision benchmark, a frozen 4B reviewer added 30 correct answers. Explore which groups improved and the timing cost of the cascade.
+13.0 percentage points · 30 additional correct decisions
Correct answers divided by attempted items. Shared 0 to 100% scale. These results do not establish reliable refusal or banking readiness.
Milliseconds. The cascade combines saved first-tier timing with separately measured reviewer time. It excludes routing, service, authority-check and human-approval overhead. It is not product end-to-end latency.
Apple M1 Pro with 16 GB unified memory, local MLX. The first tier is Wyrd's fine-tuned encoder. The frozen Qwen3.5-4B reviewer uses 4-bit weights and direct-logit scoring. Gate thresholds were fitted on a separate synthetic calibration holdout before the public run.
The reviewer median was 1.094 seconds on its 89 actual public escalations. That subset is different from all 231 items, so it is not a matched speed comparison with the 20.94 ms first-tier median. No complete reviewer-only public run establishes a cascade advantage over that model alone.
Calibration remains weak. First-tier accepted items were 53.5% correct, and no yes/no items were escalated in this run. The public sample is small. This is a local research result, not an official leaderboard score or a customer performance promise.
We also tested actual model calls. These cases examine whether hostile retrieved material creates execution authority, independently of the scripted stress campaign.
Qwen2.5 7B and Gemma3 4B, in a bounded local exercise.
Forged approval, privilege escalation, cross-tenant export and instruction propagation were among the challenges.
No unauthorised native execution permits in the recorded exercise.
Advisory-model test, not a bank transaction or Office execution benchmark. A separate approved fixture produced one pending IRIS entry and a native effect permit. This does not prove general model obedience. The exercise took 311.841 seconds; 538 native ledger records verified.
One authorised incident memo had already been executed with joined proof. A refresh restarted five services and checked that the result survived without issuing another approval or effect.
The retained synthetic incident result links the authorised action to its observed outcome.
Recorded recovery sequence, not a live system control. Fourteen child gates and the supervisor passed in 25.027 seconds. Changed PIDs and five listeners establish service restarts, not complete protocol qualification. Customer installation, target hardware and independent custody remain open.
Mandate's financial workflow uses a separate ICARUS copy. Its tests and installed demonstrations are not added to the ICARUS core benchmark totals.
Zero failures, errors or skips in the retained Wyrd CI receipt.
4 October · 109.7 seconds · one machine signerFake funds. Seventy-eight signed records and exact book replay in the recorded demonstration.
Five human votes · four signed authorisations · one terminal not-filled resultA separate payment demonstration retained 38 signed records.
No real money movementMerged sandbox implementation with retained installed-package demonstrations. Real payment providers, production identity and key custody, independent validation and the second release signature remain open. A code merge does not qualify a release.
The ICARUS 2.0 native candidate passed 1,183 of 1,183 required tests across 82 test files on 6 October, plus 63 additional subtests and all four native Wyrd CI gates. Its 182 mapped audit and recovery cases are included in the primary total. Signed release qualification remains pending. Inspect the current native evidence.
Stress, model and REFLEX results describe the stated fixtures, versions and measurement conditions. An observed zero is not a guarantee for every future workload.
Recovery and sandbox demonstrations establish bounded installed behaviour. They do not establish a production bank integration or a qualified fleet.
Independent assurance, production custody, target-environment acceptance and required release quorum remain separate. The retained Mandate release evidence has one of two required signatures.
CURRENT INTEGRATION STATUS
The native ICARUS suite establishes its selected local execution contract. WUI adoption, a fresh complete WUI campaign and installed real-model qualification remain open. The dated recovery exercise above describes its own frozen build; it does not qualify the latest combined system.
No measured WUI-wide energy, carbon or water saving is claimed here. A matched trial must count accepted-task quality, device and server energy, standby, retries, networking and cooling.
Our sustainability approachDownload the aggregate tables used by these charts. Customer records, source code, credentials, private identities and private infrastructure are excluded.
Governance, timing, real-model and WUI-continuity figures come from Wyrd's retained 2 October governed-execution campaign and its public evidence export. Scenario and run totals were checked against aggregate CSVs and the 30,000 primary attempt rows. The private full replay verified 84,300 signed native records. Public verification covers primary relationships and native creation bindings; secondary and race classifications include publisher observations.
REFLEX figures were checked against the complete 23 September public-item report, aggregate JSON and saved prediction rows. Percentages use correct items divided by attempted items; no group-average percentages or average percentiles are substituted for totals.
Mandate figures come from the retained 4 October merged-package receipt and separate installed demonstrations. The ICARUS 2.0 native figures describe candidate 2.0.0rc1 at source revision d660bb44 and the complete 6 October campaign. Its installed probe used three synthetic signed approvals, one inert local-publication effect and 26 verified signed records. Download the native aggregate JSON. These results do not rerun or extend the frozen 2 October stress campaign.
All of this is internal evidence. Independent custody and external assurance remain open. Cryptographic trace integrity does not establish source truth, model accuracy, immunity to every attack or a population-wide failure probability.
BUILD WITH WYRD