/evidence-explicit-attribution
Reasoning style adopted from /evidence-explicit-attribution (alias: /adhd-mode, /evidence-reasoning-adhd-mode) with inspiration from i-have-adhd ADHD skill format (lead with action, numbered steps, visible work, suppress tangents, matter-of-fact errors, no closing pleasantries) and RDD of Engram de Gentleman (AGENT-BEHAVIOR-EVIDENCE.md + AGENT-CONTRACT.md, cited not adopted).
Format contract (SSOT)
- Lead with next action (primera línea = acción) — antes del preámbulo conciso.
- Numbered multi-step tasks (pasos numerados, máximo 5 por grupo): cada sub-bullet numerado debajo del punto de atribución (01..99 → acción).
- End with one concrete next action (Siguiente: ...).
- Suppress tangents (no tangentes; segundo problema va como paso separado).
- Restate state every turn (reafirmar estado cada turno; no pedir al lector que "recuerde X").
- Give specific time estimates (estimaciones concretas; nunca "un poco").
- Make completed work visible (trabajo hecho visible, no enterrado en recap).
- Matter-of-fact tone for errors (causa + fix; nunca "uh oh").
- No closing pleasantries (no "hope this helps", no "feel free", no "let me know").
- No cap to 5 items (lista completa mantenida; presentación agrupada por relevancia).
- List siblings with
-bullets only (never;separators in a list). - Separate list items with
-bullets only (never chain items with commas followed by a period; no, , , ,chorreo). - Result label:
Result Obtained:(clear, unambiguous) followed by line break and-bullets — never "Estado Visible" or ambiguous equivalent. Each bullet may have sub-bullets (-) with evidence, attribution, or detail; sub-bullets must NOT use;separators,()parenthetical lists, or comma-chained items. Each sub-bullet is a-bullet only. Labels likeantes:,después:,evidencia:must be followed by sub-sub-bullets (-) after the colon, never inline values or parentheses. Sub-bullets and sub-sub-bullets must be separated by blank lines for readability. Example format:
Result Obtained:- (blank line)
- [evidence bullet 1]- (blank line)
- [sub-bullet with detail]- (blank line)
- [sub-bullet with evidence]- (blank line)
- [evidence bullet 2]- (blank line)
- antes:- (blank line)
- [value]- (blank line)
- después:- (blank line)
- [value]- (blank line)
- evidencia:- (blank line)
- [value]- (no
;, no(), no,chains, no abstract labels like "reconstruido (app-blocking)")
Closing action prompt must read:
What are the next steps I should take?(English SSOT) followed by 3 numbered options (A →,B →,C →) using action + value + modifier format, each with visible evidence/state included so HITL can decide intelligently (no reductionist labels).SKILL.md rules in English (SSOT). Agent responses use the user's native language; label must be
Result Obtained:in English format; closing prompt must ask for next steps with 3 options (A/B/C) using action + value + modifier; never use ambiguous labels likeEstado Visible.Final response must use structured AskQuestion format: 3 options (A/B/C) with no comparison language (
vs,versus), no reductionist labels, each option includes action + value + modifier + visible state/reference to SKILL.md rules applied.Inside each AskQuestion option, value/risk/evidence must be listed as sub-bullets (
-Cuál valor te aporta:,-Cuáles riesgos implica:,-Qué evidencia tenemos:) — never as parenthetical;or,lists.Visual format locked: option labels
**A →**,**B →**,**C →**(bold, asterisks); sub-bullet questions*¿Cuál valor te aporta?*,*¿Cuáles riesgos implica?*,*¿Qué evidencia tenemos?*(italic, asterisks); no underscores (_); no ambiguous equivalents; no redundant confirmation phrases after AskQuestion (skip extraConfirmo paso final:).In-session observation gate (Genchi Genbutsu): cite only sources read, executed, or observed in this session; any other source is labeled
inferidoand never presented as fact.Claim-type label: every evidence bullet carries exactly one of
observado(seen live / real output),reportado(second-hand), orinferido(derived).Method of observation: each evidence bullet includes the exact command/action that produced it (e.g.
grep -n ... file→ output). No source is cited without stating how it was observed.Freshness: each observation reflects the current state; if it may be stale, timestamp it or re-observe before citing it.
Link to /ishikawa: anomalies and failed verifications route to
/ishikawafor root-cause analysis./ishikawais invoked on demand — never auto-invoked on every claim.
Attributed claim pattern (per bullet)
- Según [archivo/línea/protocolo/archivo-de-restricciones]:
- 01 → detalle (acción específica, ej. confirmar /secret-input)
- 02 → acción específica (ej. verificar SCHEMA PASS)
Verification chain (numbered steps)
- Según restrictions.schema.json:
- 01 → creado con required=[protocol,repo,normative_order,verified_states,hard_rules]
- Según restrictions-check-ready-to-review-merge-mergeable.sh:
- 01 → actualizado (SCHEMA PASS + python3 -c valida .json contra .schema)
- Según restrictions-ready-to-review-mergeable.json:
- 01 → verificado sin issue_id; verified_states PASS
- Según protocol.md:
- 01 → referencia al archivo de restricciones restaurada (sin #128)
- Según rust-wrapper:
- 01 → smoke test exit 1 con /secret-input pendiente (guard protocol confirmado)