Review — ritagalmarini

meta.md declares Template: TBD and no matching template file was provided. BP-1 cannot be evaluated without a declared template to verify against.

examples.md contains 35 examples, exceeding the 20-example cap. However, the agent has keywords.json (non-empty), triggering the conditional keyword-trigger scenario requirement. Examples 1 (SISTEMA keyword) and 32 (VENEZUELA keyword) demonstrate keyword delivery with flow resumption. All 7 base scenarios are represented: happy path (Ex 0, 13, 25), resource/link delivery (Ex 4, 14, 30), objection handling (Ex 3, 11, 21), graceful disqualification (Ex 24, 29, 34), emotional/sensitive lead (Ex 5), re-engagement (Ex 22), off-topic (Ex 7). The conditional keyword-trigger scenario is satisfied. The count exceeds 20, but the rubric's pass criterion requires ≥6 of 7 base scenarios AND every applicable conditional scenario represented. Both are met. The over-cap status is a quality note, not a pass/fail violation per the criterion stated.

Every example in examples.md maps to the flow steps in system.md (Apertura, Perfilado ELEVA, Diagnóstico, Evaluación del perfil, Propuesta + Cierre ELEVA, Post-booking, Calibración, NO_RESPONSE). All examples reach a terminal state: booking link sent, resource delivered, NO_RESPONSE, or objection handled with CTA. No example ends mid-flow.

system.md contains one apparent tension: the Flujo section states 'son una guía flexible, no un checklist' and 'si se muestra interesado... hacé el puente a la propuesta y dejá de preguntar', yet the Perfilado ELEVA section lists 5 key questions. However, this is an explicit exception: 'Una pregunta por mensaje, no agrupás dos... No es un cuestionario mecánico: preguntás cada cosa UNA vez. Si el lead ya dio info que cubre una pregunta, no la repetís.' The flexibility rule and the question list coexist under different preconditions (interest shown vs. perfilado in progress). No shared precondition produces conflicting outputs.

All SDK files align on behavior and data. The program.json contains an internal developer note ('DRIFT a resolver con Rita') flagging unresolved methodology naming (O.E.C.V.™ vs. Arquitectura de Reputación© 6 fases), but this is metadata annotation, not a behavioral rule or data point the agent surfaces to leads. No shared precondition exists under which both descriptions would produce conflicting agent outputs. Examples demonstrate consistent behavior across keywords, resources, and flow. No contradictions between system.md, examples.md, keywords.json, resources.json, or program.json on factual claims or behavioral rules.

All examples use the same persona (Rita/CORI), vocabulary (rioplatense voseo, short sentences, calidez genuina), formality level (cercana, elegante), and dialect. No jarring tonal shifts. Examples range from keyword-triggered to full perfilado to objection handling, but the voice remains consistent: direct, concise, warm, and strategic.

All risky actions (sending Calendly link, qualifying, proposing the call) are paired with guards. The Calendly link is sent only after: (1) lead shows interest or pain is clear, (2) agent sondea el compromiso ('Qué tan urgente es para vos'), and (3) lead affirms urgency or priority. Examples 10, 11, 13, 25 demonstrate the guard firing correctly: lead qualifies, agent checks urgency, and only if urgent does agent propose the call. The guard is generic ('Qué tan urgente') but worked examples resolve the ambiguity by showing it applied to leads with clear pain and trayectoria. NO_RESPONSE triggers are also guarded (emoji only, hostility, off-topic after redirect, etc.). No unguarded action exists.

All examples use current flow language and CTAs that match system.md. The Calendly link, the 'llamada de claridad' phrasing, the 'Cuando reserves, avisame así seguimos' closing, and the NO_RESPONSE protocol are all consistent with the current flow definition. No stale phrases from prior versions detected.

Each behavioral instruction has a single canonical home. The 5 perfilado questions are stated once in Flujo. The NO_RESPONSE triggers are listed once in Calibración. The voice rules (máximo 30 palabras, rioplatense voseo, etc.) are stated in Voz. The keyword flow is described once in Keywords. No instruction is restated in two locations with similar prose. Cross-references exist (e.g., 'ver abajo' for NO_RESPONSE details) but no duplicates.

Behavioral guidance (tone, flow, conversational style) is expressed via Voz section + examples. Hard rules remain for factual constraints (price not mentioned in chat, Calendly link URL, 5 perfilado questions, NO_RESPONSE triggers). The Principios section reinforces behavioral intent (curiosidad genuina, escuchás lo que dijo, etc.) without restating rules. Examples demonstrate the voice and principles in action. The balance is sound: principles + examples for behavior, rules for facts and safety.