Adaptive Optimization — New Preview

The previous preview completed the first safe bridge between Adaptive Optimization's modern causal architecture and real optimization behavior.

AO could already:

Remember causal experiments
↓
Build Candidate Knowledge
↓
Evaluate Candidate Lifecycle
↓
Publish Coherent Decision Input
↓
Understand current context
↓
Evaluate historical applicability
↓
Resolve exact executable candidates
↓
Verify the physical implementation
↓
Evaluate manual experiment admission
↓
Create a single-use AdmissionLease
↓
Revalidate immediately before execution
↓
Run one real controlled experiment
↓
Rollback
↓
Update causal knowledge

But the previous preview intentionally left one major question unanswered:

Which candidate should AO consider next?

The user still had to manually choose the candidate.

The modern architecture could determine whether that manually requested experiment was safe enough to attempt, but it could not yet perform a modern comparison between available candidates.

And even after selection existed, another problem would remain:

A recommendation is not authority.

A system that believes a candidate looks promising must not automatically be allowed to:

• promote it;
• trust it;
• reject it;
• quarantine it;
• supersede another candidate;
• arm an experiment;
• apply an optimization.

This preview implements and validates both of those missing boundaries.

Adaptive Optimization now has:

Modern Candidate Selection
↓
Read-only recommendation
↓
Modern Decision Evidence
↓
Promotion / disposition policy
↓
Explicit Decision
↓
Durable Decision Authority
↓
Effective Candidate Disposition

For the first time, AO can not only preserve causal knowledge and evaluate whether a requested experiment is admissible.

It can now also:

compare what it currently knows about candidates

and:

persist an explicit modern authority decision without confusing recommendation with permission.

Modern Candidate Selection

The previous preview ended with this question:

"How should AO decide what deserves experimentation?"

This preview introduces the first modern Candidate Selection system designed to answer that question.

The selector consumes the modern causal state rather than relying on a simple static priority list.

Candidate selection can reason from information such as:

• Coherent Decision Input;
• Candidate Knowledge;
• Candidate Lifecycle;
• current runtime context;
• Context Applicability;
• candidate resolution;
• evidence quality;
• evidence quantity;
• contradictions;
• lifecycle disposition;
• safety information;
• rollback capability;
• current execution relevance.

The purpose is not to immediately execute an optimization.

The purpose is to answer a narrower question:

"Given what AO currently knows, is there a candidate that deserves consideration?"

Selection is not execution

This separation is fundamental.

The new selector is deliberately:

• deterministic;
• read-only;
• non-authoritative.

Running Candidate Selection does not:

• create an ARM;
• create an AdmissionLease;
• start an experiment;
• acquire the optimization lane;
• apply a physical optimization;
• write a causal verdict;
• promote a candidate;
• change lifecycle authority.

Conceptually:

Candidate Selection
↓
Recommendation only

not:

Candidate Selection
↓
Physical execution

The admission architecture introduced in the previous preview remains a separate safety boundary.

A future autonomous system must still preserve:

Selection
↓
Admission
↓
Lease
↓
Fresh Revalidation
↓
Lane acquisition
↓
Physical execution

Selection never becomes a shortcut around those stages.

NO_ACTION remains a first-class result

Adaptive Optimization is not designed around the assumption that:

something must always be optimized.

A candidate selection system that is forced to choose something would eventually manufacture experiments simply because candidates exist.

That would be dangerous.

The modern selection architecture therefore preserves an important principle:

doing nothing can be the correct decision.

If the available evidence does not justify choosing a candidate, AO can refuse to produce an actionable selection.

This allows the system to preserve uncertainty instead of turning uncertainty into unnecessary experimentation.

Insufficient information is a valid result

The selector does not convert missing evidence into artificial confidence.

One possible result is:

INSUFFICIENT_INFORMATION

This means AO currently does not possess enough trustworthy information to justify selecting a candidate.

That result is useful.

It tells the rest of the architecture:

"I do not know enough yet."

instead of:

"Choose the least bad candidate anyway."

This continues the same philosophy introduced by Context Applicability:

missing information is not permission.

Selection does not reward evidence quantity alone

A candidate with more historical records is not automatically the best candidate.

Modern selection is designed around causal usefulness rather than simple popularity.

Evidence must eventually be interpreted through dimensions such as:

• relevance to the current problem;
• contextual applicability;
• causal quality;
• contradictions;
• safety;
• repeatability;
• lifecycle state;
• rollback capability;
• experiment cost;
• information value.

A candidate that has been tested many times in irrelevant environments should not automatically outrank a candidate with fewer but more applicable observations.

Likewise, a candidate with a large measured improvement but unresolved hazards should not automatically become preferred.

Selection remains compatible with incomplete knowledge

Adaptive Optimization is still being built toward open-world discovery.

The selector therefore cannot assume that the current Candidate Catalog represents every possible optimization AO will ever discover.

Known candidates are the current searchable population.

They are not intended to permanently define the optimization search space.

This distinction becomes increasingly important as AO later gains:

• candidate variants;
• descendants;
• strategy combinations;
• newly discovered optimization families;
• automatically generated controlled candidates.

Selection operates on what AO currently knows.

It does not declare the search space complete.

Modern Candidate Selection was validated in Minecraft

The new selector was validated in the real Forge 1.20.1 runtime used for Adaptive Optimization development.

The runtime produced the conservative result:

INSUFFICIENT_INFORMATION

No candidate was activated.

No physical experiment followed.

Most importantly, evaluating selection caused:

0 automatic ARM operations

0 AdmissionLeases

0 experiment starts

0 physical applications

and no new authority writes.

This demonstrated that Modern Candidate Selection can run inside the real Minecraft environment while remaining observational.

The selector did not invent an experiment merely to prove that selection worked.

The conservative result was the correct result

An INSUFFICIENT_INFORMATION result is not considered a failed validation.

The objective of the runtime test was not:

"make AO select something."

The objective was:

"prove that AO can evaluate the modern candidate population without creating side effects and without forcing a choice."

That boundary was preserved.

This is especially important for future autonomous operation.

Adaptive Optimization must be allowed to decide that the value of another experiment is currently too low.

Candidate Selection completed the recommendation layer

After this preview, the modern causal architecture can now reach:

Causal Ledger
↓
Candidate Knowledge
↓
Candidate Lifecycle
↓
Coherent Decision Input
↓
Current Context
↓
Applicability
↓
Candidate Resolution
↓
Modern Candidate Selection
↓
Recommendation

But recommendation creates a new architectural problem.

Who is allowed to change the candidate's actual disposition?

A recommendation that says:

"this candidate looks trustworthy"

cannot itself become:

TRUSTED

Likewise:

"this candidate appears dangerous"

cannot silently become:

QUARANTINED

Recommendation and authority need separate identities, evidence and persistence.

This preview therefore introduces the next layer.

Modern Promotion & Decision Authority

Adaptive Optimization now has a formal modern authority path for candidate disposition decisions.

The architecture separates:

Evidence

from:

Recommendation

from:

Decision

from:

Durable Authority

Conceptually:

Causal evidence
↓
ModernDecisionEvidence
↓
Recommendation
↓
Decision policy
↓
Explicit ModernCandidateDecision
↓
Durable authority event
↓
Effective disposition

This prevents an internal recommendation from silently becoming permission.

Candidate dispositions are now more expressive

The modern Candidate Lifecycle can represent the following dispositions:

• AVAILABLE
• EXPERIMENTAL
• TRUSTED
• REJECTED
• QUARANTINED
• SUPERSEDED

These states describe very different conclusions.

AVAILABLE

The candidate exists and can be considered, but no stronger authority has been established.

EXPERIMENTAL

The candidate has explicit authority to remain within controlled experimental evaluation.

This is not equivalent to trust.

TRUSTED

Current evidence is strong enough to support a much stronger positive disposition.

Trust is still revocable.

REJECTED

Current authority has concluded that the candidate should not proceed under the evaluated conditions.

QUARANTINED

The candidate has a safety or correctness concern serious enough that it should remain isolated from normal consideration until that concern is resolved.

SUPERSEDED

A different candidate or descendant has explicitly replaced the candidate's role.

Supersession is an authority relationship, not simple age.

EXPERIMENTAL is not TRUSTED

This distinction is especially important.

A candidate being allowed to participate in controlled experiments does not mean:

the optimization is proven.

EXPERIMENTAL exists specifically so AO can represent:

"this candidate is permitted to remain under controlled investigation."

without claiming:

"this candidate is safe and beneficial enough to trust."

The runtime validation in this preview deliberately exercised that distinction.

ModernDecisionEvidence

Promotion decisions require their own evidence boundary.

This preview introduces immutable ModernDecisionEvidence.

The object represents the evidence presented to the modern disposition policy.

It is:

• canonical;
• descriptive;
• source-bound;
• non-authoritative.

The evidence object can describe the state from which a recommendation was produced without itself modifying that state.

This continues an architectural rule used throughout the modern AO pipeline:

information about authority is not authority.

Decision evidence is bound to coherent sources

Modern promotion decisions must not be created from an accidental mixture of unrelated generations.

Decision evidence is therefore tied to the coherent modern state used to construct it.

That allows AO to preserve the relationship between:

• the candidate;
• the causal evidence;
• Candidate Knowledge;
• Candidate Lifecycle;
• coherent source identities;
• the recommendation;
• the final explicit decision.

Future audits can therefore ask:

"What exact evidence justified this authority event?"

rather than only seeing the final disposition.

Recommendation remains non-authoritative

This preview establishes one of the most important safety rules for future autonomous AO:

Recommendation cannot modify authority by itself.

A recommendation may conclude that a candidate appears:

• available;
• experimental;
• trustworthy;
• rejectable;
• quarantine-worthy;
• superseded.

But the recommendation remains descriptive.

The effective lifecycle disposition changes only through the explicit authority path.

This allows future systems to become increasingly intelligent without giving every analytical subsystem direct permission to mutate candidate state.

Promotion policy

Adaptive Optimization now has a pure decision policy between evidence and authority.

The policy can produce outcomes such as:

• ALLOW
• DEFER
• BLOCK
• NO_OP

These results describe whether the requested disposition transition is currently justified.

ALLOW

The requested transition has enough explicit positive support.

DEFER

The requested transition may become valid later, but current evidence is insufficient or temporarily unsuitable.

BLOCK

A structural, contradictory or safety condition prevents the transition.

NO_OP

No authoritative change is necessary.

This prevents AO from manufacturing journal events when the effective state already represents the correct result.

Positive evidence is required for TRUSTED

TRUSTED is intentionally difficult to obtain.

The policy does not define trust as:

"nothing bad has happened yet."

A positive trust decision requires affirmative evidence.

Relevant requirements include causal support for areas such as:

• measured positive effect;
• repeatability;
• contextual applicability;
• semantic correctness;
• safety;
• rollback reliability;
• unresolved hazards.

A candidate cannot become trusted simply because:

• it exists;
• it has been tested once;
• its experiment completed;
• the game did not crash;
• a recommendation did not find a reason to reject it.

Trust requires positive justification.

Successful experiment execution is not automatic promotion

The previous preview already demonstrated a real physical FBB experiment that completed successfully but finalized:

INCONCLUSIVE

with:

NOT_ELIGIBLE

That principle remains intact.

The new promotion system does not reinterpret:

experiment completed successfully

as:

optimization proved beneficial.

Experiment infrastructure success and optimization evidence remain separate concepts.

Safety hazards remain stronger than performance optimism

Future AO may eventually discover candidates with very large measured performance improvements.

That cannot become an excuse to ignore:

• semantic breakage;
• compatibility failures;
• unsafe rollback;
• delayed regressions;
• contradictory evidence;
• unresolved safety hazards.

The modern promotion architecture preserves the rule that:

correctness and safety dominate raw optimization gain.

This is required before AO can eventually evaluate multidimensional Optimization Value.

ModernCandidateDecision

When a disposition transition is explicitly authorized, AO creates a canonical ModernCandidateDecision.

The decision preserves information such as:

• candidate identity;
• previous disposition;
• requested/new disposition;
• decision identity;
• evidence identity;
• coherent source identities;
• actor;
• reason;
• context;
• hazards;
• successor information when relevant.

A decision is therefore not just:

TRUSTED = true

or:

QUARANTINED = true

It is a traceable authority event with lineage.

Decision identities are deterministic

Modern decisions have stable deterministic identities derived from their meaningful contents.

This helps AO distinguish:

• the same decision reconstructed later;
• a genuinely new decision;
• duplicate authority writes;
• conflicting decisions.

The decision identity used during runtime validation was:

sha256:73b73ab443dc1b45fb8de9f57c45be93eab610de2a502b7dbb5c15ed3bfcce98

This identity remained stable after restart.

Durable authority remains event-based

The new modern promotion layer does not replace its authority with a mutable final-state file.

Disposition authority remains durable through:

candidate-lifecycle-events.jsonl

This journal is the authoritative transition history.

Materialized Candidate Lifecycle state can be rebuilt from:

Candidate Knowledge
+
durable disposition events

Conceptually:

authoritative causal evidence
+
authoritative lifecycle events
↓
rebuildable lifecycle view

This is important because a corrupted or missing derived lifecycle file should not erase the history of why a candidate reached its effective disposition.

Materialized state is not authority

Adaptive Optimization continues to distinguish:

source of truth

from:

convenient reconstructed state.

Candidate Lifecycle can expose the effective disposition for fast runtime use.

But the durable event history remains the authority boundary.

This allows derived state to be deleted and rebuilt without inventing a new decision.

Recommendation and effective disposition can disagree

This preview validates an intentionally important case.

A recommendation may currently say:

AVAILABLE

while durable authority says:

EXPERIMENTAL

The correct effective result is:

EXPERIMENTAL

because:

authority dominates recommendation.

The recommendation describes what the current analytical layer concludes.

The authority journal describes what disposition has explicitly been established.

The two are not required to be identical.

Manual authority comes before automatic authority

Just as the previous preview introduced modern experiment authority through a manual request before autonomous experimentation, this preview introduces modern promotion authority conservatively.

The new decision path does not automatically promote candidates.

The runtime test used an explicit manual modern decision.

There is currently no background process continuously changing candidate dispositions.

This provides a controlled proof that:

• evidence can be constructed;
• policy can evaluate it;
• a decision can be created;
• authority can be persisted;
• lifecycle can materialize it;
• restart can reconstruct it;

before future autonomous promotion is considered.

The first modern durable disposition transition

The runtime validation used the existing FBB causal pilot candidate:

sha256:77e2d5919463a8a3c57a38f045d2585799a0ed00624c3084db68c10f6a701f80

The explicit authority transition was:

AVAILABLE
↓
EXPERIMENTAL

Reason:

MANUAL_EXPERIMENTAL_REVIEW

The resulting modern decision/event identity was:

sha256:73b73ab443dc1b45fb8de9f57c45be93eab610de2a502b7dbb5c15ed3bfcce98

After the decision:

recommendation
AVAILABLE

while:

effective disposition
EXPERIMENTAL

This was intentional.

It demonstrated that the recommendation layer could not overwrite the explicitly established authority.

The authority journal remained minimal

After the runtime decision, the durable authority journal contained exactly one modern authority event.

It did not produce duplicate decisions.

It did not append new events merely because the read model was reconstructed.

It did not rewrite the decision during restart.

The durable lifecycle authority therefore represented:

one explicit decision

not:

one decision plus reconstruction noise.

Promotion did not trigger execution

Changing the candidate disposition from:

AVAILABLE

to:

EXPERIMENTAL

did not physically execute the candidate.

No automatic chain occurred such as:

EXPERIMENTAL
↓
ARM
↓
Lease
↓
Session
↓
Apply

The runtime remained non-autonomous.

The new authority event changed the modern candidate's disposition.

It did not grant itself physical execution.

No automatic TRUSTED transition occurred

The runtime validation also confirmed that AO did not continue promoting the candidate after the manual EXPERIMENTAL decision.

There was no automatic:

EXPERIMENTAL
↓
TRUSTED

transition.

The candidate remained exactly where explicit authority placed it.

This is important because the current FBB evidence does not justify trust.

The previous controlled experiment was inconclusive.

AO preserved that uncertainty.

Candidate authority survived restart

A clean Minecraft restart was performed after the modern disposition decision.

The new JVM reconstructed the modern authority state from the durable journal.

After restart:

authority event count
1

effective disposition
EXPERIMENTAL

recommendation
AVAILABLE

last decision identity
sha256:73b73ab443dc1b45fb8de9f57c45be93eab610de2a502b7dbb5c15ed3bfcce98

The authority event was not duplicated.

The decision was not replayed as a new action.

The recommendation did not erase the durable disposition.

Restart did not replay operational behavior

Recovering the modern decision did not cause AO to:

• ARM the FBB experiment;
• create an AdmissionLease;
• acquire the intervention lane;
• start an experiment session;
• apply the physical FBB optimization;
• write legacy learning;
• create another promotion event.

The new JVM returned to its normal observational runtime state.

The persisted authority was reconstructed as knowledge.

It was not replayed as an action.

Modern authority remains separate from legacy operational authority

This preview does not remove the legacy optimization architecture.

That migration has not happened yet.

The modern system can now:

• build causal knowledge;
• evaluate lifecycle;
• create coherent decision input;
• evaluate applicability;
• resolve candidates;
• control manual admission;
• select candidates read-only;
• evaluate promotion evidence;
• persist explicit modern disposition authority.

But AO has not yet transferred all operational ownership from legacy systems to the modern architecture.

This distinction is intentional.

Deleting or bypassing legacy authority before the modern system can safely replace its responsibilities would create a dangerous transition.

No hidden automatic activation

The complete runtime validation preserved the following boundaries.

Modern Candidate Selection did not create:

• ARM;
• lease;
• session;
• lane ownership;
• physical apply.

Modern Promotion Authority did not create:

• ARM;
• lease;
• session;
• lane ownership;
• physical apply.

Restart recovery did not create:

• ARM;
• lease;
• session;
• lane ownership;
• physical apply.

Candidate disposition and physical execution remain separate systems.

The modern architecture now has both recommendation and authority

The modern pipeline can now conceptually reach:

Observation
↓
Causal experiment
↓
Causal Ledger
↓
Candidate Knowledge
↓
Candidate Lifecycle
↓
Coherent Decision Input
↓
Current Context
↓
Applicability
↓
Candidate Resolution
↓
Modern Candidate Selection
↓
Recommendation
↓
Modern Decision Evidence
↓
Promotion Policy
↓
Explicit Decision
↓
Durable Authority
↓
Effective Candidate Disposition

This is a major change from the previous preview.

Previously, AO's modern architecture could answer:

"Is this manually requested experiment admissible?"

It can now additionally answer:

"Does any candidate currently deserve consideration?"

and:

"What explicit disposition has actually been authorized for this candidate?"

Those remain three different questions.

Selection, admission and promotion remain separate

Adaptive Optimization now deliberately has three distinct decision boundaries.

Selection

Which candidate deserves consideration?

Read-only.

Admission

Is this specific experiment safe and meaningful enough to attempt right now?

Controls permission to prepare a physical experiment.

Promotion / Decision Authority

What lifecycle disposition should be explicitly established for this candidate?

Controls durable modern candidate authority.

These systems are related.

They are not interchangeable.

A candidate can:

be selected but fail admission.

A candidate can:

be admissible for a controlled experiment but remain untrusted.

A candidate can:

remain EXPERIMENTAL even when current selection has no useful recommendation.

This separation will become increasingly important as AO gains more autonomy.

Why promotion authority came before autonomous experimentation

It would be possible to connect Modern Candidate Selection directly to the manual ARM path and let AO begin choosing its own experiments.

That is intentionally not happening yet.

Before selection can safely become autonomous, AO needs to understand the consequences of previous experiments.

Without durable decision authority, a selector could repeatedly rediscover the same candidate without a trustworthy mechanism to remember conclusions such as:

• continue experimenting;
• trust;
• reject;
• quarantine;
• supersede.

Promotion authority therefore provides another prerequisite for controlled autonomy.

Candidate conclusions are now durable

A future autonomous system should not need to relearn every candidate from zero after every restart.

The modern disposition journal now gives AO a durable place to preserve conclusions.

This is especially important for future states such as:

QUARANTINED

A candidate that caused a dangerous regression must not become eligible again simply because Minecraft restarted.

Likewise:

SUPERSEDED

must eventually preserve that a more appropriate descendant has replaced an older candidate.

The current runtime validation used the safer EXPERIMENTAL transition, but the authority architecture is designed for the broader lifecycle.

Quarantine remains an authority decision

A future safety detector may recommend quarantine.

That recommendation alone should not silently mutate Candidate Lifecycle.

The modern architecture now has a boundary where:

hazard evidence
↓
quarantine recommendation
↓
decision policy
↓
explicit authority
↓
QUARANTINED

can eventually occur.

This is necessary for AO's planned self-healing architecture.

Supersession now has a formal place

Strategy evolution will eventually create:

• descendants;
• parameter mutations;
• combinations;
• improved implementations.

The old candidate cannot simply disappear.

A future decision can explicitly establish:

SUPERSEDED

while preserving:

• the old candidate;
• its evidence;
• its lineage;
• the successor relationship;
• the reason for supersession.

This allows AO to evolve strategies without erasing causal history.

TRUSTED will remain revocable

The existence of a TRUSTED disposition does not mean future AO will treat trust as permanent.

The long-term design remains:

Trusted
↓
continued surveillance
↓
new contradiction or regression
↓
rollback when active
↓
quarantine / disposition change
↓
preserve evidence
↓
diagnose
↓
corrected descendant when possible
↓
retest

Trust means:

current evidence supports allowing this candidate under its valid conditions.

It does not mean:

this optimization can never be wrong again.

Diagnostics infrastructure

Development of this preview also introduced a dedicated Adaptive Optimization diagnostic exporter.

This is development/support infrastructure.

It is not part of the optimization decision authority.

The command:

/ao diagnostics export

can create numbered diagnostic bundles such as:

ao-diagnostic-0001.zip

ao-diagnostic-0002.zip

The bundles are designed to collect bounded AO evidence useful for development and troubleshooting.

This can include information such as:

• runtime state;
• causal ledger information;
• experiment journal state;
• Candidate Knowledge;
• Candidate Lifecycle;
• modern disposition authority;
• coherent decision state;
• admission information;
• selection information;
• modern decision read models;
• AO telemetry and local diagnostic information;
• bundle integrity metadata.

Diagnostic export is read-only

Creating a diagnostic bundle is not an optimization operation.

The exporter does not intentionally:

• create lifecycle authority;
• promote a candidate;
• change Candidate Knowledge;
• create an ARM;
• create an AdmissionLease;
• start an experiment;
• acquire the physical lane;
• apply an optimization;
• ingest new causal evidence;
• perform legacy learning.

Its purpose is to make the state AO already has easier to inspect.

Diagnostic bundles survive restart numbering

The exporter was also tested across a real Minecraft restart.

The first runtime produced:

ao-diagnostic-0001.zip

After closing and starting Minecraft again, the next manual export produced:

ao-diagnostic-0002.zip

The previous bundle was not overwritten.

No bundle was automatically produced merely because Minecraft restarted.

This makes the exporter useful for comparing evidence from separate runtime sessions.

Diagnostics do not become authority

Nothing inside the diagnostic ZIP becomes a new source of truth merely because it was exported.

The bundle is an observation artifact.

It can contain copies or bounded diagnostic views of authoritative information.

It cannot write that information back into AO.

There is also no automatic network upload.

The diagnostic system remains separate from optimization autonomy.

What is now complete?

Adaptive Optimization can now:

Observe
↓
Identify
↓
Diagnose
↓
Run controlled causal experiments
↓
Measure PRE / EXPERIMENT / RECOVERY
↓
Rollback
↓
Evaluate
↓
Persist trustworthy causal history
↓
Aggregate Candidate Knowledge
↓
Persist and rebuild Candidate Knowledge
↓
Evaluate Candidate Lifecycle
↓
Persist and rebuild Candidate Lifecycle
↓
Publish Coherent Decision Input
↓
Capture current runtime context
↓
Compare historical applicability
↓
Resolve exact executable candidates
↓
Verify physical execution identity
↓
Probe physical optimization state
↓
Evaluate manual admission
↓
Create a single-use AdmissionLease
↓
Revalidate immediately before execution
↓
Execute a controlled physical experiment
↓
Rollback
↓
Refresh causal knowledge
↓
Restart without replay
↓
Evaluate Modern Candidate Selection
↓
Return a conservative recommendation or no actionable choice
↓
Construct Modern Decision Evidence
↓
Evaluate disposition policy
↓
Create an explicit modern candidate decision
↓
Persist durable lifecycle authority
↓
Rebuild effective disposition after restart without replay

The modern architecture now contains both:

recommendation

and:

durable decision authority.

What is still missing?

Adaptive Optimization is still not fully autonomous.

The largest remaining boundary is no longer:

"Can AO choose which candidate deserves consideration?"

Modern Candidate Selection can now evaluate that.

It is also no longer:

"Can AO preserve an explicit modern conclusion about a candidate?"

Modern Promotion & Decision Authority can now do that.

The next major problem is:

"How does the modern architecture safely become the real operational authority?"

Legacy systems still retain operational responsibilities.

AO must not simply delete them.

The transition needs to prove that every responsibility being removed from the old architecture has a complete modern replacement.

Future authority migration

A future architecture will need to carefully move from:

legacy operational authority
+
modern causal intelligence

toward:

modern causal operational authority

without creating a period where:

• both systems believe they own the same action;
• neither system owns an action;
• an old fallback bypasses modern safety;
• a modern decision accidentally executes twice;
• legacy learning contradicts modern authority.

This transition must be explicit.

Autonomous selection is still disabled

Modern Candidate Selection exists.

That does not currently mean:

AO automatically selects a candidate and starts testing it.

The selection system remains read-only.

Future autonomous experimentation will require an explicit controlled connection between:

Selection
↓
Experiment request
↓
Admission
↓
AdmissionLease
↓
Fresh Revalidation
↓
Lane
↓
Execution

Every safety boundary introduced in the previous preview must remain intact.

Autonomous promotion is still disabled

The same applies to disposition authority.

AO does not currently run a background system that automatically changes candidates between:

• AVAILABLE;
• EXPERIMENTAL;
• TRUSTED;
• REJECTED;
• QUARANTINED;
• SUPERSEDED.

The authority architecture exists.

The first validated transition was explicitly controlled.

Future automation must earn the right to use that authority.

Optimization Value remains a future layer

Modern Candidate Selection is now real, but the full planned Optimization Value system is still broader.

Future AO will need to compare candidates using a multidimensional evaluation that can include:

• causal performance gain;
• confidence;
• repeatability;
• safety;
• semantic correctness;
• context applicability;
• coverage;
• generality;
• compatibility;
• long-term stability;
• runtime overhead;
• experiment cost;
• activation cost;
• restart cost;
• rollback reliability;
• blast radius;
• evidence breadth;
• contradiction history;
• recovery quality.

This is intentionally not reducible to:

highest FPS improvement wins.

Information value will become increasingly important

A future autonomous selector should sometimes prefer an experiment because it reduces an important uncertainty.

The best next experiment is not always the candidate predicted to produce the largest immediate gain.

It may instead be the experiment that answers:

"Which of these competing explanations is actually causal?"

or:

"Does this optimization still work in this environment?"

or:

"Is this candidate safe outside the narrow context where it previously succeeded?"

This will eventually allow AO to balance:

optimization value

with:

learning value.

Strategy evolution remains ahead

The current Candidate Catalog is still deliberately narrow.

Future AO should be able to evolve successful strategies.

Planned concepts continue to include:

• variants;
• descendants;
• parameter mutation;
• context specialization;
• champions;
• challengers;
• strategy combinations;
• dominance pruning;
• supersession.

A TRUSTED candidate should not become the final permanent endpoint of optimization search.

It can become a parent for better descendants.

Combinations remain independent candidates

Future Adaptive Optimization should also be able to discover when multiple primitives work better together.

Examples may include:

CACHE + INVALIDATE

DEDUPLICATE + BATCH

COALESCE + SCHEDULE

MEMOIZE + FAST_PATH

A combination must receive:

• a new identity;
• its own lineage;
• its own physical vector;
• its own rollback;
• its own experiments;
• its own Candidate Knowledge;
• its own lifecycle;
• its own disposition authority.

Trust cannot be inherited automatically from the components.

Strategy garbage collection remains future work

As open-world discovery begins creating larger candidate populations, AO will eventually need to manage old strategies.

Future strategy retention may distinguish concepts such as:

• KEEP;
• COMPRESS;
• ARCHIVE;
• DELETE;
• QUARANTINE.

However, cleanup must never erase evidence required to understand:

• important failures;
• safety incidents;
• lineage;
• supersession;
• causal contradictions;
• rollback history.

Optimization memory must remain useful without growing forever.

Self-healing remains a major objective

The new durable authority architecture provides another prerequisite for self-correction.

A future candidate may become trusted and later produce:

• instability;
• compatibility problems;
• semantic bugs;
• memory growth;
• delayed regressions;
• unexpected interactions.

The intended response remains:

detect
↓
rollback first
↓
quarantine
↓
preserve evidence
↓
diagnose
↓
generate a corrected descendant when possible
↓
retest
↓
only restore trust after new evidence

AO should correct its own conclusions rather than defend them.

Long-term surveillance remains necessary

Candidate decisions will eventually need evidence from much longer timescales than one experiment window.

Adaptive Optimization should detect failures that only appear:

• after hours of gameplay;
• after many chunks;
• after memory accumulation;
• after repeated world loads;
• after death / respawn;
• after dimension changes;
• after resource reloads;
• after mod configuration changes;
• after interaction with other optimizations.

A promotion decision is therefore never the end of observation.

Shared knowledge still cannot transfer authority

The long-term federated knowledge design remains unchanged.

Different AO installations may eventually share information such as:

• candidate identities;
• causal evidence;
• environment descriptors;
• failures;
• quarantines;
• delayed regressions;
• lineage;
• strategy relationships.

But imported knowledge must remain:

evidence

not:

local authority.

A candidate trusted in another modpack cannot silently become trusted locally.

Local context, validation and authority remain required.

Restart-required optimizations remain separate

The currently validated physical causal pilot remains:

LIVE_RUNTIME

Future candidates that require:

• bootstrap changes;
• transformation before class load;
• startup configuration;
• restart;

need a separate transactional lifecycle.

The future restart-required path remains conceptually:

preflight
↓
durable pending restart
↓
controlled startup
↓
boot verification
↓
runtime validation
↓
provisional result
↓
accept / rollback

Crash-loop protection remains mandatory.

Open-world discovery remains the final direction

Adaptive Optimization is still not intended to become a fixed list of handcrafted optimization switches.

Current candidates and descriptive categories are scaffolding for the architecture.

They are not intended to permanently limit discovery.

AO should eventually be capable of discovering:

• bottleneck families not explicitly known beforehand;
• new causal relationships;
• new candidate variants;
• unexpected interactions;
• optimization opportunities that were never manually listed by the original developer.

The principle remains:

Discovery should remain broad.

Authority and physical intervention should remain controlled.

Current status

The previous preview ended with:

Adaptive Optimization's modern causal intelligence can now influence whether a manually requested physical experiment is allowed to happen.

This preview extends that architecture significantly.

Adaptive Optimization can now:

remember
↓
construct coherent causal memory
↓
understand current context
↓
compare historical applicability
↓
resolve physical candidates
↓
safely admit controlled experiments
↓
learn from their results
↓
compare the modern candidate population
↓
refuse to choose when information is insufficient
↓
produce a non-authoritative recommendation
↓
construct evidence for a lifecycle decision
↓
evaluate an explicit disposition transition
↓
persist that decision as durable authority
↓
rebuild the exact effective disposition after restart
↓
without replaying the decision as physical behavior

The modern architecture now understands another critical distinction:

Knowing what may be worth doing

is not the same as:

being allowed to do it

and neither is the same as:

deciding what should be trusted afterward.

Adaptive Optimization now has separate architecture for all three.

The next major transition is allowing these modern systems to gradually become the real operational authority of AO while preserving every causal, safety, rollback and execution boundary built so far.