Prospective sensing · separate authorization · bounded response
AEGIS has been tested as a separable training-control system: sense a state transition, decide whether it warrants action, then apply a bounded response. The strongest evidence is end-to-end containment for abrupt high-loss disturbances across Qwen and Llama. The current policy is a utility demonstration—not an optimal controller.
Value comes from avoided restart, rollback, recovery, or degradation exposure when a qualifying incident occurs.No universal savings rate is asserted.
one governed system
3×What the evidence validates
The research program now tests three different questions independently. A useful monitor must notice a transition, distinguish observation from authorization, and demonstrate that a bounded response can change downstream outcomes.
Claim discipline: these layers are related, but they are not interchangeable. Detection does not automatically authorize intervention; containment does not prove the detector chose the optimal time; observed specificity does not prove universal safety.
Confirmed containment cohorts
The frozen system detected and acted in the registered transfer studies. This supports bounded architecture and scale tolerance—not parameter-count invariance, universal model agnosticism, distributed-training efficacy, 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.
Denominator discipline: this console does not headline a cumulative run total. Historical runs, development smokes, reused controls, prospective arms, and confirmation seeds overlap in ways that make one grand number easy to misread.
Dose response, opened up
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.
Exploratory transfer—not pooled confirmation
New full-parameter 7B and vision studies add useful boundary evidence. They do not enter the 13-pair confirmed headline because their gates, confounds, and endpoints differ.
Registered negative results
AEGIS is not being described as optimal because the repository contains direct evidence against that stronger story. These tests constrain what the system can honestly be said to do.
Operational integration
AEGIS monitors training health in real time through proprietary model-internal signals alongside ordinary run telemetry.
The agreed authorization gate routes a qualifying event into the operator-selected response contract.
Choose alert, checkpoint, quarantine, halt, adaptive braking, or a staged combination. Operator authority remains intact.
Conditional economic translation
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.
Trace the claim