Examining Discovery Spaces

Context

Operations and DiscoverySpaces

• discoveryspaces (or spaces for short) define a set of points (entities) and how to measure them. They also contain the results of the measurements
• operations operate on discovery spaces either selecting or measuring points or analysing existing measurements

Terminology: Distinguishing Entities in a DiscoverySpace

When working with the data from a discoveryspace the following distinctions are important.

• Measured: These entities have been measured by an operation on the space
• Unmeasured: These entities have not been measured by an operation on the space

The samplestore used by a discovery space is shared. This means there may be relevant measurement data in the samplestore for entities in the space but that measurement has not been performed by an operation on the space (it was performed on another).

• Matching: Data in the spaces samplestore that matches the space definition - it includes measured entities
• Missing: Entities that have no matching data in the samplestore.

Why is it useful to work with matching data?

1. Allows using the discoveryspace as a view to fetch particular data without having to perform operations on it
   • Concrete example: You create a discoveryspace that is a subspace of another sampled spaced to analyze it. You can perform analysis on existing data even though no operation has been run on the new discoveryspace.
2. Memoization: You can understand if there are memoization opportunities that would speed up a operation on the space.

Pre-requisites: The Space Identifier

To apply this skill you need either:

(a) a space id; (b) explicit instruction to examine “the latest” space

In the case of (b) get the actual identifier:

uv run ado show related space --use-latest

Tips

Avoiding refetching YAML

ado get … -o yaml writes YAML to stdout by default. Prefer --output-file PATH (with the same -o yaml) to save it once and reuse the file instead of calling ado get repeatedly for the same resource.

Large output files

The output produced for a given -o/--output format can be very large (for example from show entities). Use --output-file with the path where the output should be saved, and when inspecting these files:

• Use wc to count the file size first before using head/tail/cat etc. on it.
• Use head -n1 to get column headers, this will not be large
• Avoid head -n > 1 unless you have a specific need e.g. checking if file is corrupted
• Avoid tail unless you have a specific need
• Prefer python e.g. pandas.read_csv for any detailed analysis on the file.

Workflow

Run Step 2 and 3 first. Then steps 4,5 and 6 can be run in parallel.

Step 1: Get Space YAML

uv run ado get space SPACE_ID -o yaml --output-file SPACE_ID.yaml

Extract and summarise:

• Resource identifier and metadata (name, description, labels)
• sampleStoreIdentifier: the sample store backing this space
• entitySpace: dimensions — property names, types (categorical / discrete / continuous), and their domains/values
• experiments: actuator and experiment identifiers that define what can be measured, and which target properties each experiment produces

Step 2: Sampling coverage and related resources

Execute

uv run ado show details space SPACE_ID

This outputs two sections:

DETAILS — sampling coverage:

• Total entities in the space
• How many have been measured
• How many have failed measurements
• How many are unmeasured
• How many are matching

Compare measured vs total to understand exploration progress. Compare measured vs matching to understand memoization opportunities. Also, a signal that other overlapping spaces exist.

RELATED RESOURCES — all operations and stores linked to this space.

Performance note: ado show details space is slow as it fetches and aggregates entity data. Use only when sampling coverage is needed.

Step 3: Check for existing report

• Check if there is an existing report for this space in reports/<ado_context_name>/
• If yes, check if either of the following are true:
  • New operations have been run on space since report
  • The number of measured entities has increased
• If neither of above are true, ask the user if they want to write a new report or use existing
  • As nothing has changed, the only purpose of creating a new report is if a different agent is being used

Step 4: Find Similar spaces

ado get space --matching-space-id SPACE_ID --details finds spaces with the same entity structure. Use this to understand research progression and why this space was created.

uv run ado get space --matching-space-id SPACE_ID --details

Step 5: Export Measurement Data

Note: Keep in mind the guidelines on large output files for the following.

uv run ado show entities space SPACE_ID \
  --include measured \
  --property-format target \
  -o csv --output-file SPACE_ID_entities.csv

This writes the data to SPACE_ID_entities.csv. If you find SPACE_ID_entities.csv already exists do not use it, as data may be stale

You can also get lists of all unmeasured or missing entities, though this is not typically required unless you want to analyse the unsampled portion.

Perform an analysis of the measurements, checking e.g. distributions of metrics, metric outliers, correlations between metrics. Take into account the domain of the experiment and meaning of metrics when looking for patterns.

Step 6: Examine Related Operations

For each related operation (output in step 2), use the examining-ado-operations skill to understand what each operation did and what it produced.

Note: Do not analyze the data in the operations, or do detailed diagnoses. Just enough for summary.

Producing a Report

Structure the report as:

1. Overview: What the space represents. Infer from metadata, dimensions, and experiments. Short and narrative.
   • Space summary – ID, metadata, entity count, dimensions (parameters and their types/values)
   • Measurement space – experiments, target properties
2. Related Spaces (Optional): If there are related spaces, describe them and how they relate.
3. Sampling coverage – sampled vs unsampled vs missing counts; progress assessment
4. Data summary – distributions of measured properties, notable performers, outliers, correlations
5. Related operations – which operations ran on this space and their status