Computation and realization / Guided exploration

Which options must survive?

A choice that is best today can become costly tomorrow. The deeper question is not which design wins now, but what must remain available for the decisions still ahead.

By F. M. Cinnor

The question

Optimization cannot recover an option that has disappeared.

Architecture, technology, and operating requirements do not arrive at the same time. A realization discarded under one set of assumptions may be the one a later decision needs.

CRG starts with the computation and its legal realizations. It asks which representation of those possibilities is sufficient for the future decision class, and when reducing that representation is a safe commitment.

Read the Safe-Commitment Law

An inspectable example

One computation. Three possible realizations.

A latency-lean choice, a balanced choice, and an energy-lean choice satisfy the same assumed contract. Lower resource burden is better. None dominates another.

The decision

Additive priceBottleneck

Resource price weighting80% latency / 20% energy

The retained family

A Latency-leanB BalancedC Energy-lean

Retained Removed Normalized units; not physical measurements.

Full-family optimumA / 3.2

Retained-family optimumA / 3.2

Commitment regret0

The retained family preserves this decision. That does not establish preservation of every future decision.

Every option remains visible for comparison, even when removed from the retained family.
Realization	Latency	Energy	Cost	Retention	Optimal in
A / Latency-lean	2	8	3.2	Retained	Full family + Retained family
B / Balanced	5	5	5	Removed	—
C / Energy-lean	8	2	6.8	Removed	—

Change the price.

With an 80% latency weighting, A wins at 3.2. Retain A, then shift to 20%: C would win at 3.2, but A now costs 6.8. The commitment regret is 3.6.

Change the decision class.

A and C together preserve every nonnegative additive-price optimum in this example. But minimizing the larger resource burden selects B: 5 instead of 8. Removing B incurs regret of 3.

A constructed teaching example, not a CRG benchmark, chip model, or issued certificate. The computation, legality, hierarchy, and resource accounting are held fixed; only the declared signature objective and retained family change.

The formal distinction

Safe for which future decisions?

For this finite example, let the full family contain A, B, and C, and let a nonempty retained family be a subset. The signature gives each realization's latency and energy burdens.

C_w(P)=wL(P)+(1-w)E(P),\quad 0\leq w\leq1

D(P)=\max\{L(P),E(P)\}

\operatorname{Regret}_{C}(\mathcal A)=\min_{P\in\mathcal A}C(P)-\min_{P\in\mathcal P}C(P)

Here, the full family is denoted by $\mathcal P$ , the retained family by $\mathcal A$ , and L and E are dimensionless burdens. The paper uses infima over the declared legal realization family; minima suffice for these three points.

The theory distinguishes an upper obligation region from its convex linear-cost body. Upper closure allows provisioned resource budgets above an attainable signature; convexification is invisible to additive prices, but not to every monotone objective.

\Gamma_{\mathcal A}=\operatorname{cl}\operatorname{up}\sigma(\mathcal A)

K_{\mathcal A}=\operatorname{cl}\operatorname{conv}\Gamma_{\mathcal A}

Nonnegative additive prices: Preserving the full linear-cost body preserves every optimum in this class. A and C are sufficient in the example.
Continuous monotone objectives: Preserving the full upper obligation region preserves this broader class. The bottleneck decision can detect the missing B.
Realization-sensitive decisions: A signature projection may omit layout, event structure, or other distinctions. Such decisions require information beyond the retained coordinates.

The claim is not that Pareto reasoning always fails, or that all options must always be kept. The required preservation depends on the decisions still in scope.

Safe-Commitment Law, PDF page 8 · Geometry and decision classes, PDF pages 24–27

Published validation

From a small example to a declared Transformer co-design case.

The released case study tests retention in a controlled six-realization Transformer family across 250 randomized price and objective changes. It compares a current winner, an initial two-point PPA Pareto set, and two CRG retention policies.

Released case study, v3.2.0, PDF page 13.
Retention policy	Missed optimum	Mean relative regret
Winner only	58.0%	7.018%
Initial PPA Pareto set	57.6%	6.965%
CRG declared-ledger set	0.4%	0.021%
CRG universal-monotone set	0%	0%

The two-realization declared-ledger set was constructed to intersect the optima of that ledger. The five-realization universal-monotone set retained the broader decision-sufficient region. These are different preservation commitments, not interchangeable certificates.

A controlled model-level experiment, not fabricated-silicon validation or a universal comparison with production design-space exploration. Physical calibration, wider workloads, and decision coverage remain separate questions.

Inspect the experiment and its boundaries

Into engineering practice

Keep the question connected to the machinery.

CRG System carries declared choices, evidence, and commitment into replayable engineering workflows. The interactive example above illustrates a principle; it does not run the released system.

Inspect the formal contribution

Six manuscripts establish the theory, foundations, scaling, design calculus, structural results, and Transformer case.

Read the manuscripts ↗

Explore the engineering context

Examine the built system

The operational commitment layer connects theory to bounded assessments, independently checkable evidence, and authorized decisions.

Explore CRG System ↗

Our complete Research in the Open commitment