What's new since the previous development update

This update continues the work started by MODERN_EXPERIMENTAL.

Rather than adding another large user-facing optimization feature, this preview focuses on strengthening the connection between:

generated strategies
↓
physical execution
↓
runtime measurements
↓
causal results
↓
future benchmark / learning evidence

The goal is to make sure that when the modern system eventually learns from a real generated experiment, it can prove exactly:

what was executed, where it was executed, and what was actually measured.

Generated experiments can now retain exact execution provenance

AO can now capture immutable provenance for a generated strategy immediately before its experiment begins.

For a valid v2 generated registration, the execution record can preserve:

• experiment identity;
• candidate identity;
• registration identity;
• exact StrategySpec identity;
• canonical runtime context;
• runtime side;
• adapter/source identity.

These facts are combined into a canonical identity so later systems can verify that they are still talking about the exact strategy that was physically executed.

This helps prevent a later candidate, regenerated definition or changed runtime registration from being mistaken for the original experiment.

Runtime measurements can now retain their original historical context

Experiment samples can now carry a historical runtime-evidence sidecar.

Instead of storing only:

"this metric had this value"

AO can preserve the runtime observation associated with that exact measurement sample.

This includes information such as:

• world identity;
• dimension;
• reload/lifecycle context;
• player/workload state;
• sampling state;
• metric availability;
• runtime session identity;
• server tick sequence;
• capture time.

The sample and its historical runtime observation are bound together using a canonical identity.

This is an important distinction:

AO no longer has to assume that the runtime state observed later is the same state that existed when a historical sample was measured.

Measurement windows can now preserve exact per-sample evidence

The PRE, EXPERIMENT and RECOVERY windows can now retain their corresponding historical runtime evidence alongside their normal measurements.

AO preserves the same ordering between:

measurement sample
and
historical runtime evidence

and requires a complete one-to-one relationship before treating that historical evidence as complete.

If the source cannot provide that provenance, AO does not manufacture it.

Legacy/unbound sample sources can still operate, but their missing historical provenance remains explicitly unavailable.

Completed runtime experiments now expose immutable measurement facts

The experiment coordinator can now expose a read-only representation of the most recently completed runtime experiment.

That completed measurement contains the real closed:

PRE

EXPERIMENT

and:

RECOVERY

windows, together with:

• experiment identity;
• causal record identity;
• PRE boundary;
• physical apply timestamp;
• final experiment verdict;
• learning eligibility.

AO also exposes the historical sample evidence associated with those windows when it is complete.

These projections are:

read-only

They do not start another experiment, perform new sampling, reconstruct old measurements or grant additional optimization authority.

Generated execution can now be matched to the experiment that actually happened

A new read-only binding layer has been added for completed generated experiments.

Its purpose is to prove that several previously separate pieces of evidence all describe the same real event:

generated candidate
+
exact v2 registration
+
StrategySpec
+
physical binding
+
execution provenance
+
causal record
+
completed measurement windows

The binding is accepted only when those identities remain consistent.

AO can reject the handoff when it detects conditions such as:

• missing completed measurement;
• missing generated execution snapshot;
• experiment identity mismatch;
• causal identity mismatch;
• missing historical v2 provenance;
• stale or replaced registration;
• strategy identity mismatch;
• runtime context / side mismatch;
• invalid physical binding;
• invalid measurement boundaries;
• degraded generated registry state.

This gives the modern system a much stronger basis for eventually turning a completed generated experiment into reusable benchmark and learning evidence.

Historical facts are not reconstructed from the current runtime

This work closes an important conceptual gap.

A future learning system should not do something like:

experiment finishes
↓
look at the runtime now
↓
assume that was the state during the experiment

The newer path increasingly works as:

capture historical state when the sample happens
↓
retain exact execution provenance
↓
close real measurement windows
↓
bind those historical facts together
↓
allow later learning only when identities match

This is especially important for long-running or highly dynamic modpacks where runtime conditions can change substantially during an experiment.

This does not grant new productive authority yet

These additions strengthen the evidence chain.

They do not mean that generated strategies now automatically receive full productive learning or promotion authority.

The new completed-experiment binding is currently a read-only foundation.

The broader modern path is still being connected, and the previous restriction remains important:

AO will not invent missing authoritative evidence just to complete the pipeline.

In particular, the modern generated-promotion path remains intentionally constrained where required authoritative inputs are still unavailable.

MODERN_EXPERIMENTAL remains experimental

This update does not change the runtime-mode policy introduced previously.

LEGACY

remains the default.

MODERN_EXPERIMENTAL

continues allowing more of the modern architecture to participate during real gameplay while the legacy optimizer remains responsible for much of AO's current practical optimization.

The modern system is still being connected gradually.

What this update represents

The previous preview began allowing modern AO to participate naturally during runtime.

This update begins strengthening what happens after a real experiment starts producing evidence.

The direction is increasingly becoming:

observe
↓
generate candidate
↓
prove candidate identity
↓
physically execute
↓
capture historical runtime evidence
↓
measure PRE / EXPERIMENT / RECOVERY
↓
produce causal outcome
↓
prove that all of those facts belong to the same experiment
↓
future benchmark / learning authority

The important part is that AO is moving toward learning from what actually happened rather than reconstructing what it thinks happened afterward.

Current state

Adaptive Optimization is continuing the transition from:

modern architecture around a legacy optimizer

toward:

a modern optimizer with a verifiable evidence chain from observation to execution to learning.

The newest work strengthens:

• generated execution identity;
• historical runtime provenance;
• measurement-window provenance;
• completed runtime experiment projections;
• generated strategy ↔ experiment binding;
• causal evidence continuity;
• fail-safe rejection of mismatched historical facts.

The system is still intentionally conservative.

Having the information available does not automatically grant authority to use it.

What remains

The next major step is to connect these stronger historical and execution bindings into the remaining productive modern learning/benchmark path.

That means continuing toward:

real generated experiment
↓
authoritative completed evidence
↓
benchmark / learning integration
↓
validated lifecycle consequence
↓
safe autonomous reuse

without bypassing:

• experiment-cost authority;
• causal evidence requirements;
• runtime scope;
• lifecycle correctness;
• rollback guarantees;
• provenance validation;
• restart safety.

The modern optimizer is getting closer to being able to prove not only that an optimization exists, but that:

this exact optimization produced this exact measured result under this exact runtime context.

That evidence chain is necessary before AO can safely give the modern system more autonomous control.