A AEGIS containment evidence
AEGIS overview
Evidence Console v2.0 · July 29, 2026

Registered disturbance containment · Qwen + Llama · approximately 1B–4B

Detect the disturbance.
Limit the damage.

AEGIS reduced downstream damage in prospective matched studies across two model families. It remained inactive in the registered clean-control cohort, transferred through Qwen3-4B, and showed clear limits as disturbance severity increased.

Commercial modelConditional

Value comes from avoided restart, rollback, recovery, or degradation exposure when a qualifying incident occurs.No universal savings rate is asserted.

prospective matched evidence

2×

model families

fresh-seed confirmation through 4B parameters
5 / 5Qwen pairs benefited at primary containment milestones
5 / 5Llama pairs benefited at every reported milestone
3 / 3Qwen3-4B pairs showed downstream benefit
0 / 5clean runs authorized an intervention

Architecture and scale transfer

One monitoring-and-control loop, tested across model families and scale.

Evidence boundaryApproximately 1B → 4Btwo model families · ordinary schedules

The frozen system detected and acted without model-specific recalibration in the registered transfer studies. This supports bounded scale tolerance—not parameter-count invariance, distributed-training readiness, or frontier-scale validation.

Outcome discipline: the cards retain each study's own registered endpoint. They are not pooled into one synthetic “success rate,” because milestone loss gaps and damage-area reduction are related but not identical estimands.

Dose response, opened up

Containment is strongest before the actuator saturates.

Prospective Qwen dose confirmation

Residual downstream damage

Shadow damage Governed damage

Bars show the registered downstream damage-area metric. Smaller is better. The result establishes containment of optimizer damage; it does not mean AEGIS repairs the source data.

Operational integration

One health layer. Your intervention contract.

01

Observe

AEGIS monitors training health in real time through proprietary model-internal signals alongside ordinary run telemetry.

02

Route

The agreed authorization gate routes a qualifying event into the operator-selected response contract.

03

Respond

Choose alert, checkpoint, quarantine, halt, adaptive braking, or a staged combination. Operator authority remains intact.

Integrated demonstration: adaptive braking proves that monitoring, authorization, and bounded response can operate as one system; it is one policy option, not the prescribed deployment model. The frozen gate also authorized zero interventions across five clean controls and five difficult-but-valid data shifts.

Conditional economic translation

Estimate avoided incident exposure—not generic training efficiency.

The model is intentionally simple: annual training compute budget × the share exposed to qualifying incidents × the share of that exposure a validated containment policy could avoid. The last two inputs belong to the operator and must be measured in a pilot.

Illustrative avoided exposure$625Kscenario only; not an observed AEGIS savings rate or frontier-scale forecast

Trace the claim

Aggregate receipts with explicit limits

Download manifest
Frozen public evidence identity Technical evaluation or shadow-mode pilot: founder@arcafutura.com
Supported: early detection and bounded containment for a registered high-loss disturbance class. Open: low-loss harmful regimes, reliable recurrent protection, distributed training, frontier scale, and customer-specific avoided-cost validation.

Claim boundary

What this proves—and what it doesn’t.

AEGIS shows that real-time monitoring, authorization, and bounded response can work as one system. The operator chooses how that system responds.

Prospective matched studies support early detection and reduction of downstream optimizer damage for a registered high-loss disturbance class across Qwen and Llama, with transfer through Qwen3-4B.