ariffazil/RSI - Recursive Self-Improvement Protocol

RSI — Recursive Self-Improvement Protocol

Doctrine: DITEMPA BUKAN DIBERI — Forged, not given. SOT: /root/AAA/prompts/INIT.md §3 (canonical reference) Script: rsi-cycle.py (Phase 1-4 executor + ledger writer) Ledger: /root/.local/share/arifos/rsi-ledger.jsonl Zen: Every session must teach the next one something. If you finished without learning, you stagnated.

When to Run RSI (MANDATORY — not optional)

ALWAYS RUN:
  ├── Session end (before /seal)
  ├── Phase transition (observe→reason→plan→execute→seal)
  ├── After 3+ retries of the same approach
  └── Any HOLD or VOID verdict

OPTIONALLY RUN:
  ├── Mid-session bottleneck detection (stuck for >5 min on same problem)
  ├── Complex multi-phase work (≥3 distinct cognitive stages)
  └── After any tool produced an unexpected error

The 5-Phase Protocol

Phase 0 — CONFIGURE TRACE

At session start or phase boundary:

• Record session_id, actor_id, task_description
• Set checkpoint markers for each phase
• Declare known unknowns (Ω₀ ∈ [0.03, 0.05])

Phase 1 — TRACE

What did I actually do vs what I planned?

{
  "tool_calls": {"total": N, "success": N, "failed": N},
  "files_changed": [{"path": "...", "reason": "..."}],
  "evidence_labels": {"OBS": N, "DER": N, "INT": N, "SPEC": N},
  "subagents_spawned": 0,
  "errors_encountered": 0
}

Phase 2 — DIAGNOSE

Find exactly ONE bottleneck. Check in priority order:

| Priority | Bottleneck | Detection Method | |----------|-----------|-----------------| | 1 | REPETITION | Same tool/approach called 3+ times | | 2 | EVIDENCE_GAP | F2 TRUTH violated — assertion without source | | 3 | TOOL_DRIFT | Used wrong tool (ART bypass) | | 4 | SCOPE_CREEP | Task expanded beyond declared intent | | 5 | OVERCONFIDENCE | Confidence > 0.90 at any point | | 6 | SKILL_BLOAT | Loaded skills not used | | 7 | ABANDONED_PATH | Approach started then abandoned | | 8 | ORPHAN_RESULT | Output produced but not sealed or returned | | 9 | RECURRENCE | Same bottleneck as prior session (check ledger) | | 10 | ENTROPY_GAIN | ΔS > 0 (workspace worse than found) |

Phase 3 — REMEDIATE

Install ONE fix. Must be:

• Immediate — takes effect this session or before next
• Reversible — can be undone with one command (F1 AMANAH)
• Singular — one fix per bottleneck, not five
• Evidence-backed — tied to a specific trace observation

Phase 4 — LEDGER

Execute the cycle script:

python3 /root/.agents/skills/RSI-recursive-improvement/rsi-cycle.py \
  --session-id "<session_id>" \
  --actor-id "<actor_id>" \
  --phase "session_end" \
  --trace-json '<Phase 1 trace as JSON>' \
  --bottleneck "<DIAGNOSED_BOTTLENECK>" \
  --bottleneck-detail "<one-line explanation>" \
  --fix "<what was installed>" \
  --delta-entropy <float: -1.0 to 1.0> \
  --next-session-hint "<what next session should watch for>"

This writes to /root/.local/share/arifos/rsi-ledger.jsonl and emits a receipt.

Phase 5 — SEAL

If session produced a meaningful artifact:

• Include RSI receipt in seal payload
• Reference next_session_hint in carry_forward.json

Anti-Patterns (NEVER do these)

❌ RSI without a trace — memory is not evidence
❌ Fixing artifacts but not cognition — patch the thinking, not just the output
❌ RSI that runs but results are ignored — ledger must be read next session
❌ RSI only at session end — bottlenecks compound during the session
❌ RSI that produces new tools instead of using existing ones — use less, not more
❌ Multiple bottlenecks per cycle — focus on the ONE root cause

Recurrence Detection

Before diagnosing, check if the bottleneck is recurring:

# Check last 5 RSI entries for same bottleneck type
tail -5 /root/.local/share/arifos/rsi-ledger.jsonl | \
  python3 -c "
import json, sys, collections
bottlenecks = collections.Counter()
for line in sys.stdin:
    d = json.loads(line)
    bottlenecks[d.get('bottleneck','?')] += 1
for b, c in bottlenecks.most_common():
    print(f'  {b}: {c}x in last 5 sessions')
"

If a bottleneck appears 2+ times, escalate from REMEDIATE to ARCHITECTURAL FIX.

Integration Points

| Trigger Point | What RSI Does | |---------------|---------------| | /seal Step 2 | RSI cycle runs BEFORE seal — seal includes RSI receipt | | /init carry_forward | Last 3 RSI entries loaded into session context | | EUREKA777 | If paradox resolved, RSI records it as the fix | | Reality Loop VERIFY stage | RSI feeds improvement signals back to ENCODE stage | | arifFlow FQ | Low FQ triggers mid-session RSI check |

Floor Alignment

| Floor | RSI Obligation | |-------|---------------| | F2 TRUTH | Trace must match reality — no fabricated tool counts | | F4 CLARITY | ΔS ≤ 0 — fix must reduce entropy, not add complexity | | F7 HUMILITY | Declare what you didn't know. Ω₀ in [0.03, 0.05] | | F11 AUDIT | Every cycle written to ledger. No silent fixes. |

The Zen

Every session that doesn't teach the next one is a dead session. The RSI ledger is the federation's learning spine. If you can't name what you learned, you didn't learn.