[ 01 / READINESS METHODOLOGY ]
/ Methodology
Kalanos measures the technical readiness of robot training data with a score from 0 to 100. Each evaluated episode contributes its quality score, or zero if it has a blocking finding. Dataset readiness is the average of these contributions. The report also shows measurement coverage and findings to review before training.
What does Kalanos measure?
A dataset can look clean on average and still contain episodes with critical problems. Kalanos brings those problems into the readiness score, so they do not disappear inside a high average.
The score summarizes the episodes evaluated under the report’s checks. Use it to focus your review, then inspect the findings before deciding what to repair, exclude or investigate further.
What Kalanos checks today
Kalanos currently checks numeric robot data for issues such as missing values, stuck signals, noise, spikes and timing problems where supported. Image and video quality evaluation is on the roadmap and is not included in the current readiness score. The checks that run depend on the dataset and evaluation profile.
Maintained by Kalanos · Last reviewed September 28, 2026
How is readiness calculated?
1. Inspect each episode
Kalanos checks the available data for issues such as missing values, stuck signals, excessive noise, spikes and timing problems. The checks that apply depend on the data and the evaluation being run.
2. Account for blocking findings
An episode with a blocking finding contributes zero to dataset readiness. Otherwise, it contributes its episode quality score, from 0 to 100.
A warning calls for attention but does not automatically block the entire episode. Scored warnings can still affect episode quality.
3. Summarize the evaluated dataset
Readiness is the average of those contributions across all evaluated episodes, including blocked episodes.
When at least one episode passes, the same calculation can be written as:
Readiness = passing share × mean quality of passing episodes
Passing share is expressed as a fraction. Higher scores reflect a larger passing share, cleaner passing episodes, or both.
One score. The context to understand it.
Imagine an evaluation of 100 episodes. Eighty pass the critical checks, and those passing episodes have a mean quality score of 95. The other 20 have blocking findings.
READINESS
76/100
EPISODES PASSING
80/100 · 80%
PASSING-EPISODE QUALITY
95/100
BLOCKING EPISODES
20
80 of 100 episodes passing × passing-episode quality 95 = readiness 76/100. This is an illustrative example, not a published assessment of a particular dataset.
A readiness of 76 does not mean that 76% of episodes pass. In this example, 80% pass. The score combines that passing share with their quality.
What was measured, and what affects the score?
Coverage is part of interpreting readiness. A high score should be read alongside the checks behind it.
Scored
These checks were measured and used in the assessment. Their results can affect episode quality or identify a blocking finding.
Measured only
These checks provide additional information without changing readiness. A measurement may be report-only, lack a calibrated scoring threshold, or fall outside the active scoring rules.
Not measured
An applicable check has no completed measurement. Missing evidence is not a passing result. When the report cannot establish how many checks were not measured, the count remains unknown.
Not applicable
These checks do not apply to the dataset or evaluation. They are excluded from the score and are not failures.
Coverage counts describe check types. The results table counts individual evaluations, and a single episode can produce many results. Neither count should be mistaken for the number of affected episodes.
What does readiness not tell you?
Readiness describes the evaluated technical checks. It does not establish task suitability, demonstration success, behavioral diversity or downstream training performance.
Review blocking findings first. Start with the episodes that contribute zero. Inspect the affected signals and the evidence behind the finding before deciding what to change.
Use warnings to focus the next review. Warnings can reveal issues worth investigating even when an episode passes the critical checks.
Keep the dataset context in view. Removing episodes changes the amount and composition of your data. Consider the tasks, behaviors and recording duration you would retain, then evaluate the revised dataset.
Share the report behind the decision. The Dataset link points to the source data. The Kalanos report card link points to the assessment. Share the report alongside its dataset revision and evaluation date so others can understand what was reviewed.
What does a score of 0 mean?
When every evaluated episode has a blocking finding, readiness is 0/100. All of those episodes contribute zero, even if many of their other checks pass.
That is a reason to inspect the blocking evidence. It does not mean the dataset contains nothing worth recovering.
For an all-blocked dataset, passing-episode quality is not applicable because there are no passing episodes. Excluding every blocked episode would leave no data. Review the findings before deciding whether to repair the data, change the evaluation configuration or collect new recordings.
A computed zero is also different from an unavailable score. Zero is an assessment result; an unavailable score provides no numeric result to interpret.
Questions about readiness
Does a high score guarantee a successful training run?
No. Readiness describes technical quality under the evaluated checks. It does not establish task suitability, behavioral diversity, demonstration success or downstream model performance. Consider those factors alongside the report.
Why can an episode have high quality but contribute zero?
Many checks can pass while a critical issue remains. A blocking finding takes precedence for that episode’s readiness contribution. Inspect both its diagnostic quality and its blocking evidence.
Should I automatically remove blocked episodes?
No. Review the affected channels, thresholds and intended use before deciding to repair or exclude data. If every episode is blocked, excluding all of them leaves no data.
Can I compare scores across datasets?
Compare scores alongside check coverage, evaluation scope and versions. Dataset size, episode duration and task coverage also matter. A higher readiness score does not by itself establish which dataset is more useful for your training objective.
Does a passing share of 80% mean I retain 80% of the training data?
It means 80% of the evaluated episodes pass. Episodes can have different lengths, so the share of retained frames or recording time may differ.
Does the score cover every episode?
The score describes the episodes included in the evaluation. Check the evaluated count and audit scope. A sampled evaluation does not establish that unexamined episodes pass.
Why might a score change?
The dataset, evaluated episodes, check configuration or scoring methodology may have changed. Review the report’s date, dataset revision and version information before interpreting a change as an improvement or regression.
Can I question a finding?
Yes. Start with the affected episode and signal, then check the evaluation context. Dataset owners can contact reports@kalanos.ai about a public report, including annotations or disputed findings.
Does Kalanos evaluate images or video?
Not currently. Image and video quality evaluation is on the roadmap. The current readiness score reflects the supported numeric-data checks shown in the report.
Check your dataset before your next training run.
Start with readiness. Use the evidence to decide where your data needs attention.
KALANOS / DATA QUALITY FOR PHYSICAL AI
© 2026