semiAIfoundry

System milestone

The commitment layer for AI-native chip design.

Generation is becoming abundant. Elimination is becoming more consequential. CRG Systems makes design-space reduction explicit, scoped, and independently replayable, so autonomous design systems can explore broadly without silently erasing choices future decisions may still need.

From broad design exploration to governed commitment A large connected field of design alternatives passes through the CRG commitment boundary. A retained family continues with a replayable record while evidence and change can reopen the decision.
Explore broadlyCommit within scopeReopen when reality changes
The defining systems question

What is the workflow justified in forgetting?

Orchestration decides what runs.

EDA determines what passes.

Commitment infrastructure makes narrowing explicit, scoped, and replayable.

The bottleneck has moved.

AI-native chip design is entering an age of abundance. Agents can generate more RTL variants, mappings, schedules, verification strategies, floorplans, package configurations, and repair paths than any engineering team can manually inspect.

Every autonomous design loop eventually narrows. Candidates are merged. Dominated points are pruned. Search stops. An architecture freezes. At that moment, the workflow makes a consequential claim, often silently: the distinctions it erased will not be needed by the decisions still ahead.

That claim needs infrastructure.

Verification asks whether a chosen design passes. Commitment asks whether choosing it destroyed information the next decision will need.

CRG Systems provides that missing layer. It gives AI-native workflows Realization Memory: a durable account of the alternatives, identities, evidence, decisions, and reopening conditions behind consequential reductions.

System position

The layer between possibility and commitment.

CRG sits beneath agents and alongside existing engineering tools. It does not replace generation, EDA, simulation, signoff, or organizational authority. It makes the narrowing between them inspectable.

01 / ExploreModels and agents
Generative modelsDomain generatorsSearch and optimizationMulti-agent systems
02 / GovernCRG commitment layer
Realization identityDecision scopeCertificate, witness, or refusalReopening conditions
03 / RealizeEDA and physical workflows
Evaluation and simulationVerification and signoffFoundry and packagingDeployment qualification
Evidence and change

Reprice, revalidate, reopen, or regenerate as the operating context changes.

Agents expand the search. CRG governs the commitments. Qualified engineering systems determine what becomes physically realizable.
Capstone milestone

From a formal law to an agent-native system.

CRG Systems v1.5.0 closes the initial theory-to-platform arc. Each stage is public, inspectable, and connected to the next.

  1. 01
    Formal theory

    Define what a safe reduction must preserve.

    A six-manuscript technical corpus formalizes computations, realization families, decision-visible geometry, and the Safe-Commitment Law.

    Decision-preservation laws
  2. 02
    Reproducible validation

    Test what breaks when narrowing happens too early.

    Frozen studies measure winner misses, regret under change, and the difference between repricing yesterday's survivors and reopening the family.

    Public protocols and results
  3. 03
    Operational system

    Carry exact semantics into deployable software.

    The cumulative 1.x runtime binds identity, retention, evidence, change, continued validity, portable records, and independent verification.

    Replayable runtime artifacts
  4. 04
    Agent-native boundary

    Put justified commitment in the agent's critical path.

    v1.5 adds causal receipts, bounded coordination, canonical package resolution, execution contracts, generated clients, and an Embedded Developer Kit preview.

    Integration-ready interfaces
Validation

The cost of forgetting is measurable.

The validation program asks the question optimizers usually leave behind: what must survive reduction so that future decisions remain sound?

26,624retention scenarios
4Transformer workloads
1,024capability-change scenarios
01 / Construct

Build the policy without seeing the test.

Five legal, nonvertical realizations of a Transformer microblock were evaluated across four workloads. Each policy could learn only from 512 construction scenarios per workload.

02 / Stress

Ask whether the retained set survives new conditions.

Disjoint seeds generated held-out, technology-shift, objective-shift, joint-shift, and controlled-perturbation cases. The headline comparison uses 1,024 held-out and 1,024 technology-shift scenarios per workload.

03 / Reopen

Change what is legally possible.

A separate 1,024-scenario track introduced a newly legal vertical realization, then compared rescoring yesterday's survivors with reopening and rebuilding the realization family.

Winner miss
The later best legal realization was absent from the retained set.
Relative regret
The normalized loss between the best retained choice and the true later best.
Matched budget
The exact minimax baseline and CRG ledger cover retained the same two-to-four candidates on all four workloads.
Later-winner miss rate

How often the retained set failed to include the later winner.

Held-out conditions

Initial winner only48.4%
CRG / minimax, matched budget0%

Technology shift

Initial winner only59.7%
CRG / minimax, matched budget11.0%
Capability change

Mean relative regret after the legal realization family changed.

100% change detection97.1% newly legal realization wins
Reprice retained options only27.5%
Reopen the realization family0% observed

No rescoring of yesterday's survivors can recover options that were never carried forward.

What the results say Keeping today's winner is brittle. A decision-scoped set preserved every later winner under held-out conditions at the same two-to-four-candidate budget, while a technology shift exposed the limit of that construction scope. Universal-monotone retention kept all five nondominated candidates and recorded no misses in its tested unchanged-family scope. When a new capability changed the legal family, reopening restored zero observed regret; repricing the old survivors could not.

What they do not say This is exact, finite-registry validation of retention and reopening semantics. It is not silicon PPA signoff, physical-implementation qualification, or a universal result across all chip designs. Construction and evaluation ledgers used disjoint, frozen seeds.

Read the frozen protocol ↗ Inspect the retained results ↗

Cumulative build

Six releases. One commitment stack.

v1.5 is consequential because it completes a connected architecture rather than adding an isolated agent feature.

  1. v1.0 / v1.0.1

    Exact foundation

    Identity, numerics, realization generation, geometry, certification, change, portable bundles, durable audit, and separate verification.

  2. v1.1.0

    Safe commitment

    Registry intake, decision-scoped retention, explicit pruning conditions, bounded-regret evidence, and portable records.

  3. v1.2.0

    Challenge and change

    Phase and stability analysis, certified deltas, comparison, and Decision Challenge.

  4. v1.3.0

    Qualified evidence

    Evidence applicability, uncertainty, qualification, selective revalidation, and controlled replacement.

  5. v1.4.0

    Durable validity

    Continued validity, lifecycle authority, reopening, recovery, migration evidence, and resilience.

  6. v1.5.0

    Agent-native boundary

    Design-space reduction, causal receipts, bounded coordination, package resolution, restricted execution, and end-to-end agent surfaces.

Exact identity governed retention challengeable change qualified evidence durable validity and authority agent-native realization memory

Platform depth

An integrated system, not a thin wrapper.

The v1.5 distribution exposes one canonical semantic system across product, integration, storage, lifecycle, and verification surfaces.

17component packages
100public schemas
106API operations
83executable CLI leaves
17Studio workspaces
7reference packs
Reference compatibility47 / 47

mandatory cases passed across 13 declared profiles

Independent implementation10 / 10

applicable cases passed by the zero-dependency Node.js verifier

The published status is compatible, not officially conformant: every mandatory case passes, while no external governance authority is represented as having issued a designation. Read the v1.5.0 release record ↗

Conventional build equivalent

The depth of the cumulative platform is measurable in years, not feature sprints.

An internal greenfield engineering-equivalence assessment estimates the conventional effort required to reproduce a comparable concept-to-v1.5 platform, including theory-to-system translation, exact computation, product surfaces, security, release discipline, and independent verification.

30–42FTE-years
30–42calendar months
$10–18Mestimated 2026 U.S. build effort

Engineering-equivalence planning estimate, not historical expenditure, audited cost, product pricing, company valuation, customer adoption, or realized return. The range reflects a technical-baseline, work-breakdown, and sensitivity method using U.S. labor and employer-compensation benchmarks.

v1.5 agent boundary

A decision travels with its proof.

A proposed reduction becomes a bounded, portable statement carrying the derived realization family, declared decision scope, retained and pruned realizations, resource limits, evidence, and an independent reconstruction path.

Authority boundary

Assessment is not authorization. A CRG result does not authorize execution, publication, tape-out, or repository mutation. Accountable organizational policy remains in control.

Declared request

An agent or orchestrator supplies the problem, not the conclusion.

  • Operator graph

    The computation and legal transformations.

  • Explicit bounds

    Exact inputs, resource limits, and completeness basis.

  • Decision scope

    The future observations the reduction must preserve.

  • Qualified evidence

    EDA, simulation, and source-adapter results with provenance.

CRG commitment boundary

May this distinction be discarded?

  1. 01

    Derive the finite closure, legal realization family, and exact geometry.

  2. 02

    Test the proposed reduction against the declared future-decision scope.

  3. 03

    Bind the result to evidence, resource limits, and reopening conditions.

Typed outcome

Uncertainty never collapses into an unsupported success.

  • Certificate

    The reduction is justified within its declared scope.

  • Regret witness

    A future decision can distinguish an option that would be pruned.

  • Typed refusal

    Evidence, completeness, support, or resource limits are insufficient.

Change routing

The return path depends on what changed.

Prices or weightsReprice
Evidence validityRevalidate
Legality or capabilityReopen
Generation boundaryRegenerate
CRG governs the permission to forget. It does not choose what to build or who may execute.
Portable commitment record

Independently replayable by design.

The verifier reconstructs supported claims rather than trusting producer status.

Registered family
Identity and completeness basis
Decision scope
Future observations to preserve
Reduction
Retained and pruned realizations
Result
Certificate, witness, or refusal
Validity
Evidence dependencies and reopening conditions
Verification
Code-separated reconstruction path

Build your differentiation above CRG. Do not rebuild the trust substrate beneath it.

Adoption path

Scale begins with one commitment boundary.

The initial 1.x foundation is complete and built for scale. v1.5.0 is ready for bounded integration into AI-native design workflows: begin where an agent or optimization process can erase a choice, declare what future decisions the reduction must preserve, and independently replay the result.

Expand from that tested boundary as the organization qualifies its tools, policies, evidence, and deployment environment.

01

Choose one consequential reduction.

02

Inspect the certificate, witness, or refusal.

03

Replay independently, then change an assumption.

CRG Systems is source-available for permitted research, educational, and other noncommercial use with attribution. Commercial use requires a separate written commercial license from Semi AI Foundry, LLC. The release license controls.

Explore freely.
Eliminate with reasons.
Commit within scope.
Reopen when reality changes.