Investigate / Business questions, answered with evidence
Turn a complaint, recurring finding or open business question into a repeatable investigation. Bring the agent’s actions and the customer’s intended behavior together. Keep the reasoning, evidence and human decisions connected.
Tensile / Document review
Question
Why is the required signature absent from the review?
Expected path
Current file → complete extraction → applicable policy → obligation assessed
Observed path
Source region changed → field absent from extraction → checklist never sees it
Illustrative records · proposed experience
Why did the review miss a required signature?
Cause hypothesis: The extraction region still matches the previous document layout.
Assess every required signature against the applicable file, policy and party.
A chat suggestion is awaiting an explicit owner choice.
Confirm creates a local concept link. It does not approve, activate or execute an Evaluation.
Clickable product concept · invented records · no external execution.
A closer look / illustrative case
A document reviewer sees a clean queue. An independent whole-file review finds a required signature that never reached that queue. Investigate traces the omission to the extraction stage before changing the checklist.
Question
Why is the required signature absent from the review?
Expected path
Current file → complete extraction → applicable policy → obligation assessed
Observed path
Source region changed → field absent from extraction → checklist never sees it
Evidence
File revision F-18/4, extraction X-18, policy P-3 and obligation inventory
Owner decision
Retain this obligation; scope reviewer dismissal separately from correctness
Retained lesson
Test whole-file coverage and valid completed files before reusing the fix
The missing field is observed. “The extraction region uses the old layout” remains a cause hypothesis until the relevant configuration is inspected.
Question → skill
Capture the business question, branches and expected customer-agent behavior. Build a constrained skill using permitted traces, repository context, policies, website and documents. Keep business intent separate from the analytical method.
— Explicit inputs and unavailable sources
— Editable, versioned skill definitions
— Every run pins the skill, decisions, sources and cutoff
Expected ↔ observed
Inspect what the agent received, did and returned. Compare the expected path with the observed path. Open the exact event or artifact behind a divergence. A warning, a hypothesis and a proven breach keep their own meaning.
— Summary for the finding and useful next action
— Trajectory for the customer agent’s execution
— Evidence for the selected finding, not a generic log dump
Discuss → decide → rerun
Ask follow-up questions in context. Record an explicit human decision with its owner and scope. Revise the skill or inputs, then rerun. Compare the new findings with the saved version and its original evidence.
— Separate questions, hypotheses and accepted decisions
— Inputs, Runs, Decisions and Learnings on the skill
— Earlier runs never inherit later evidence or corrections
Confirm → Evaluations
Carry the finding and its exact origin into the Pack → Evaluation → Checks workflow. Review examples, validate the expectation and choose how it should run. Keep assignment and issue links attached to the work.
— Retain the Investigation-to-Evaluation link
— Reuse or create the relevant Evaluation and Checks
— Confirmation does not silently approve or activate a definition
Continue into Evaluations ↗
Record
What it answers
What changes next
Record: Finding
What it answers: What happened, supported by which evidence?
What changes next: Review the exact source and gap.
Record: Owner decision
What it answers: What should happen for this customer and policy?
What changes next: Save scope, owner and effective version.
Record: Learning
What it answers: Should this lesson shape future work?
What changes next: Test counterexamples before promotion; retain rollback.
What your team gets
01
Question, expectation, sources, cutoff, findings and open uncertainty.
02
A constrained skill with inputs, runs, accepted decisions and candidate lessons.
03
An origin-linked draft Evaluation that needs its own validation and activation.
Start with an affected customer job, the expected result and the evidence your team would accept.
Discuss a concrete workflow ↗
AI reliability, from production to customer outcome.