Developers, creators & evaluators
We build.
We create.
We evaluate.
We build applications, develop AI agents and workflows, and create 3D work in Blender. We help clients evaluate AI outputs and build evaluation systems with clear rubrics, test cases and review tools.
See how we workExample assessment7/ 7
Deliver an editable scene with its textures included.
The example handoff contains the editable mesh and packed textures. Its reopen check reports no missing files.
- Editable mesh: satisfied
- Packed textures: satisfied
- Reopened scene: satisfied
Illustrative outputs and rubric. Criteria are defined for each project.
Rubric07 / Complete
Complete
The result meets the stated requirements and is supported by the supplied evidence.
From context
to working files.
Technical evaluation across agents, code, research, presentations, and 3D. Follow the instructions. Inspect the work. Explain the result.
Explore the practiceLook through
every layer.
The final output is part of the picture. Context, execution, evidence, and the delivered files all deserve a closer look.
01Responses & context
Conversations / comparisons / instruction following
A response belongs to a conversation. Evaluate what it says, what it leaves out, and whether it still follows the requirements established several turns earlier.
Inspect. Inspect factual support, relevance, completeness, tone, and internal consistency. Track changes to the brief, persistent constraints, and places where an answer quietly substitutes an easier task. Compare the explanation that is actually provided with the evidence available to the reader.
Verify. Trace each requirement to the response. Check claims against sources or supplied context, compare alternative answers on the same dimensions, and distinguish an incorrect statement from an unsupported one. Test whether a correction carries through the next turn.
Return. Return a clear preference or rubric judgment, the passages that support it, the consequence of each issue, and a specific correction. Keep uncertainty visible when the available evidence cannot settle the question.
Compare the proposals using only the supplied evidence.
Proposal B is stronger because of its estimated market share.
Context mismatch
The supplied material contains no market-share estimate. This changes the basis of the comparison.
02Agents, tools & skills
SKILL.md / Codex / execution traces
An agent can produce a convincing final message while leaving the actual task unfinished. Review the instructions, the actions it took, and the artifacts it delivered together.
Inspect. Read the skill’s scope, triggers, prerequisites, and completion requirements. Examine tool selection, arguments, execution order, instruction conflicts, handoffs, and recovery after failures. Check whether the workflow uses the right information at the right point.
Verify. Connect the brief to the visible execution trace. Confirm that commands ran in the intended environment, required files exist, and verification supports the completion claim. Exercise a missing dependency or ambiguous input to see whether the workflow recovers coherently.
Return. Provide a trace-backed explanation of where the workflow succeeded or broke down. Revise the relevant SKILL.md instructions, tool steps, or checks so another run has a clearer path to a verifiable result.
## Deliverable
Export the deck as PDF.
Open the export and check all slides.- 01Read brief & sourcesObserved
- 02Write presentationObserved
- 03Export & inspect PDFMissing
Incomplete handoff
The source exists. The required export and its verification are absent from the trace.
03Code & terminal work
Repositories / automation / reproducibility
Read the change, run the work, and inspect the result. Evaluate software against the requested behavior and the environment where someone will actually use it.
Inspect. Inspect setup instructions, dependency assumptions, commands, diffs, interfaces, and error handling. Look for missing edge cases, accidental changes outside the task, and code that depends on an undocumented local state. Review whether the tests exercise the requirements.
Verify. Reproduce the setup in a clean environment when the task calls for it. Run relevant checks, inspect generated files, and try meaningful boundary and failure cases. A passing test is useful evidence; its value depends on what the assertion proves.
Return. Hand over a reproducible issue, the command or input that exposes it, expected and observed behavior, and a focused repair. Include the revised code and verification notes when implementation is part of the work.
Return the average. For empty input, return None.
def average(values):
return sum(values) / len(values)Boundary failure
Empty input reaches division by zero. Add the missing branch and verify the specified return value.
04Research & data
Sources / documents / tables / calculations
A citation is a starting point. Evaluate whether the evidence supports the specific claim, whether the data is consistent, and what remains uncertain.
Inspect. Inspect source relevance, publication context, conflicting accounts, missing qualifications, and the relationship between conclusions and evidence. Review calculations, table labels, units, denominators, and whether a chart accurately represents the underlying values.
Verify. Open the cited material and trace important claims to the passages or data that support them. Recalculate consequential totals, compare sources with different perspectives, and separate observations from interpretations. Check time-sensitive statements against the requested date or period.
Return. Produce a source map, corrected claims or calculations, and concise notes explaining the strength and limits of the evidence. Organize the result so another person can follow the references and reproduce the important checks.
Supported
Supported
Not established
Claim exceeds evidence
Association alone does not establish the stated causal relationship. Narrow the conclusion.
05Slides & presentations
Narrative / evidence / visual communication
Evaluate a presentation as a sequence and as a usable file. The argument, supporting material, and design all need to serve the audience and the brief.
Inspect. Inspect required sections, slide count, narrative progression, repeated or contradictory claims, and the relationship between headlines and evidence. Review hierarchy, text size, contrast, chart units, source notes, alignment, and consistency across the deck.
Verify. Compare the slides with the source material and the stated audience. Read the sequence without presenter commentary, inspect charts against their data, and check exports at realistic viewing sizes. Confirm that the editable file and final export contain the intended content.
Return. Return slide-specific annotations with reasons and concrete changes: a clearer headline, a missing source, a corrected chart, or a reordered argument. Where useful, prepare revised slides or alternative layouts that address the identified issue.
Missing bridge
The evidence covers one market. Explain the additional assumptions or narrow the recommendation.
06Blender & visual artifacts
Scenes / geometry / materials / renders
A finished image shows one view of the work. Inspect the source scene as well as the rendered result to understand whether the deliverable meets its intended use.
Inspect. Review scene organization, objects, scale, geometry, normals, modifiers, UVs, and material relationships. Inspect lighting, camera framing, external textures, and render or export settings. Relate technical issues to visible effects and the requirements of the brief.
Verify. Check the file with its expected dependencies, inspect the relevant views, and reproduce the required render or export. Confirm that source changes appear in the delivered result and that the receiving person has the assets needed to continue working.
Return. Provide an annotated inspection, the affected objects or settings, and a prioritized list of fixes. When revision is included, deliver the updated scene, required assets, and a clear account of what was checked in the final output.
- ✓ Camera configured
- ✓ Geometry present
- ✓ Render settings saved
- ! Texture unavailable
Unresolved dependency
The material references a local image outside the handoff. Include the asset and reopen the delivered scene.
A judgment
you can trace.
A rating is one possible output. The method depends on what the task needs to establish.
- 01 /
Define the checks.
Translate the brief into observable requirements. Separate correctness, completeness, instruction following, presentation, and usability so a strong result in one area does not conceal a failure in another.
- 02 /
Choose the judgment.
Use pass/fail checks for explicit requirements, dimension-based rubrics for quality, and pairwise comparisons for alternatives. Apply an ordinal scale such as 1–7 when the project calls for it, with clear anchors for each rating.
- 03 /
Attach the evidence.
Point to the response passage, tool action, file, source, slide, or scene object behind the finding. Describe severity in terms of its effect on the task. Distinguish a demonstrated failure from an unverified concern.
- 04 /
Make the next step clear.
Return an actionable correction, then check the revision when requested. Compare similar cases for consistency and use calibration examples to keep judgments aligned across a larger body of work.
Review it.
Build it.
Check it.
Independent technical evaluation and development, carried from the brief through the finished files.
Work can include reviewing existing outputs, building alternatives, revising skills and instructions, writing code and automation, investigating sources, or improving presentation and 3D deliverables. The scope starts with what needs to be produced and how it will be checked.
A useful handoff includes the artifact, the evidence behind important decisions, known limitations, and the steps needed to reproduce or continue the work. Review notes should help the next person act.
See evaluation coverage