What's new since the previous development update

This preview focuses less on adding another large collection of isolated systems and more on beginning to expose the modern architecture to normal runtime use.

The biggest addition is a new optional runtime mode:

MODERN_EXPERIMENTAL

The modern system can now be enabled optionally

Adaptive Optimization now supports two runtime modes:

LEGACY

and:

MODERN_EXPERIMENTAL

The default remains:

LEGACY

so existing installations continue using the same behavior unless the user explicitly enables the experimental mode.

The setting is stored in:

runtime.properties

using:

runtime.mode=LEGACY

or:

runtime.mode=MODERN_EXPERIMENTAL

If the configuration is missing, invalid or cannot be read, AO automatically falls back to:

LEGACY

EXPERIMENTAL does not mean dangerous

MODERN_EXPERIMENTAL is called experimental because the modern optimization system is:

not fully connected yet

—not because enabling it is expected to damage a world or make AO unsafe to use normally.

The experimental runtime path remains heavily restricted.

It currently uses modern systems to:

• observe real runtime conditions;
• detect optimization opportunities;
• create modern optimization questions/hypotheses;
• validate provenance;
• exercise the modern promotion pipeline;
• collect diagnostics about what the modern system would do.

Failures are handled fail-safe, and invalid modern state does not automatically receive execution authority.

The current runtime diagnostics explicitly keep productive generated promotion disabled while an authoritative experiment-cost source is still missing.

AO will not invent that missing evidence simply to let a candidate pass.

The legacy optimizer is still active

Enabling:

MODERN_EXPERIMENTAL

does not mean AO suddenly abandons the legacy optimizer.

AO is still a hybrid system.

The practical optimization that already exists continues to run, while the modern architecture performs additional runtime analysis around it.

Conceptually, the current preview is closer to:

legacy optimization
+
modern runtime observation
+
modern hypothesis generation
+
modern safety/authority infrastructure

rather than:

legacy replaced completely by modern AO

That migration is still in progress.

Early testing can show additional performance improvement

Early runtime testing has shown that enabling the experimental modern path can produce additional performance improvements in some environments.

However, this should still be treated as:

experimental performance behavior

rather than a guaranteed FPS increase.

The legacy optimizer is still contributing significantly to current performance, and the modern system is not yet fully productive end-to-end.

Results may therefore vary depending on:

• modpack;
• world;
• runtime conditions;
• shaders;
• active workload;
• session length.

The purpose of this preview is partly to begin gathering more real runtime evidence about that behavior.

Modern runtime analysis now runs naturally during gameplay

When MODERN_EXPERIMENTAL is enabled, AO can periodically inspect the current runtime instead of requiring a manual development probe.

The modern path can now:

observe runtime
↓
detect optimization opportunities
↓
bind them to the current side/context
↓
construct modern candidate observations
↓
generate bounded optimization hypotheses
↓
evaluate the existing promotion architecture

This is one of the first steps toward making the modern optimizer participate naturally during normal Minecraft sessions.

Modern observation remains bounded

The experimental modern evaluation is rate-limited rather than running continuously every tick.

It also verifies that runtime observations carry the expected canonical provenance before allowing them further into the modern pipeline.

If provenance does not match, the operation is blocked.

Unexpected failures become:

FAILED_SAFE

instead of granting optimization authority.

Durable generated strategies now re-enter more safely after restart

This preview also improves the connection between:

Candidate Registry

and:

standalone durable StrategySpec storage

during startup.

AO now validates generated strategies against their durable standalone definitions before allowing them back into the runtime candidate catalog.

It can distinguish:

exact registry + durable strategy match
→ accepted

durable envelope missing but exactly reconstructible
→ accepted as reconstructible

corrupt durable strategy
→ blocked

registry / durable strategy conflict
→ blocked

invalid or incomplete registration
→ blocked

The bridge itself is read-only.

It does not silently rewrite corrupted or conflicting strategy history.

Restart-required strategies can no longer be accidentally hot-armed

Another important safety improvement closes a gap around generated strategies that require a Minecraft restart.

Previously, the general ARM path needed stronger enforcement of the difference between:

LIVE_RUNTIME

and:

RESTART_REQUIRED

That check now exists directly in the productive boundaries.

A strategy that is not valid for live runtime execution receives:

RESTART_REQUIRED_NOT_HOT_ARMED

instead of being armed as though it could safely execute immediately.

The protection is checked again before physical execution

The restart requirement is not checked only once.

AO checks compatibility:

before ARM

and again:

immediately before experiment start / physical execution

including after the final runtime recapture.

If the strategy no longer qualifies for hot runtime execution, AO:

• cancels the reservation;
• removes the live ARM;
• does not start the experiment;
• does not physically apply the strategy.

This further strengthens the rule:

restart-required means restart-required.

The modern mode remains intentionally incomplete

There is an important boundary in this preview.

The modern experimental runtime can now naturally reach:

runtime observation
↓
optimization opportunity
↓
generative reasoning
↓
strategy preview
↓
promotion evaluation

but the final productive generated-promotion path is still deliberately blocked where authoritative information is missing.

In particular, AO currently refuses to fabricate an experiment cost just to complete the transition.

So:

MODERN_EXPERIMENTAL is usable, but the modern optimizer is not finished.

That is exactly why it remains marked:

EXPERIMENTAL

and why:

LEGACY

remains the default.

How to enable it

The experimental modern runtime can be enabled by changing:

runtime.mode=LEGACY

to:

runtime.mode=MODERN_EXPERIMENTAL

inside AO's:

runtime.properties

configuration.

Restart Minecraft after changing the mode so AO can load the new runtime configuration cleanly.

Users who do not want to test the modern path do not need to change anything.

What this preview represents

The previous updates built much of the modern optimization architecture.

This preview begins answering a different question:

"Can normal players actually start using parts of it during real gameplay?"

The answer is now:

yes, experimentally.

The legacy optimizer remains the main practical foundation.

The modern architecture can now begin observing and reasoning alongside it during normal runtime.

Current state

Adaptive Optimization can now increasingly operate as:

proven legacy optimization
+
experimental modern runtime intelligence

The legacy system continues providing the optimization behavior that already works today.

The modern system is progressively adding:

runtime observation
↓
causal reasoning
↓
generated hypotheses
↓
durable strategy identity
↓
safe authority boundaries
↓
restart-aware execution
↓
future autonomous optimization

without pretending that unfinished pieces are already complete.

What remains

The next major goal is to continue moving from:

modern analysis and preview

toward:

fully productive modern optimization

while preserving:

• legacy performance improvements;
• rollback safety;
• evidence requirements;
• restart correctness;
• scope-aware authority;
• low runtime overhead.

The modern system must earn more control gradually rather than receiving it simply because the architecture exists.

Where AO is heading

Adaptive Optimization is now moving from:

a legacy optimizer surrounded by modern systems

toward:

a modern optimizer that can safely take over those responsibilities itself.

This preview is one of the first releases where users can explicitly choose to let that transition begin during normal gameplay.

LEGACY remains the stable default.

MODERN_EXPERIMENTAL is available for users who want to test the newer runtime architecture and potentially benefit from the additional performance observed during early testing while helping expose the remaining gaps before the modern system becomes the default.