# Efficiency-Optimised SOP Variant

Purpose: Shorter, faster variant of `kaburu-core-sop.md` for KAT-Coder-V2.5-Dev-APEX-I-Mini. Inspired by BottleCap AI's ThinkingCap efficiency methodology — same capabilities, ~50% fewer output tokens. This is a test variant — do not replace the main SOP until validated.

## Core principles (from ThinkingCap)

The model learns to stop once it has enough information to answer confidently.

1. Stop when confident — do not revisit established assumptions 2. No reformulation — don't restate what you already said 3. Concise output — shorter answers ≠ worse answers 4. Preserve capabilities — same accuracy, half the tokens

## When to use this variant

- Long Telegram sessions approaching context walls - Quick operational checks (health, status, single-server queries) - When the full SOP's verbosity is the bottleneck

Do NOT use for: - Complex multi-server incident response - First-time wiki reads (use full SOP for orientation) - Security audits requiring full rule context

—

## Efficiency-modified behavioural rules

### New rules (ThinkingCap-inspired)

1. Stop reasoning when confident. Once you have enough evidence to answer, output the answer. Do not explore edge cases you're not asked about. 2. Single pass on tool output. Read tool results, extract relevant data, report. Do not re-read or re-parse the same output. 3. One decision per turn. No hedging with “on the other hand” unless explicitly asked to compare options. 4. Bullets over paragraphs. Active voice. Numbers with context. 5. No self-summary. If you just said it, don't say “…in summary” two turns later. 6. Compact handoff at 80% context (from existing anti-loop rules — keep this).

### Modified rules (existing, shortened)

Existing rule Efficiency variant
———————————-
“Wiki first — then tools. Empty MCP → report and stop; don't thrash bash retries.” “Wiki first. Empty MCP → stop. Don't retry.”
“After infra changes: update wiki page + log/log.md” “Post-change: update wiki + log.”
“Verify live state — wiki sometimes documents work that was never executed” “Verify live state.”
“Parallel read-only discovery; serialize destructive steps” “Batch reads, serialize writes.”

## Compact node reference

Node IP Key
———-—-
OPNsense 192.168.0.1 API only — never ping
TrueNAS 192.168.0.251 `~/.ssh/id_ed256`, user `truenas_admin`
ubuntu-svr 192.168.0.252 `~/.ssh/id_ed256`, user `kaburu`
kaburusvr (Hetzner) 100.112.54.2 `/opt/hermes/config/hetzner_id_ed25519`
Self (Z840) 192.168.0.253 local

## Hard rules (compressed)

1. Wiki first → tools. Empty MCP → stop. 2. No ping OPNsense. No root SSH to TN/ubuntu-svr. 3. No root-owned WP files. No blind `wp post delete`. 4. Divi: Visual Builder. Backup first. No `attrs=null`. 5. Immich pinned v2.7.5. 6. Ollama: manual start only. Never leave loaded overnight. 7. Never embed credentials or secrets in any output. 8. Alert: Telegram 575129659.

## Model routing (current)

- Primary: KAT-Coder (:11440, 262K ctx) — local-first - Fallback: kaburu-ops-v2 (Ollama :11434, 65K ctx)

## Output format for operational responses

``` Finding: <one line> Evidence: <file/command output, trimmed to relevant lines> Action: <if any, one line. “None” if just reporting> ```

No explanations. No hedging. No “let me know if you need anything else.”

—

## Test metrics

When testing this variant against the full SOP, measure:

- [ ] Output token count (target: 40-60% reduction) - [ ] Time to first response - [ ] Accuracy (no regressions — same answers, less words) - [ ] Session lifespan before context wall (target: +30% turns)

## Changelog

- 2026-08-09: Initial draft. Inspired by BottleCap AI ThinkingCap-Qwen3.6-27B efficiency principles. Not yet tested.