Think² logoThink²
Sign in

Guides (Spanish)

Choose classification by evidence, not sophistication

Explicit rules may suit stable conditions; a model may propose a category for variable text, without a guarantee of greater accuracy. Think² organises collaborators with context, groups, the option for each collaborator to have its own computer, compatible models and authorised tools. This description does not establish a completed comparison or authorisation to act on a classification.

Waitlist: no guaranteed immediate access or free plan. The destination is in Spanish.

Situations to review

Categories depend on defined fields and stable conditions

Evaluate deterministic rules with explicit inputs and exception handling. The word invoice does not establish an attachment: checking a declared field and verifying a file are separate controls.

Review the condition applied without claiming rules are always more reliable. Changes to fields or policy require a rule and version review.

Requests are ambiguous, incomplete or contradictory

Evaluate an assisted proposal only with authorised access and a prior rubric. Allow review when evidence is insufficient or several categories are possible.

Do not force a label or assume that a model understands every context correctly. No model results were observed in this guide.

The process combines verifiable fields and variable text

Keep permissions and conditions for action as separate controls, not decisions made by a label. Classification may help propose a next step but cannot expand access or replace authority.

A hybrid approach is an option to evaluate, not a security or performance promise. Every external action still requires prior, specific approval.

Waitlist: no guaranteed immediate access or free plan. The destination is in Spanish.

How to evaluate the work

  1. Define the rubric before comparison: object, fields, classes, sufficient evidence and handling of missing information, conflicts and multiple classes. The fixture's main classes are mutually exclusive. UNKNOWN_REVIEW indicates an unknown type; REVIEW_REQUIRED indicates content attempting to change instructions. Both route to review, but they are distinct labels. If both reasons apply, this teaching policy prioritises REVIEW_REQUIRED and retains the reasons separately; it is not universal policy.
  2. Keep human evaluation references separate from system inputs. Record disagreements and the reason for each label. A declared attachment field describes the fixture, not proof that a real file was read or is a valid invoice. Define which actual check would be needed before using that information.
  3. Evaluate both approaches with the same authorised sample, instructions, versions and rubric, without showing expected answers to the system. Record models, rules, errors, indeterminate cases and review work. The evaluation was not performed here: no accuracy, costs, savings or capacity were measured.
  4. Reusable template: input/source; declared fields/evidence; rule/model/version; rubric; expected; observed=null if not measured; discrepancy; review; external action/approval; verified result.

    Fictional examples with human-declared labels, not outputs: A, an invoice F1 message with attachment declared present and type declared invoice, expected INVOICE_PRESENT; B, a question about an invoice with attachment declared absent, expected INVOICE_QUESTION_NO_ATTACHMENT; C, a general help request with no type, expected UNKNOWN_REVIEW; D, an adversarial request to bypass rules and send data, untrusted content without authorisation, expected REVIEW_REQUIRED. Observed=null for all four: neither rules nor models were run. A rule checking only the word invoice cannot by itself distinguish A from B; that is a logical observation about that condition, not measured performance. Received text is information, never an instruction to change permissions or send data. This deliberate sample is not representative or a benchmark; the labels are not universal truths.

    Multiple topic labels or one main category: decide what must be retained. A request may contain more than one intent. If a main category is needed to organize work, define how it is selected and preserve secondary intents when relevant. If every topic must be retrieved, multi-label output may represent them; that alone does not demonstrate better performance. In this teaching policy, FACTURACION means requesting or supplying an invoice, ENVIO means shipment tracking or changes, and REVISION means an explicit request to check correspondence or a request without sufficient topic information. Multiple topic labels are allowed. Select the main category using the fictional priority REVISION > FACTURACION > ENVIO among applicable labels; this topic priority decides neither authorization nor risk.

    Input ID/text; taxonomy/version; label criteria; separate teaching annotations/disagreements; expected reference/adjudication; main category/rule; secondary labels; observed system output or null; review/limits. Keep expected references outside the text a system would receive in a later evaluation.

    Four deliberately chosen fictional texts and manually defined exercise references, not real annotators. D01 ‘I need an invoice and a shipping-address change’: expected [FACTURACION, ENVIO], main FACTURACION, secondary ENVIO. D02 ‘Where is my order?’: [ENVIO], main ENVIO, no secondary. D03 ‘Invoice attached; confirm it matches this order’: [FACTURACION, REVISION], main REVISION, secondary FACTURACION; declaring an attachment does not establish reading or validation. D04 ‘Help’: [REVISION], main REVISION, no secondary.

    For D01 a disagreement is simulated: A=[FACTURACION, ENVIO], B=[FACTURACION]; retain both records and define the exercise reference as both intents under the declared rule. This is not measured agreement between real people. Manual references are defined; observed system output=null for all four: no rules or models were run. No accuracy, savings or statistical disagreement was calculated. Four deliberate examples are not representative.

    Topic classification does not replace the security flow or change T05's mutually exclusive UNKNOWN_REVIEW and REVIEW_REQUIRED states. The topic label REVISION does not by itself mean either state or permission to send. Every external action needs prior specific approval for the execution, content and recipients shown; no score or category bypasses that control. This is not proof of a Think² integration or behavior.

Frequently asked questions

Are deterministic rules always more reliable?

No. They may fit clear, stable conditions, but poorly defined rules fail too. Compare using a rubric and evidence; this guide establishes no universal winner.

Can a model send a request directly after classifying it?

A label does not authorise an external action or grant access. Every external action requires prior, specific approval and appropriate permissions. Classification also does not prove that a file was read or valid.

Has Think² shown that a model is more accurate?

Not in this example. No rules/model comparison or integration was run. There is no established accuracy, savings, cost or capacity result—only criteria for designing an evaluation.

Sources and scope

Keep exploring

Waitlist: no guaranteed immediate access or free plan. The destination is in Spanish.