Skip to content
finalstate.aiVerifiable Space: Where the Answer Can Be Checked
103Atlas

103The combination

Verifiable Space: Where the Answer Can Be Checked

Verifiable space is work where a valid check can evaluate a proposed answer within the available budget and time.

A lock dial aligns and its bolt retracts.An illustrative lock separates searching for a combination from testing one. The bolt retracts only after the dial reaches the correct alignment.

Checkable ground

Name the claim before choosing the check

Two candidates reach a check; one meets its rule and one fails. A third remains unscored.Neutral candidates receive a status only after evaluation. A check mark means this rule was met; a cross means this rule failed. The unscored candidate has neither mark. These are illustrative states, not a measured yield.CheckUnscoredMeets ruleFails rule
Schematic rule check. Scores and measurements need their own units and interpretation; unscored does not mean incorrect.

A verifier may establish correctness under formal rules, return a score, or measure a property. Its result supports only the claim it tests. For repeated automated sifting, the check must also be affordable and finish in time. The generation–verification gap asks a separate question: can the system select the good candidate it was capable of generating?

The evaluator in the loop

AlphaEvolve runs programs against task-specific criteria

Candidate programs enter an evaluator. A failed program stops; a retained program joins the database feeding the next generation.An illustrative correctness screen sits inside a scored evolutionary search. Failed checks can stop a program. Stored programs and evaluation results inform later prompts. No score improvement or success rate is plotted.RunRejectRetain
Novikov et al., AlphaEvolve (2025), §§2.1, 2.4–2.5. A schematic evaluation loop, not a performance chart.
  • Generate a candidate program
  • Execute the evaluator and record its results
  • Use stored programs and feedback in later generations

The AlphaEvolve white paper describes correctness checks, quantitative metrics and optional model-generated feedback. Those signals have different evidential roles. Evaluation may take seconds or substantial computation.

What the calculation establishes

GNoME screens structures, then calculations test stability predictions

One candidate crystal passes through a DFT calculation and receives a predicted-stability result.The same schematic unit cell appears before and after a density functional theory evaluation. Its attached check concerns predicted thermodynamic stability, not synthesis or usefulness. One example illustrates the relationship without implying a discovery rate.DFTPredicted stable
Merchant et al., Nature 624 (2023), 80–85, “Overview of generation and filtration.” Schematic unit cell; no measured energies or yield shown. Computational stability is not a synthesis certificate.

In Merchant et al., Nature (2023), graph networks filter candidates before density functional theory calculations evaluate selected structures. The result concerns predicted thermodynamic stability under the computational method. Synthesis and usefulness require further evidence. Error rates also matter when the desired target is rare: see the base-rate trap.

Find versus check

A proposed solution can make the checking task smaller

The same six-vertex graph appears twice, first with available edges and then with a proposed path visiting every vertex once.The proposed path is a certificate that can be checked against the graph. The picture supplies no search-time estimate and does not establish P unequal to NP.FindCheck
Illustrative Hamiltonian-path certificate. Cook, The P Versus NP Problem, §1 and p. 5; historical foundation: STOC 1971, DOI 10.1145/800157.805047.

Given a proposed route through a graph, check that every step uses an edge and every vertex appears once. Finding such a route is a different task. Cook's formal account defines efficient verification using certificates; P versus NP remains open. A small example does not prove a universal search-time gap.

What the bench decides

AlphaProteo's binding claims meet the laboratory

A designed binder is taken to a laboratory assay, which produces a measurement record.Separate design and assay objects show the need for experimental work. The assay produces a record without an invented affinity value or automatic successful docking. Binding under specified conditions does not by itself establish biological function or practical usefulness.DesignAssay record
AlphaProteo (2024), §2 and Supplement §S1.4; selected kinetic affinity measurements used bio-layer interferometry. Schematic assay; no response curve or affinity value shown.

The 2024 AlphaProteo report combines computational design with experimental binding and affinity assays. Groups at the Francis Crick Institute further validated selected binders. The announcement also reports failure on TNFα. An external measurement is not automatically a cheap, autonomous check.

The evaluator gives search a direction

A valid, affordable evaluator lets a system compare candidates on a stated criterion. Without that check, the proposals can still inform a person or a slower experiment. They remain unresolved on that criterion; missing cheap verification does not settle out-of-distribution judgment.

The edge of the map

Unscored, failed and falsely flagged are different states

A checked result and a false alarm sit inside the evaluator's scope. An unscored candidate remains outside it.The unscored candidate keeps a neutral outline. A separate flagged candidate is marked false only after follow-up. The three cases illustrate different evidence states; their count does not represent prevalence or test accuracy.CheckedUnscoredFalse alarm
Illustrative evidence states, not a frequency estimate. A valid test still needs deployment-specific prevalence and error rates.

A candidate without an applicable check remains unscored. A candidate that fails a valid check has evidence against that claim. A false alarm passes a screening threshold but proves to be a non-target on follow-up. Rarity can make false alarms dominate. Out-of-distribution cases and Knightian uncertainty require their own judgments.

Carry the verifier

A result travels with the check that supports it

A candidate is physically attached to an evidence record with fields for claim, check, cost and recheck conditions.The attached fields remain open for the owner to complete. A result cannot outrun the scope, cost or continuing validity of its evaluator. This is a record to complete, not an approval stamp.ClaimCheckCostRecheck

Record the claim, evaluator, cost and conditions that would require another check. An opportunity may have a deadline; a formal proof does not expire simply with time. Carry that evidence into sizing a bet, deciding when to abstain, and paying for verification. If the record is incomplete, keep the candidate unresolved.

Read the transcript

01 · THE COMBINATION

A steel safe, a numbered dial, and an unknown combination. Finding it may require many trials. Near misses give little help. But a candidate can be tested: the right sequence reaches the mechanism, the bolt retracts, and the door can open. The lock supplies a check on this attempt. It does not supply the next combination to try. Search and checking are different jobs.

02 · CHECKABLE GROUND

A proposed answer needs a check matched to its claim. Some checks establish correctness under formal rules. Others return a score or a measurement. Apply the check to the stated claim. A result may meet the rule or fail it; a candidate without an applicable check remains unscored. For an automated sift, the evaluator must also fit the available budget and time. A system that generates a good candidate may still fail to select it. That selection problem is the generation–verification gap.

03 · THE EVALUATOR IN THE LOOP

AlphaEvolve, announced by Google DeepMind in 2025, executes candidate programs against task-specific criteria. Programs that fail early tests can be stopped. Stored programs and their results inform the next generation. The evaluators can check correctness or measure performance. The white paper also permits model-generated feedback for properties such as simplicity; that feedback has a different evidential role from an execution test. Evaluation may take seconds or hours. Its affordability depends on the task and the value of continuing the search.

04 · WHAT THE CALCULATION ESTABLISHES

GNoME, reported in Nature in 2023, uses graph networks to screen candidate crystal structures. Selected structures then undergo density functional theory calculations. The resulting energies support a prediction of thermodynamic stability. The calculation remains an approximation with a specified method and comparison set. It does not establish that a structure can be synthesized or will be useful. Computational screening can reduce the experimental burden, but its compute cost still belongs in the search budget.

05 · FIND VERSUS CHECK

For some problems, a proposed solution comes with a certificate that makes checking straightforward. Consider a route through a graph. Check that each step uses an available edge and that every vertex appears exactly once. That checks this proposed route. It does not explain how to find one efficiently in every graph. Cook's formal account of P versus N P makes that distinction precise through polynomial-time checking. Whether efficient verification always implies efficient solution remains open.

06 · WHAT THE BENCH DECIDES

AlphaProteo provides a different kind of check. Researchers produce selected candidates and test them in the laboratory. The assays measure binding and affinity under specified conditions. Groups at the Francis Crick Institute further validated selected binders. The 2024 report includes an unsuccessful target, T N F alpha, as well as successful designs. Binding alone does not settle every question about biological function or practical usefulness. And laboratory work is not automatically cheap or autonomous. Its cost and turnaround determine where it fits in a search.

07 · Advertisement · ElevenLabs Studio

The last mile of content is often sound: the right voice, the clean take, the version that works in another market. ElevenLabs brings AI voice, music, dubbing, and sound tools into one production flow, so teams can shape audio with the same care they bring to the story. Create the version people can actually hear.

08 · DIRECTED SEARCH

A useful evaluator gives search a direction. A program can be compared on a measured criterion; a crystal can be assessed for computational stability; a binder can be tested under assay conditions. These results support different claims. The person designing the search must specify which result counts, whether the evaluator is reliable, and what repeated checks will cost. Where those conditions do not hold, proposals can still support human judgment or a slower experiment.

09 · THE EDGE OF THE MAP

Without an applicable check, a candidate remains unscored. That is different from failing a valid check. A test can also flag a non-target, which follow-up then identifies as a false alarm. When the target is rare, false alarms can dominate the queue even with a low false-positive rate. The boundary is therefore specific to the claim, evaluator and deployment. Out-of-distribution cases and Knightian uncertainty require further judgment; neither is defined simply by the absence of a cheap test.

10 · CARRY THE VERIFIER

Keep the evidence attached to the result. Name the claim, the evaluator, the cost of repeating it, and the conditions that would require another check. An opportunity may expire. A formal proof does not become invalid just because time passes; its formal assumptions still matter. If a result lacks the evidence needed for its claimed status, return it as unresolved. The owner can commission another test, narrow the claim, or decline to use it.

01 / 10 · THE COMBINATION0:00 / 5:38