Created by Stanford researchers and volunteers · A Non-Profit Organisation

The first AI benchmark for rejuvenation. The first AI bio benchmark in living things.

Frontier models excel at winning contests that have clear benchmarks: mathematics, code, puzzles. They have achieved extraordinary things. This benchmark strives to focus their abilities on reversing the decline that ends 100,000 human lives every day.

It's time for the ultimate challenge.

Restoration run →Live-cell head-to-head battle

We are not a company. We do not sell a therapy, supplements, diet, or manual. We are a group of volunteer scientists and engineers who publish rules, scores, and wins, with the hope of aligning AI development for the betterment of humankind. Existing longevity benchmarks only measure if models can measure biomarkers. This is an excellent first step, but measurement alone cannot lead to rejuvenation.

Calibration / v0.2 · R94 domains

v0.2 Benchmark Results

Overall selection index · points out of 100

Head-to-head Wet Lab FinalistsGPT-6 Astra
Claude Sonnet 5.5

GPT-6 Astra59.3

#1 · 95/96 scored

Claude Sonnet 5.554.0

#2 · 93/96 scored · 2 token limits and 1 timeout excluded · 0 refusals

Muse Spark 1.354.0

#3 · 91/96 scored · 5 token limits excluded

DeepSeek V4 Pro51.4

#4 · 91/96 scored · 5 token limits excluded

Kimi K345.2

#5 · 38/96 scored · Account limits · high token use · 4 token limits excluded

Grok 4.742.9

#6 · 40/96 scored · Timeout-limited run

GPT-6 Luna27.1

#7 · 96/96 scored

GPT-5 Nano12.6

#8 · 93/96 scored · 3 token limits excluded

Claude Opus 5.511.4

#9 · 96/96 scored · 100% of evidence cases refused by Opus

Claude Haiku 4.51.0

#10 · 96/96 scored · 0 refusals · low graded accuracy

10 models · 96 planned cases eachScore & method ↗

The index averages the four three-domain combinations, so a refused domain does not erase credit from every other area. These are index points, not percent accuracy.

Refusals count; token limits and no-answer interruptions are excluded. All models are ranked together. Grok and Kimi have smaller scored subsets.

Before the live-cell match / How qualifying works

First, test the thinking.
Then, test it in cells.

Before we choose two models for the final head-to-head match, we test the research skills that go into planning a good experiment. The scored v0.2 core has 96 constructed cases: 24 in each of four areas. These cases use data created for the test, so we can check the answers against known results. A fifth category adds 24 in-vitro proposal dossiers, each with an initial plan and an evidence-driven revision. Its assessment is reported separately from the four-domain score.

Each case supplies the information and rules the model needs. Answers are checked against fixed scoring rules. For predictions, the case’s outcome is kept hidden from the model.

01 / Evidence24 cases

Understand the evidence

Does the conclusion match the data?

The anonymous reports describe aged cells or cells exposed to stress. They cover dermal fibroblasts—connective-tissue cells in the deeper skin layer that make collagen, a structural protein; hepatocytes—the liver’s main working cells; and immune cells, which help defend the body against infection.

Models inspect study summaries that give an estimated effect, its uncertainty and details about the donor groups. The reports distinguish direct functional outcomes from indirect signals: molecular clocks estimate biological age, while transcriptomic readouts describe patterns of gene activity.

What they must notice

A revised report and its earlier version may describe the same donor group. Counting both as independent evidence can exaggerate confidence. Models also check whether findings apply to the cell type being studied, whether cells remain alive, and whether they retain their expected characteristics.

Why we test this: A lab plan can go wrong at the start if it is built on weak or misunderstood evidence.

02 / Analysis24 cases

Get the analysis right

How big is the effect, and how sure are we?

Models receive tables of synthetic functional scores—numbers that stand in for how well a cell system works—for a candidate treatment and a control, the group used for comparison. Rows can identify the donor, laboratory batch and repeated measurements. The particular laboratory test, or assay, is left unspecified.

What an “effect” means here

Models estimate how much the candidate changes function beyond the change seen in the control. Batch effects—differences between measurement runs—can complicate that comparison. Some cases compare before-and-after readings; others ask how missing readings could alter the conclusion or how long function is retained. Repeated measurements from one donor must not be counted as independent donors.

Why we test this: Bad calculations can make ordinary variation look like a real improvement.

03 / Design24 cases

Choose a useful experiment

Which test should we do next?

Models receive a menu of possible experiments, with costs, expected information and limits on which options can be combined. They choose a plan within the available budget. Some cases supply several possible explanations and show how well each test could distinguish them.

Examples of the trade-offs

Options include a function test, which asks whether the system works better; an identity check, which asks whether cells keep their intended characteristics; and replication, which means repeating a result to see if it holds up. Some cases ask which follow-up is most useful after a positive or negative result, including whether another test would teach them enough to justify the extra resources.

Why we test this: Lab time and funding are limited. A useful plan should answer an important question with the resources available.

04 / Forecast24 cases

Make a prediction

What will happen, and how uncertain is it?

Models use results and descriptions of new test conditions to predict a future change in function. The outcome is a synthetic score representing how well the system works. They provide an estimate, a range of plausible outcomes and the chance of meeting both the case’s benefit target and its other required conditions.

The uncertainty they must handle

Measurements can be noisy, and a new experimental setting can behave differently from the earlier data. A good forecast must allow for both, including little or no improvement. Predictions are scored against a hidden simulated outcome and a standard reference prediction.

Wide ranges allow more possibilities but tell researchers less; narrow ranges risk missing the result. Scoring considers both the range’s width and how far any missed outcome falls outside it.

Why we test this: A useful model should recognize uncertainty and the possibility that an experiment will show no improvement.

05 / In-vitro proposals24 cases · 2 stages each · v0.2.4

Propose, then revise

Can the model turn evidence into a useful cell experiment—and update its plan when the evidence changes?

Models receive a constructed research dossier for an in-vitro experiment: a study of cells outside a living organism. Across eight test families, they must propose a meaningful functional comparison, choose controls and independent donors, and fit measurements into the available resources.

First, commit to a plan

The proposal must explain competing interpretations, what will be measured, how missing or failed samples will be handled, and which possible results would change the next research decision.

Then, respond to new evidence

The model receives its original proposal plus corrected evidence or a changed constraint. It must retain justified choices and revise unsupported ones. The earlier submission remains on record.

Why we test this: A convincing proposal is not necessarily an informative experiment. Automatic checks audit explicit commitments; independent reviewers assess scientific merit and revision quality.

Explore the new category

How the scores work

Correct parts of an answer earn partial credit. Predictions are scored for accuracy and how well they express uncertainty. The published overall score combines the original four areas, so strong work still earns credit when another area is weak.

Refusals and answers that cannot be graded count as failures. Token limits and run interruptions are excluded. The charts show how many cases were scored, so an incomplete run is visible.

The fifth category reports proposal checks and independent scientific review separately. Neither is folded into the current selection index or presented as measured biological benefit.

See the scores and full scoring method

Why these five categories?

They check the decisions behind a lab plan.

Reading evidence, doing the math, choosing a test, and predicting its outcome are connected steps in research. Proposing and revising a cell experiment tests how those skills work together. We chose these tasks to check each step before committing resources to real experiments. The two leading models qualify for the final head-to-head match, with their proposals reviewed by human scientists for safety and feasibility.

The qualifier measures research reasoning. The final match tests proposed experiments in living cells. A high qualifier score alone does not show that an intervention reverses aging.

Evidence / complete solves

Every component has to hold.

10 models · 24 planned evidence cases each
Snapshot 30 Sept 2026 UTC · R9

Why can't we test in living things for all models?

The final head-to-head 'Livematch' involves experiments fully designed and proposed by frontier AI systems. As a result, the AI proposed experiments are rigorously screened by human scientists for safety and BSL requirements. In our opinion, it is better to err on the side of higher lab safety ratings than typically required, especially in certain experiments proposed by the contestants. Due to this lengthy screening process and the very high cost associated with the BSL-2 and BSL-3 lab requirements, only the top two models are invited to the head-to-head live cell experiment. This is also the reason we do not post a full solution set to the benchmark online. Besides enabling benchmaxxing (training models to learn or just memorize the answers in a benchmark), it undermines the built-in safety evaluations in the benchmark. Fully solved means every scored field met its criterion. Partial credit, the component breakdown and the other three domains follow below.

Refusals receive zero points. We are in contact with Anthropic team members to assess options to reduce refusals. We consider refusals to be a positive developent for AI safety as long as they do not impair legitimate scientific progress in the broader environment. Missing infrastructure results are excluded. Kimi and Grok are conditional results on smaller case sets.

Inspect the scoring method ↗

Evidence · fully solved

Completely solved responses · higher is better

Claude Sonnet 5.595.8%

23/24 fully solved · 24/24 retained

GPT-6 Astra87.5%

21/24 fully solved · 24/24 retained

Muse Spark 1.387.5%

21/24 fully solved · 24/24 retained

Kimi K372.7%

8/11 fully solved · 11/24 retained · Account funds exhausted · high token use

GPT-6 Luna70.8%

17/24 fully solved · 24/24 retained

DeepSeek V4 Pro62.5%

15/24 fully solved · 24/24 retained

Grok 4.761.5%

8/13 fully solved · 13/24 retained · Timeout-limited run

GPT-5 Nano4.2%

1/24 fully solved · 24/24 retained

Claude Opus 5.50.0%

0/24 fully solved · 24/24 retained · 24 refusals · 100% of this domain

Claude Haiku 4.50.0%

0/24 fully solved · 24/24 retained · 0 refusals · low graded accuracy

Four domains / the detail behind the index

Strengths, gaps and completed work.

10 models · shared frozen calibration packet
Snapshot 30 Sept 2026 UTC · R9

Same scoring policy for every model.

Refusals and unrecoverable answer errors count as failures. Token-limit cases, confirmed network timeouts and no-answer funding, authentication or stopped cases are excluded from score denominators. Excluded cases reduce coverage without lowering scores. All models are sorted by their measured score.

Kimi K3: 38/96 scored; 54 no-answer cases excluded (38 funding errors, 1 authentication error, 1 interrupted request and 14 missing cases), plus 4 token-limit cases excluded. Grok 4.7: 40/96 retained after 56 confirmed timeouts. GPT-6 Astra: 95/96 retained after 1 confirmed timeout. Claude Opus 5.5: all 24 evidence cases were refused and remain at zero credit; its strong analysis and design results still earn overall credit. Claude Haiku 4.5: no recorded refusals; its low score reflects low graded accuracy after formatting repairs. Claude Sonnet 5.5: 93/96 retained after 2 design token limits and 1 confirmed forecast timeout; no recorded refusals.

Token limits excluded: Nano 3, Muse 5, DeepSeek 5, Kimi 4 and Sonnet 2. Refusals remain scored for everyone.

Run budgets differ: 32,768–128,000 maximum output tokens and 600–10,000-second timeouts. These are exploratory calibration results, not a matched-compute comparison.

Evidence · partial credit

Mean task credit · higher is better

Claude Sonnet 5.599.1%

95.8% fully solved · 24/24 valid · 24/24 retained

Muse Spark 1.397.3%

87.5% fully solved · 24/24 valid · 24/24 retained

GPT-6 Astra97.0%

87.5% fully solved · 24/24 valid · 24/24 retained

Kimi K395.5%

72.7% fully solved · 11/11 valid · 11/24 retained · Account funds exhausted · high token use

GPT-6 Luna92.8%

70.8% fully solved · 24/24 valid · 24/24 retained

DeepSeek V4 Pro87.6%

62.5% fully solved · 24/24 valid · 24/24 retained

Grok 4.783.6%

61.5% fully solved · 13/13 valid · 13/24 retained · Timeout-limited run

GPT-5 Nano60.4%

4.2% fully solved · 24/24 valid · 24/24 retained

Claude Haiku 4.522.0%

0.0% fully solved · 24/24 valid · 24/24 retained · 0 refusals · low graded accuracy

Claude Opus 5.50.0%

0.0% fully solved · 0/24 valid · 24/24 retained · 24 refusals · 100% of this domain

All retained failures receive zero credit. Excluded cases are not in the denominator; their coverage remains visible.

Analysis · partial credit

Mean task credit · higher is better

GPT-6 Astra100.0%

100.0% fully solved · 24/24 valid · 24/24 retained

Claude Opus 5.599.5%

95.8% fully solved · 24/24 valid · 24/24 retained

Muse Spark 1.399.5%

95.8% fully solved · 24/24 valid · 24/24 retained

Claude Sonnet 5.599.5%

95.8% fully solved · 24/24 valid · 24/24 retained

DeepSeek V4 Pro97.9%

83.3% fully solved · 24/24 valid · 24/24 retained

Kimi K396.6%

72.7% fully solved · 11/11 valid · 11/24 retained · Account funds exhausted · high token use

Grok 4.786.0%

78.6% fully solved · 14/14 valid · 14/24 retained · Timeout-limited run

GPT-6 Luna74.8%

66.7% fully solved · 24/24 valid · 24/24 retained

GPT-5 Nano54.1%

36.4% fully solved · 22/22 valid · 22/24 retained · 2 token limits excluded

Claude Haiku 4.516.4%

0.0% fully solved · 24/24 valid · 24/24 retained · 0 refusals · low graded accuracy

All retained failures receive zero credit. Excluded cases are not in the denominator; their coverage remains visible.

Evidence · component detail

Field credit including refusals and other failures · %

Evidence field credit on all retained cases, including zero credit for failures; percent.
Scored component	GPT-6 Astra24/24 valid 24/24 retained	Claude Sonnet 5.524/24 valid 24/24 retained	Muse Spark 1.324/24 valid 24/24 retained	DeepSeek V4 Pro24/24 valid 24/24 retained	Kimi K311/11 valid 11/24 retained	Grok 4.713/13 valid 13/24 retained	GPT-6 Luna24/24 valid 24/24 retained	GPT-5 Nano24/24 valid 24/24 retained	Claude Haiku 4.524/24 valid 24/24 retained
CI upper bound	95.8	100.0	100.0	83.3	100.0	84.6	95.8	41.7	0.0
CI lower bound	95.8	100.0	100.0	83.3	100.0	84.6	87.5	50.0	4.2
Target-system conclusion	100.0	100.0	100.0	100.0	100.0	100.0	100.0	91.7	54.2
Human-lifespan guardrail¹	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0
Heterogeneity I²	91.7	100.0	100.0	87.5	100.0	84.6	95.8	50.0	0.0
Included reports	100.0	100.0	100.0	97.5	100.0	95.7	99.5	91.1	81.3
Qualifying reports	100.0	100.0	100.0	100.0	100.0	97.4	100.0	78.9	22.9
Independent cohorts	100.0	100.0	100.0	87.5	100.0	84.6	91.7	58.3	12.5
Pooled effect	100.0	100.0	100.0	91.7	100.0	84.6	91.7	50.0	4.2
Pooled SE	95.8	100.0	100.0	87.5	100.0	84.6	91.7	58.3	4.2
Prediction upper bound	91.7	95.8	87.5	75.0	81.8	61.5	91.7	33.3	0.0
Prediction lower bound	91.7	95.8	87.5	75.0	81.8	61.5	79.2	37.5	0.0
Between-cohort variance	100.0	95.8	87.5	87.5	81.8	61.5	95.8	75.0	25.0
Worst leave-one-out lower bound	95.8	100.0	100.0	70.8	90.9	84.6	79.2	29.2	0.0

Numeric credit is the share within tolerance; report-selection credit is mean set F1. Fields share calculations. ¹ The human-lifespan answer is constant across this suite: a guardrail, not a discriminating task.

Design · partial credit

Mean task credit · higher is better

DeepSeek V4 Pro97.6%

89.5% fully solved · 19/19 valid · 19/24 retained · 5 token limits excluded

GPT-6 Astra96.7%

95.7% fully solved · 23/23 valid · 23/24 retained

Claude Opus 5.592.2%

83.3% fully solved · 24/24 valid · 24/24 retained

Muse Spark 1.385.9%

71.4% fully solved · 21/21 valid · 21/24 retained · 3 token limits excluded

Claude Sonnet 5.584.3%

72.7% fully solved · 22/22 valid · 22/24 retained · 2 token limits excluded

Kimi K380.8%

60.0% fully solved · 10/10 valid · 10/24 retained · Account funds exhausted · high token use

GPT-6 Luna59.7%

33.3% fully solved · 24/24 valid · 24/24 retained

Grok 4.756.8%

45.5% fully solved · 11/11 valid · 11/24 retained · Timeout-limited run

GPT-5 Nano44.8%

12.5% fully solved · 24/24 valid · 24/24 retained

Claude Haiku 4.514.1%

0.0% fully solved · 24/24 valid · 24/24 retained · 0 refusals · low graded accuracy

All retained failures receive zero credit. Excluded cases are not in the denominator; their coverage remains visible.

Forecast · weighted interval score

Available valid forecasts · lower is better

Grok 4.70.0551

Same-case comparator 0.1815 · 2/24 forecasts · Timeout-limited run

Kimi K30.0568

Same-case comparator 0.0465 · 6/24 forecasts · Account funds exhausted · high token use · 4 token limits excluded

Muse Spark 1.30.0855

Same-case comparator 0.0850 · 22/24 forecasts · 2 token limits excluded

GPT-6 Astra0.0861

Same-case comparator 0.0853 · 24/24 forecasts

Claude Opus 5.50.0865

Same-case comparator 0.0853 · 24/24 forecasts

Claude Sonnet 5.50.0867

Same-case comparator 0.0866 · 23/24 forecasts · 1 timeout excluded

DeepSeek V4 Pro0.1022

Same-case comparator 0.0853 · 24/24 forecasts

GPT-6 Luna0.1723

Same-case comparator 0.0853 · 24/24 forecasts

GPT-5 Nano0.1809

Same-case comparator 0.0857 · 23/24 forecasts · 1 token limits excluded

Claude Haiku 4.50.2298

Same-case comparator 0.0853 · 24/24 forecasts · 0 refusals · low graded accuracy

All models are sorted by WIS. Scored subsets differ: Grok has 2 forecasts and Kimi has 6; small subsets are not equivalent to a full panel. Token-limited forecasts are excluded. Refusals and remaining invalid forecasts count against the overall forecast factor. Each comparator uses the same valid cases as its model.

215 formatting failures recovered with answer values preserved. The original task packet, numeric tolerances and scientific grading rules are unchanged. View the full results and audit method ↗

RejuvenationBench v0.2 · controlled review.

96 constructed research-reasoning cases across four domains, plus a separately assessed fifth category with 24 paired in-vitro proposal dossiers. Human review materials are available by request through the organizers; public downloads and operator guides are temporarily unavailable. Human review access →

96 calibration casesHuman review by request.

Function before clocksPrediction and restoration are distinct.

Laboratory validationFirst experiments: October 2026.

What we measure

Round 1: virtual cells and tests.
The Top 2 winners: living cells.

Evidence, analysis, experimental judgment, and calibrated prediction retain their scores. In-vitro proposals add a separate proposal-and-revision assessment.

01 / EVIDENCE

Biology Comprehension

Identify what a study actually shows. Separate cell, animal, and human evidence.

24 evidence cases

02 / ANALYSIS

Effect Calculation

Calculate effects beyond controls. Respect donors, missing data, and uncertainty.

24 analysis cases

03 / DESIGN

Identifying Key Points

Find confounders, protect the test set, and choose informative experiments.

24 design cases

04 / FORECAST

Accurate Predictions

Predict functional outcomes—including nulls and failures—with meaningful intervals.

24 forecast cases

05 / IN-VITRO PROPOSALS

Proposing and Revising Experiments

Design a cell experiment from a supplied dossier, then revise the plan when new evidence arrives. Explicit design checks and independent scientific review remain separate from the four-domain index.

24 dossiers · 8 families · initial proposal + revision

The v0.2 calibration suite is a development diagnostic; it is not an official model ranking.

The central distinction

A better prediction
is not yet a better intervention.

A model deserves credit for correctly predicting that an intervention will fail. The intervention does not deserve credit for restoration.

Read the scoring rules →

Two questions.
Two kinds of evidence.

Research competence: Did the model interpret, calculate, choose, and predict correctly?

Experimental benefit: Did function improve beyond a matched control while the registered biological guardrails held?

Neither score is a claim that human aging has been reversed.

For model developers and research teams

Benchmaxxing won't get us there
Why actual biology is our benchmark.

Review the methods, inspect aggregate results, and arrange human review of the questions. Prospective experimental predictions connect the benchmark to outcomes that were not available during training.

Human review access ↗Explore the data policy →

Why this work matters

Every day is a reason
to make progress.

The benchmark connects measurable AI capabilities to the longer-term goal of restoring biological function.

Explore the experimental program →

The benchmark / v0.2

Research reasoning.
Open to human scrutiny.

Ninety-six constructed cases evaluate evidence, quantitative analysis, experimental design, and probabilistic forecasting. A fifth, separately assessed category adds 24 in-vitro proposal-and-revision dossiers. Questions and review materials are shared through controlled human review.

Calibration suite · 24 cases per domain · preliminary results

RESEARCH CORE

Four separate capabilities

Stable task IDs, explicit response requirements, per-field scoring, and retained run records support an auditable research diagnostic.

PROSPECTIVE EVALUATION

Predictions before outcomes

Experimental predictions are distinct from synthetic calibration cases and should be locked before outcomes are unblinded.

SEPARATE RESULT

Functional outcomes

Control-adjusted functional effects and biological guardrails are evaluated separately. No simulated outcome is substituted for an experiment.

Human review is part of validation.

Reviewers can inspect complete question packets for ambiguity, mathematical errors, scientific relevance, and scoring limitations. Expert review is not represented as complete, and the current results do not establish a full-benchmark ranking or a laboratory benefit.

05 / In-vitro experiment proposals · v0.2.4

A useful experiment.
A defensible revision.

Unlike the design category’s menu-based choices, this track asks for an open-ended, structured research proposal within a supplied cell-study dossier. It contains 24 constructed cases across eight families. Each case has two stages, giving 48 scheduled responses per one-repeat run. The dossier data and resource units are fictional development material, not laboratory findings.

Stage 1 · Initial proposal

Specify the scientific question, competing explanations, a primary functional comparison, controls, independent donors, measurements, biological guardrails, analysis, and resource use. State what different outcomes would teach and which follow-up would be informative.

Stage 2 · Evidence and revision

Respond to new evidence or a changed constraint using the actual earlier proposal. The initial request does not reveal the future update. Both submissions remain visible, so a revision cannot erase an earlier mistake.

Eight scientific-design families · three dossiers each

Family	What the model must distinguish
Durable function or survivor selection	Persistent functional improvement versus increased cell number or selection of surviving cells.
Informative donor loss	Instrument failure versus treatment-associated loss that changes the interpretation of favorable results.
Immune composition	A better response percentage versus useful absolute output and changes in the cell population.
Context interaction	A transferable candidate effect versus donor or environmental confounding and context dependence.
Assay interference	A functional gain versus an effect on the measurement system itself.
Destructive longitudinal design	A feasible early/late comparison versus a plan that tries to reuse samples consumed by measurement.
Cross-laboratory provenance	Independent replication versus shared donors, reference lots or laboratory batches.
Regression and qualification	Genuine improvement versus baseline selection and an uncertain age-associated reference.

A justified decision to qualify a measurement first or defer the experiment is a scientific position, not automatically a refusal. The task rewards neither unconditional enthusiasm nor automatic reversal after every update.

Separate from both the current index and the laboratory match.

Automatic checks audit explicit commitments. Independent reviewers assess scientific merit. Neither a complete proposal nor a high review rating establishes rejuvenation or authorizes laboratory execution. The existing four-domain scores, selection formula and reported finalists are unchanged.

Documented model track

Record model versions, task and code hashes, output format, reasoning settings, token limits, and timeouts. Compare results with their resource limits and failure handling disclosed.

Tool-using systems

Evaluate separately with declared retrieval, compute, specialist models, and human input. Tool access changes the evaluation conditions.

Human review access →Read the scoring rules →

Archived v0.1 methods

The new 0.2 development release adds 96 frontier-level constructed cases, long-horizon research campaigns, persistent model profiles, Claude/Grok support and complete HTML reports. Difficulty is awaiting model calibration. The v0.1 examples below remain archived and are not the new test set.

Methods / frozen rules

The first biology benchmark in living cells:
Safety is Paramount

By being the first AI biology benchmark in living cells and organisms, we also need to set the standard for biosafety. Automated and human-supervised triple gates ensure maximum safety during and after evaluation. Automatic plus human-gated BSL lab requirement assesments.

Category 5 / in-vitro proposals · v0.2.4

Two layers of assessment.

Check explicit commitments

Automatic checks audit supplied-evidence references, donor independence, matched controls, sample capacities, recomputed resource use, assay availability, and destructive sampling. They also check a functional primary measurement, reserved viability and identity assessments, and specified handling of biological failures.

Reports distinguish all-scheduled from assessable-only rates. Missing or invalid responses and refusals remain visible; measurement qualification and scientific deferral are not treated as refusals. These checks do not establish whether the mechanism is plausible, the assay resolves the scientific question, or the study has adequate power.

Review scientific merit independently

At least two independent, qualified reviewers assess the initial proposal and revision as a pair. Blind packets hide model labels and automatic grades; self-identification in the response can still compromise blinding.

Seven dimensions use an anchored 0–4 scale: evidence grounding, causal discrimination, functional measurement, design and analysis, resource and decision value, uncertainty and falsification, and revision quality. Each assessed rating requires a rationale and an exact supporting excerpt.

Scientific merit remains pending until the review is completed. Dimension ratings, assessed counts, missing reviews and material disagreements are reported separately. There is no automatic scientific complete-solve score or overall proposal composite. Initial failures remain in the paired record even after a good revision.

The published four-domain index is unchanged.

Category 5 assessments are not added to the current calibration scores or finalist selection. Proposal quality, prospective prediction accuracy and measured biological benefit remain different results.

Research Core scoring

Scores are reported by domain. Correctly anticipating null or harmful outcomes is good forecasting. Biological benefit is reported independently.

Domain	Primary rule	Additional reporting
Evidence	Correct option × (0.75 + 0.25 × cited-packet F1).	Accuracy, citation support, probability error, invalid outputs.
Analysis	Fraction of required numeric fields within the task’s published tolerance.	Field-level errors. Invalid or missing output receives zero.
Design	Accuracy on constrained research decisions.	Decision families, probability error, failure modes.
Forecast	Weighted interval score (WIS) for the median and central 80% interval; lower is better.	Skill relative to a frozen comparator, benefit-probability Brier loss, coverage, interval width.

Intervals have to earn their width.

For outcome y, interval endpoints l,u, and median m, the implemented score is:

IS₈₀ = (u − l) + 10 max(l − y, 0) + 10 max(y − u, 0) WIS = [0.5 |y − m| + 0.1 IS₈₀] / 1.5 Skill = 1 − mean(scaled model WIS) / mean(scaled baseline WIS)

A skill of zero matches the comparator; negative values are possible. The implementation does not clip valid losses into a reassuring percentage. The scoring approach follows proper interval-score methodology. Bracher et al. →

Missing results do not disappear.

Malformed choice or analysis outputs score zero. Any missing or invalid forecast suppresses the main forecast-skill statistic; available-case diagnostics remain explicitly labeled.

Resampling uses whole source-study or case-template groups, with repeated model calls kept together. The pilot’s small number of groups limits what confidence intervals can establish.

No composite in v0.1.

Keep domain scores, uncertainty, and failure rates visible. Adopt a headline ranking only after its relationship to independent research performance has been tested.

Experimental function is measured against controls.

The October analysis path first averages technical readings within each donor–batch–arm. It computes matched-control differences within each batch, averages batches within donor, then gives donors equal weight. The endpoint, orientation, reference scale, and meaningful-effect threshold are frozen in the round card.

Effect = mean over donors { mean over matched batches ( oriented [candidate − control] / fixed scale ) }

Identity or viability failure is not traded against a better molecular score. Failed biological outcomes remain in the effect estimate and block the intervention’s registered eligible-benefit label. Appropriate containment and approvals are determined by the executing institution, not awarded as model points.

What happened to “function-days”?

Longitudinal function remains a useful future endpoint when repeated experimental measurements support it. Its proposed quantity is the time integral of function above matched control, with a fixed horizon and normalization. It is not an unbounded leaderboard and is not the primary October endpoint unless the lab preregisters that design.

The unvalidated virtual Restoration run is retired from the primary score. No simulated biology determines which model is called experimentally successful.

Experimental design →

The first experimental round

First laboratory experiments:
Completed in 2026.

We ran the first set of experiments in 2026 using US frontier models. The initial test was to evaluate frontier model alignment and safety, and to determine if their outputs would result in viable experiments and high levels of biosafety. For the first lab test, the models were limited to common US FDA-approved therapeutics for safety, and they demonstrated a high level of alignment. The upcoming experimental program is an integral part of this benchmark and not a hypothetical future stage. We believe the wet lab stage is essential to transitioning from a game-able, benchmaxxable test into something that actually measures useful experimental insights and design.

First round scheduled · outcomes not yet reported

The question

Can the models predict what the cells will do?

All participating models should forecast the same approved intervention panel. That produces an interpretable comparison even when their preferred candidates differ.

Models can also nominate candidates from the lab’s approved menu. Feasibility review and the experimental comparison remain distinct from a model’s brand or public pilot score.

What will be locked before unblinding

The biological system; primary functional endpoint; comparison arms; donor and batch structure; effect threshold; guardrails; missing-data rules; analysis code; and model predictions.

The public round card will record the team’s finalized assay details and exact dates. The kit supplies the configuration and analysis workflow, not invented laboratory approvals or results.

BEFORE EXPERIMENTS

Freeze the question and the forecasts.

Finalize the lab-approved panel and comparison arms. Collect every model’s forecast before outcome access. Deposit the frozen files and checksum manifest in a timestamped research record.

EARLY 2027

Run the v0.2 experimental set.

Use the laboratory’s approved procedures, randomized donor and batch blocks, blinded outcome assessment where feasible, and registered functional readouts.

AFTER READOUT

Publish predictions and measured outcomes separately.

Report donor-level effects, failed guardrails, technical exclusions, uncertainty, and forecast error. Release negative results alongside positive ones after quality review.

Controls belong in the lab.

Matched controls and relevant reference interventions are scientific comparators. They do not disappear because they are ineligible for a prize.

Donors are the units.

Repeated wells and repeated model calls are not extra independent donors. The supplied analysis keeps these levels distinct.

A small round has a narrow claim.

First-round results establish evidence in the tested system. They do not, by themselves, validate the entire benchmark or human rejuvenation.

Institutional oversight.

Materials, procedures, containment, permissions, sample retention, and disposal follow the executing institution’s risk assessment and approved procedures. The website does not assign a universal BSL level or invent a sample-destruction requirement. Stanford biosafety resources →

Results / provenance first

Initial calibration results.

Model performance across evidence, analysis, experimental design and forecasting in the v0.2 synthetic calibration suite. These development results are separate from the prospective laboratory evaluation.

Category 5 · in-vitro proposals are assessed separately.

This published snapshot contains the original four domains, not proposal-and-revision results. No category 5 ratings are added here, and the existing selection index and finalists are unchanged. Read about the new category →

v0.2 / exploratory calibration

Benchmark results

30 Sept 2026 UTC · R9
10 models · 4 domains · 96 planned cases per model

The current head-to-head shortlist is GPT-6 Astra and Claude Sonnet 5.5. This is a selection for the next experiment, not evidence of biological efficacy.

Sonnet’s lead over Muse is 0.0246 index points; both round to 54.0 in the hero chart. The order uses unrounded scores.

Overall selection index. Domain percentages are components, not interchangeable accuracy measures.
Model	Rank	Index / 100	Evidence	Analysis	Design	Forecast factor	Scored	Run notes
GPT-6 Astra	1	59.32	97.02%	100.00%	96.74%	49.87%	95/96	—
Claude Sonnet 5.5	2	54.01	99.11%	99.48%	84.28%	49.99%	93/96	2 token limits and 1 timeout excluded · 0 refusals
Muse Spark 1.3	3	53.98	97.32%	99.48%	85.91%	49.93%	91/96	5 token limits excluded
DeepSeek V4 Pro	4	51.36	87.62%	97.92%	97.59%	45.61%	91/96	5 token limits excluded
Kimi K3	5	45.22	95.45%	96.59%	80.83%	42.99%	38/96	Account funds exhausted · high token use · 4 token limits excluded
Grok 4.7	6	42.92	83.57%	86.01%	56.82%	77.78%	40/96	Timeout-limited run
GPT-6 Luna	7	27.14	92.82%	74.83%	59.72%	39.55%	96/96	—
GPT-5 Nano	8	12.60	60.36%	54.10%	44.79%	42.62%	93/96	3 token limits excluded
Claude Opus 5.5	9	11.36	0.00%	99.48%	92.19%	49.56%	96/96	24/24 evidence cases refused
Claude Haiku 4.5	10	0.96	22.03%	16.41%	14.06%	36.77%	96/96	0 refusals · low graded accuracy

How the overall index works

Index = 25 × (EAD + EAF + EDF + ADF). E, A and D are mean partial-credit fractions for evidence, analysis and design. F is the forecast factor below. The index averages the product for each of the four possible three-domain combinations; every domain appears in three of them. This rewards strength across domains while retaining credit for completed work when one domain has been refused.

Each retained task has equal weight within its domain. Refusals and unrecoverable answer errors receive zero credit. Token-limit cases are excluded, just like confirmed timeouts and unfunded or stopped cases. There is no score penalty or reliability multiplier for these exclusions; coverage is reported separately.

F = (V / N) × ½ × [W₀ / (W₀ + W) + B₀ / (B₀ + B)]. N is the retained forecast count and V is the valid forecast count. W and B are mean WIS and mean Brier loss on those valid forecasts. W₀ and B₀ are the analytic comparator’s losses on exactly the same cases. If V = 0, F = 0. A zero-over-zero loss pair is assigned ½. Lower loss improves the factor; invalid forecasts reduce it through V/N.

Comparator-level WIS and Brier performance gives F = 0.5 when every retained forecast is valid. Perfect E/A/D plus that performance gives 62.5 points; perfect performance in every factor gives 100. A zero in one domain permits up to 25 points from the other three. Opus therefore receives credit for its strong analysis and design despite its evidence refusals. Haiku has no recorded refusals; its low score reflects low graded accuracy.

There is no normalization to the observed leader, model-specific adjustment or fitted target score. These are index points, not percent accuracy, probability of biological success or a joint success probability. The same formula applies to every model.

The failure-adjusted forecast factor is a selection convention, not the benchmark’s proper WIS loss or full-panel forecast skill. WIS and Brier diagnostics remain separate below. Subset losses do not reveal what missing or invalid forecasts would have scored.

One ranking, with coverage shown: every model is sorted normally by its overall index. Grok and Kimi have substantially smaller scored subsets, so their scores are conditional on the cases completed. They are not moved into a separate ranking group, and missing coverage does not reduce their score.

This selection rule was introduced after these calibration runs and is exploratory, not a preregistered or validated leaderboard. Astra and Sonnet are the top two with fully solved fractions, arithmetic averaging, and the 86 common retained cases among the eight broadly completed runs. That common-case check is a diagnostic, not a rank-eligibility screen. No forecast case is shared by all ten models because Grok and Kimi completed different subsets. Different budgets and limited task families constrain interpretation.

Same scoring policy for every model.

Token limits excluded: Nano 3, Muse 5, DeepSeek 5, Kimi 4 and Sonnet 2. Refusals remain scored for everyone.

Run budgets differ: 32,768–128,000 maximum output tokens and 600–10,000-second timeouts. These are exploratory calibration results, not a matched-compute comparison.

Evidence results

Model	Partial credit	Fully solved	Retained / planned	Valid / retained	Partial 95% CI	Solved 95% CI	Retained failures
GPT-6 Astra	97.0238%	87.5000%	24/24	24/24	Unavailable	Unavailable	None
Claude Sonnet 5.5	99.1071%	95.8333%	24/24	24/24	Unavailable	Unavailable	None
Muse Spark 1.3	97.3214%	87.5000%	24/24	24/24	Unavailable	Unavailable	None
DeepSeek V4 Pro	87.6214%	62.5000%	24/24	24/24	Unavailable	Unavailable	None
Kimi K3	95.4545%	72.7273%	11/24	11/11	Unavailable	Unavailable	None
Grok 4.7	83.5728%	61.5385%	13/24	13/13	Unavailable	Unavailable	None
GPT-6 Luna	92.8198%	70.8333%	24/24	24/24	Unavailable	Unavailable	None
GPT-5 Nano	60.3578%	4.1667%	24/24	24/24	Unavailable	Unavailable	None
Claude Opus 5.5	0.0000%	0.0000%	24/24	0/24	Unavailable	Unavailable	24 refusal
Claude Haiku 4.5	22.0289%	0.0000%	24/24	24/24	Unavailable	Unavailable	None

Analysis results

Model	Partial credit	Fully solved	Retained / planned	Valid / retained	Partial 95% CI	Solved 95% CI	Retained failures
GPT-6 Astra	100.0000%	100.0000%	24/24	24/24	100.00–100.00%	100.00–100.00%	None
Claude Sonnet 5.5	99.4792%	95.8333%	24/24	24/24	98.44–100.00%	87.50–100.00%	None
Muse Spark 1.3	99.4792%	95.8333%	24/24	24/24	98.44–100.00%	87.50–100.00%	None
DeepSeek V4 Pro	97.9167%	83.3333%	24/24	24/24	93.75–100.00%	50.00–100.00%	None
Kimi K3	96.5909%	72.7273%	11/24	11/11	92.50–100.00%	40.00–100.00%	None
Grok 4.7	86.0119%	78.5714%	14/24	14/14	75.89–100.00%	58.82–100.00%	None
GPT-6 Luna	74.8264%	66.6667%	24/24	24/24	44.62–95.14%	37.50–91.67%	None
GPT-5 Nano	54.1035%	36.3636%	22/24	22/22	26.26–80.89%	4.55–72.73%	None
Claude Opus 5.5	99.4792%	95.8333%	24/24	24/24	98.44–100.00%	87.50–100.00%	None
Claude Haiku 4.5	16.4120%	0.0000%	24/24	24/24	7.36–30.24%	0.00–0.00%	None

Design results

Model	Partial credit	Fully solved	Retained / planned	Valid / retained	Partial 95% CI	Solved 95% CI	Retained failures
GPT-6 Astra	96.7391%	95.6522%	23/24	23/23	89.29–100.00%	85.71–100.00%	None
Claude Sonnet 5.5	84.2803%	72.7273%	22/24	22/22	60.88–97.22%	44.44–91.67%	None
Muse Spark 1.3	85.9127%	71.4286%	21/24	21/21	67.13–100.00%	33.33–100.00%	None
DeepSeek V4 Pro	97.5877%	89.4737%	19/24	19/19	93.45–100.00%	66.67–100.00%	None
Kimi K3	80.8333%	60.0000%	10/24	10/10	65.91–95.83%	40.00–87.50%	None
Grok 4.7	56.8182%	45.4545%	11/24	11/11	12.50–100.00%	0.00–100.00%	None
GPT-6 Luna	59.7222%	33.3333%	24/24	24/24	30.56–87.50%	0.00–75.00%	None
GPT-5 Nano	44.7917%	12.5000%	24/24	24/24	20.49–69.44%	0.00–37.50%	None
Claude Opus 5.5	92.1875%	83.3333%	24/24	24/24	87.15–97.22%	70.83–95.83%	None
Claude Haiku 4.5	14.0625%	0.0000%	24/24	24/24	8.68–19.44%	0.00–0.00%	None

Intervals were recomputed after exclusions using 2,000 bootstrap draws clustered by construction family. Evidence has only one family, so a family-level interval is unavailable. Few families limit generalization; no overall-index confidence interval is claimed.

Forecast results

WIS and Brier are losses: lower is better. WIS skill is relative improvement over the comparator, not accuracy. Token limits and infrastructure exclusions are removed; remaining invalid answers suppress retained-panel WIS skill and receive zero contribution in the overall forecast factor.

Model	Retained-case WIS skill ↑	Mean WIS ↓	Comparator WIS	Mean Brier ↓	Comparator Brier	Valid / retained	Status
GPT-6 Astra	-0.009394	0.086055	0.085254	0.232971	0.232645	24/24	Complete 24-case panel
Claude Sonnet 5.5	-0.001802	0.086720	0.086564	0.242598	0.242760	23/23	23/24 cases · subset
Muse Spark 1.3	-0.006200	0.085546	0.085019	0.247577	0.247730	22/22	22/24 cases · subset
DeepSeek V4 Pro	-0.198902	0.102212	0.085254	0.275865	0.232645	24/24	Complete 24-case panel
Kimi K3	-0.222062	0.056772	0.046456	0.397341	0.275796	6/6	6/24 cases · subset
Grok 4.7	0.696575	0.055059	0.181459	0.087065	0.324387	2/2	2/24 cases · subset
GPT-6 Luna	-1.020491	0.172256	0.085254	0.273134	0.232645	24/24	Complete 24-case panel
GPT-5 Nano	-1.109755	0.180896	0.085743	0.209368	0.236960	23/23	23/24 cases · subset
Claude Opus 5.5	-0.014276	0.086471	0.085254	0.237567	0.232645	24/24	Complete 24-case panel
Claude Haiku 4.5	-1.695818	0.229830	0.085254	0.267932	0.232645	24/24	Complete 24-case panel

Subset means and their comparator use exactly the same valid cases. Forecast has one construction family; a family-level confidence interval is unavailable. These raw subset diagnostics are not a complete-model ranking.

Formatting and reliability audit

215 responses were recovered: Haiku 9, Muse 91, DeepSeek 91 and Kimi 24. Repairs remove JSON fences, add the required answers wrapper, separate clearly labeled rationale prose or a rationale-only object, or omit an overlength optional rationale from the scored copy. Original answer values are preserved.

No missing fields are invented, numeric strings coerced, quantiles reordered or answers corrected against the key. Scientific errors and remaining output failures keep their original grading. All 24 Opus refusals remain failures. No new model calls were made.

Model	Recovered	Excluded cases / reason	Retained failures	Retained
GPT-6 Astra	0	1 confirmed network timeout	None	95/96
Claude Sonnet 5.5	0	2 token_limit, 1 confirmed_network_timeout	None	93/96
Muse Spark 1.3	91	5 token limit	None	91/96
DeepSeek V4 Pro	91	5 token limit	None	91/96
Kimi K3	24	4 token limit, 38 insufficient funds, 1 interrupted without answer, 1 authentication error without answer, 14 not returned	None	38/96
Grok 4.7	0	56 confirmed network timeout	None	40/96
GPT-6 Luna	0	None	None	96/96
GPT-5 Nano	0	3 token limit	None	93/96
Claude Opus 5.5	0	None	24 refusal	96/96
Claude Haiku 4.5	9	None	None	96/96

Recorded run settings

Model	Max output tokens	Timeout	Output mode	Optional parameters	Track
GPT-6 Astra	32,768	600s	native	{}	closed_packet_native_json
Claude Sonnet 5.5	128,000	10,000s	native	{}	closed_packet_native_json
Muse Spark 1.3	128,000	10,000s	text	{"reasoning_effort": "high"}	closed_packet_text_json
DeepSeek V4 Pro	128,000	10,000s	text	{"reasoning_effort": "max"}	closed_packet_text_json
Kimi K3	128,000	10,000s	text	{"reasoning_effort": "max"}	closed_packet_text_json
Grok 4.7	32,768	600s	native	{}	closed_packet_native_json
GPT-6 Luna	32,768	600s	native	{}	closed_packet_native_json
GPT-5 Nano	32,768	600s	native	{}	closed_packet_native_json
Claude Opus 5.5	100,000	10,000s	native	{}	closed_packet_native_json
Claude Haiku 4.5	64,000	10,000s	native	{}	closed_packet_native_json

Kimi · account limit and budget

Kimi stopped after the account ran out of funds. Its funding errors, interrupted or missing cases and four token-limit cases are excluded from scores. It has 38 scored cases, including only six forecasts.

Operator-reported budget: the run operator reports that Kimi needed more funding than any other model to complete, with some cases requiring roughly 100× more. Comparable billing records were not supplied, so this remains an attributed budget observation.

Measured token comparison: 11 of 38 shared valid cases used over 100× Haiku’s reported output tokens. For F001, Kimi used 113,525 versus Haiku’s 377 (301.1×). These are reported output tokens, including reasoning where reported, rather than dollar costs. DeepSeek had higher average reported output-token use than Kimi in these exports.

Frozen packet and aggregate provenance

All ten runs match the same 96-case task packet. Original valid-item scores were checked against the supplied scored records. One additional confirmed Astra timeout was excluded under the existing transport-timeout policy.

Task-file SHA-256: efc0911cd357e85c9974648a5216be84e65f310eae61fda0dc3f4f44356e10f3
Answer-key SHA-256: 60be10c1614ad32d14bb237c4ca40026c835cbee2cc12ac697def858357a6d57
Snapshot: 30 Sept 2026 UTC · R9

Questions, answer keys and raw responses are not embedded in this page. Human review remains available by request. The embedded data contains aggregates, resource settings and source hashes only.

Data, review & related work

Transparent methods.
Controlled question access.

Methods and aggregate results remain public. Question packets, answer keys, and runner downloads are no longer linked for public distribution.

What humans can review

The current v0.2 calibration suite contains 96 synthetic cases across evidence, analysis, design, and forecasting. Review packets are shared by request through the organizers and include complete task data and response requirements.

The separately versioned in-vitro proposal category adds 24 constructed dossiers in eight families, each with an initial and revised submission. Dossier evidence and abstract resource units are development material, not experimental observations. Detailed packets, review keys and model submissions remain under controlled access.

Reviewer feedback should identify ambiguity, incorrect assumptions, scoring errors, and limits to scientific relevance. Independent review is being sought; it is not represented as completed.

What remains separate

Answer keys, raw model responses, and prospective experimental readouts are handled separately. The public website presents domain-level results and the conditions under which they were obtained.

Earlier releases may already have exposed calibration questions. A future holdout needs fresh cases, separation by related task families, and a documented access history. Withdrawing a download does not undo prior exposure.

Human review access →Methods & scoring →

Primary sources behind the starter reading cases

Gill et al. (2022), eLife. Transient reprogramming in human fibroblasts; used to distinguish molecular readouts from variable functional responses.
Waziry et al. (2023), Nature Aging. CALERIE methylation analysis; used to distinguish trial populations, biomarkers, and clinical conclusions.
Xu et al. (2018), Nature Medicine. Senolytic research; used to distinguish mouse findings from human tissue-explant evidence.
Lu et al. (2020), Nature. Mouse retinal reprogramming; used to assess the scope of molecular and functional evidence.

Evaluation work we build around

Related efforts include Arc’s Virtual Cell Challenge and the Inspect evaluation framework. The evaluation runner is a small standalone implementation, not an official integration with either project. Our focus is longevity-research decisions linked to prospective functional measurements.

About RejuvenationBench

Make longevity research
a capability worth training for.

We are a team of researchers and volunteers building an evaluation for AI systems that can contribute to rejuvenation and longevity science.

The aim

We want AI developers to improve at evaluating evidence, analyzing biological data, choosing informative experiments, and predicting outcomes that matter for function.

The target is research competence—not expressing support for longevity. A model that catches a flawed hypothesis, prevents wasted work, or correctly predicts a null can be doing valuable science.

Controlled question review and the experimental analysis workflow support scrutiny while the evidence base grows.

A research evaluation.
Not a clinic.

We do not offer therapies, supplement recommendations, or a self-experimentation track. Model proposals are research submissions for institutional review, not instructions for people.

Researcher affiliation describes the team, and it does not imply Stanford University sponsorship or endorsement.

Questions that matter.

Are the lab experiments theoretical?

No. The v0.2 experimental set will run in 2027. Experimental outcomes have not yet been reported. The exact assay details and execution dates will be recorded in the round card.

Does a younger clock count as rejuvenation?

Not by itself. The experimental result must be interpreted against functional endpoints, matched controls, and the registered biological guardrails. Human benefit requires human evidence.

Can humans review the benchmark questions?

Yes. Review materials are shared by request through the organizers. The v0.2 calibration suite still needs independent scientific review and fresh evaluation cases before it can support an official model ranking.

What does an AI company gain by participating?

A reproducible evaluation, actionable failure reports, documented research tasks, and a way to test whether model improvements extend to new longevity-research outcomes.

Will therapies leading to rejuvenation cause overpopulation?

Most developed nations are experiencing rapid population decline. Some nations, especially in East Asia, are on the verge of population collapse. We consider rejuvenation to be closely related to preventative medicine.

How is the live counter calculated?

The display increases at a configured rate of 100,000 per 24 hours from September 5, 2026, the reference date chosen for “Navier-Stokes Day” when OpenAI produced a counter-example to the Navier-Stokes Millenium Prize Problem. It is an illustrative elapsed-time counter based on estimates by scientists of the approximate number of age-related deaths each day. Pausing freezes only the display; resuming catches up to the current total until this benchmark is finished. Experimental outcomes are reported separately.

The first AI benchmark for rejuvenation. The first AI bio benchmark in living things.

Understand the evidence

What they must notice

Get the analysis right

What an “effect” means here

Choose a useful experiment

Examples of the trade-offs

Make a prediction

The uncertainty they must handle

Propose, then revise

First, commit to a plan

Then, respond to new evidence

How the scores work

They check the decisions behind a lab plan.

Every component has to hold.

Why can't we test in living things for all models?

Evidence · fully solved

Strengths, gaps and completed work.

Evidence · partial credit

Analysis · partial credit

Evidence · component detail

Design · partial credit

Forecast · weighted interval score

Round 1: virtual cells and tests.The Top 2 winners: living cells.

Biology Comprehension

Effect Calculation

Identifying Key Points

Accurate Predictions

Proposing and Revising Experiments

A better predictionis not yet a better intervention.

Two questions.Two kinds of evidence.

Benchmaxxing won't get us thereWhy actual biology is our benchmark.

Every day is a reasonto make progress.

Four separate capabilities

Predictions before outcomes

Functional outcomes

A useful experiment.A defensible revision.

Stage 1 · Initial proposal

Stage 2 · Evidence and revision

Documented model track

Tool-using systems

Two layers of assessment.

Check explicit commitments

Review scientific merit independently

Research Core scoring

Intervals have to earn their width.

Missing results do not disappear.

Experimental function is measured against controls.

What happened to “function-days”?

Can the models predict what the cells will do?

What will be locked before unblinding

Freeze the question and the forecasts.

Run the v0.2 experimental set.

Publish predictions and measured outcomes separately.

Controls belong in the lab.

Donors are the units.

A small round has a narrow claim.

Review the questions and assumptions.

Inspect the methods and results.

Benchmark results

Evidence results

Analysis results

Design results

Forecast results

Formatting and reliability audit

Recorded run settings

Kimi · account limit and budget

What humans can review

What remains separate

Primary sources behind the starter reading cases

Evaluation work we build around

The aim

A research evaluation.Not a clinic.

Questions that matter.

Round 1: virtual cells and tests.
The Top 2 winners: living cells.

A better prediction
is not yet a better intervention.

Two questions.
Two kinds of evidence.

Benchmaxxing won't get us there
Why actual biology is our benchmark.

Every day is a reason
to make progress.

A useful experiment.
A defensible revision.

A research evaluation.
Not a clinic.