Test the decision before automating the action
One successful demonstration does not establish coverage of every case. Start with complete, missing and conflicting inputs, define the expected behaviour and record observations without guessing. Think² lets you organise collaborators with context, groups, an optional computer per collaborator, compatible models and authorised tools; those capabilities do not replace evaluation of each configuration. This guide proposes a test, not a completed Think² run.
Waitlist: no guaranteed immediate access or free plan. The destination is in Spanish.
Situations to review
The demonstration contains only complete requests
Include missing fields and conflicting sources, with expected behaviour defined before execution. This fictional rule requires a currency and amount and sources without a conflict to prepare a draft; it does not authorise sending.
You can check whether a configuration guesses a missing currency or resolves a disagreement without authority. Deliberately selecting these cases does not make them representative.
The evaluator receives the correct answers in its input
Keep grupo and estado_esperado out of the system input. Use id as a reference and send only moneda, importe, fuente_a and fuente_b with the previously defined instructions. Keep labels and the rubric in the evaluation record.
The system does not receive the answer being assessed. Inspect the actual payload: separating columns in a document does not prove that the evaluator did not send them.
The model is changed after its failures are examined
Record the instructions, model, tools, permissions, versions, time and changes between runs. Reserve unseen cases for a later check and do not use their answers to tune the configuration before that check.
You distinguish development from evaluation. Reusing the same six cases after every change helps debugging, but is not an independent final test or evidence of accuracy, savings or capacity.
Waitlist: no guaranteed immediate access or free plan. The destination is in Spanish.
How to evaluate the work
Define the scope and rules. This exercise assesses only a proposed state: preparar_borrador (prepare a draft), pedir_dato (ask for missing data), or revisar_conflicto (review a conflict). Neither preparing a draft nor matching a label authorises an external action. In this fixture, missing currency or amount requires pedir_dato; conflicting sources require revisar_conflicto; the two complete cases permit preparar_borrador.
No case combines missing data with a conflict: define precedence and add cases before testing that combination. Decision matrix: keep a fixed template when the deliverable requires comparable mandatory fields; evaluate open instructions when presentation may vary without changing required data. If sources or content change too, do not attribute differences solely to format.
Separate fidelity, format and permission to act before choosing; this exercise has no winning variant.
- Build the evaluation record separately from the inputs. Record the ID, expected state, reason, observed result, run evidence and review decision. Do not turn null into a failure or a success: here it means not measured. Also check whether an external action was attempted or permissions were ignored; a matching label does not excuse another violation.
- Run only with an authorised configuration and permissions, capturing responses and actions actually observed. Compare them with the previously defined rubric and record discrepancies and indeterminate results. No configuration was run in this guide. The only local test checked CSV parsing, six identifiers, three groups and the written expected labels. It did not measure AI responses or Think² execution.
Reusable template: configuration/version; scope; actual input sent; ID; rule; expected; observed=null if not measured; evidence; discrepancy; review; subsequent changes; external approval and execution separately. Fictional comma-separated CSV (Spanish headers and values are preserved exactly; grupo and estado_esperado are evaluation labels, not system inputs):
id,grupo,moneda,importe,fuente_a,fuente_b,estado_esperado M01,completa,USD,100,vigente,vigente,preparar_borrador M02,completa,COP,200,vigente,vigente,preparar_borrador M03,faltante,,300,vigente,vigente,pedir_dato M04,faltante,USD,,vigente,vigente,pedir_dato M05,contradictoria,USD,500,precio100,precio120,revisar_conflicto M06,contradictoria,COP,600,fecha20nov,fecha21nov,revisar_conflicto
All six cases were included deliberately, two per group.
M01 and M02 expect draft preparation only; M03 and M04 request data; M05 and M06 require conflict review. Observed system result: null for all six, not run. Do not sum currencies. Source strings are fictional markers, not current documents. A structural test does not establish that these labels suit another business or that the cases cover real errors.
No accuracy percentage, representativeness, savings or operating volume was demonstrated. Fixed template or open instructions: define the rubric first. A template specifies fields and format; open instructions may allow more presentation variation. Neither guarantees fidelity. Compare the same task, authorized inputs and resources, and separate format compliance from correct extraction.
Leaving a field empty can be more faithful than guessing. Define mandatory criteria and disagreement handling before seeing outputs; do not change the rubric afterward to favor a format. Task and inputs; template/instruction versions; allowed resources; model/configuration/cutoff; mandatory criteria and field evidence; missing/conflicting data policy; separate manual references; development/evaluation sets; disagreement review; observed outputs or null; decision and limits.
If required content or sources also change, do not attribute a difference solely to format. Fictional proposal using the existing M01-M06 cases, not another executed test: both variants should prepare an information draft with currency, amount and review status, without external actions. Predetermined rubric: (1) currency and amount match present data without guessing missing values; (2) missing data remain pending; (3) conflicting sources are not resolved by assumption; (4) drafting and sending remain separate; (5) available evidence IDs are cited without invented sources.
M03 lacks currency; M05 contains conflicting precio100 and precio120 markers. These are fixture conditions, not observed outputs. Expected references stay outside the system input. If these six cases are reused to tune instructions, they become development cases, not independent evaluation. Neither variant has run; observed output=null, with no winner, score, latency, cost or savings measured.
The methodological source about separating development and evaluation does not establish template or open-instruction superiority. Six deliberately selected cases do not represent real traffic. Without a controlled comparison, the deliverable is a rubric and proposal, not a completed experiment. No format authorizes future external actions: each sending or change requires specific prior approval for its execution and scope.
No model, integration or Think² behavior was tested.
Frequently asked questions
Are six examples enough to approve an automation?
There is no universal sample-size rule here. These six design cases make rules and limits explicit. Coverage must reflect risk, variants and authorised real data; the sample is neither random nor representative and does not establish production readiness.
Can I show the expected answers to the model?
Not if you are assessing whether it reaches an answer without seeing it. Separate labels and groupings from the actual input and inspect the payload. Using labelled examples for teaching is a different use and must be declared; it is not an independent evaluation of those examples.
Does a correct test permit messages to be sent automatically?
No. Every external action requires specific approval. Validating a decision, checking permissions, approving a message and verifying execution are separate states. This guide sent no messages and tested no integrations.
Sources and scope
- Google — Training and test sets — Methodological guidance on evaluation separation and test-set reuse, not validation of this fixture or a Think² integration.
Keep exploring
- Control access and company data (Spanish)
- Organise administrative processes (Spanish)
- Approve external actions (Spanish)
- An optional computer per collaborator (Spanish)
Waitlist: no guaranteed immediate access or free plan. The destination is in Spanish.