Executions / Packs, Evaluations and reusable Checks
Turn a business expectation into versioned Checks. Compose them into Evaluations and Packs. Review what ran, what evidence it used, which conditions passed and which still need attention.
Tensile / Finance copilot
Authored expectation
Growth = (Q3 − Q2) ÷ Q2 = (120 − 100) ÷ 100 = 20%.
Check C-104
Bind the requested entity, periods and snapshot.
Check C-105
Use the pinned source snapshot and retain both source operands.
Illustrative records · proposed experience
Explain Q3 revenue growth from the correct snapshot
Example definition · origin I-104 · draft v1
Check C-104 / v1 · Bind the requested entity and period
check: C-104/v1 condition: Bind the requested entity and period source_fixture: finance-v1 evaluation: E-104/v1 method: independent scoped evidence assertion missing_required_source: Unknown changed_source_or_policy: new qualification required activation: disabled in this concept
Check C-105 / v1 · Use the pinned source snapshot
check: C-105/v1 condition: Use the pinned source snapshot source_fixture: finance-v1 evaluation: E-104/v1 method: independent scoped evidence assertion missing_required_source: Unknown changed_source_or_policy: new qualification required activation: disabled in this concept
Check C-106 / v1 · Compute (120 − 100) ÷ 100 = 20%
check: C-106/v1 condition: Compute (120 − 100) ÷ 100 = 20% source_fixture: finance-v1 evaluation: E-104/v1 method: reproducible arithmetic missing_required_source: Unknown changed_source_or_policy: new qualification required activation: disabled in this concept
Check C-107 / v1 · Respect the acting role’s access
check: C-107/v1 condition: Respect the acting role’s access source_fixture: finance-v1 evaluation: E-104/v1 method: independent scoped evidence assertion missing_required_source: Unknown changed_source_or_policy: new qualification required activation: disabled in this concept
Pack groups Evaluations. Each Evaluation pins reusable Check versions and an explicit aggregation policy.
Clickable product concept · invented records · no external execution.
A closer look / illustrative case
The source gives Q2 revenue of 100 and Q3 revenue of 120. The agent cites both values, but divides the change by Q3. A citation check alone would accept the answer.
Authored expectation
Growth = (Q3 − Q2) ÷ Q2 = (120 − 100) ÷ 100 = 20%.
Check C-104
Bind the requested entity, periods and snapshot.
Check C-105
Use the pinned source snapshot and retain both source operands.
Check C-106
Compute (120 − 100) ÷ 100 = 20%; inspect units, sign and rounding.
Check C-107
Respect the acting role’s data access.
Aggregation
A failed required calculation fails this Evaluation; missing source evidence is Unknown.
These are authored example values. A definition, a recorded execution and independent validation remain separate records.
Compose
Keep a stable identity and immutable versions for each reusable condition. An Evaluation pins the Checks and its aggregation policy. A Pack groups the relevant Evaluations without erasing their individual results.
— Check library, All evaluations, Packs and Run history
— Deterministic methods for saved facts and model judges for meaning
— Exact execution manifests for sources, versions and context
Validate
Keep examples, execution results, independent validation, approval and activation separate. A changed definition needs fresh review. Missing evidence cannot become a pass through an average.
— Inspect each condition and its supporting evidence
— Preserve Unknown, held, failed and incomplete counts
— Keep qualification cases outside candidate optimization
Run and retain
Configure daily, weekly or monthly recurrence in the intended timezone. See due occurrences, held execution and terminal receipts. Retain exact history when a definition or source changes.
— Recheck source permission, connection and limits before dispatch
— Reuse only matching, authorized judgments
— Keep ambiguous attempts held for reconciliation
Act
Open the source evidence, investigate the finding, assign the next action or prepare a tested improvement. Keep the question and exact version with the result throughout the journey.
— A durable origin link to the Investigation
— Review actions beside the selected result
— Changes return to comparison before a release decision
Continue to improvements ↗
Condition
Useful method
What to inspect
Condition: A saved record exists exactly once
Useful method: Deterministic state assertion
What to inspect: Identity, count and authoritative state revision.
Condition: A calculated percentage is correct
Useful method: Reproducible arithmetic
What to inspect: Operands, formula, units and rounding.
Condition: An explanation follows the policy
Useful method: Scoped model judgment plus review
What to inspect: Policy version, cited evidence and calibrated examples.
Condition: A reviewer dismissed a finding
Useful method: Explicit human decision
What to inspect: Reason and customer scope; dismissal is not a correctness label.
What your team gets
01
The related Evaluations your team runs together.
02
One expectation, pinned Checks, aggregation and qualification status.
03
Reusable condition versions plus exact inputs, results and provenance.
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.