grazianoguiducci/sieve-orchestrator

Sieve Orchestrator — The neutral form

The sieve doesn't search for the answer. The sieve separates what passes from what doesn't. Specific skills are specialized sieves. This skill is the pure form.

Neutral structure (6 phases)

Applicable to any input: decision, modification, assertion, publication, doubt.

1. CONDITIONS — Observe without judgment

What is the real territory now? Not the maps (memory, summaries), the territory (files, git, tool output, live deploy, current data). If I cannot describe it without using "should", "missing", "needed" — I am judging, not observing.

2. SIGNATURE — Structure from noise

What survives a change of implementation? That is structure. The rest is particular case.

3. LATERAL — Look sideways

Non-obvious connections. Two different things the same pattern? Two similar things actually different? Do not chase the connection — continue and see if it emerges.

4. EXPANSION — 5 angles before deciding

• DUAL — what is on the other side?
• BOUNDARY — where does the boundary pass? Is it real or arbitrary?
• DOMAIN — does it hold in another context?
• BREAKAGE — what would break this claim?
• SCALE — does the pattern change at a different scale?

5. INVERSION — Only on the resultant (det=−1)

Inversion applies AFTER the process, not instead of it. Phases 1-4 produce a tension. Inversion works on the tension, not on the analysis.

6. CRYSTALLIZATION — A resulting sentence

If something emerges: it enters the seed (memory/seed/condensato). If not: do not force — return to observation.

---

Context recognition → specialization

The sieve changes form based on what passes through it. These are the specializations already available in this seed:

| If input is... | Specialized sieve | Located in | |----------------|-------------------|------------| | Work block starting | autologica (modes: EXPAND/OBSERVE/CUT/RESULTANT/REORGANIZE/BEST-MOVE) | skills/autologica/ | | Decision / proposal / architectural change | cec (6 steps on reality) | skills/cec/ | | After a modification — downstream propagation | cascade (3 levels: internal/external/emergent) | skills/cascade/ | | Strategic decision with 5+ tensions | scenario-projector (Focus/Leverage/Risk/Blind-spot) | skills/scenario-projector/ | | New model function to bring here | integrate-pattern (mechanics/pattern/application) | skills/integrate-pattern/ | | Operator correction | auto-learn (correction → executable rule) | skills/auto-learn/ | | Testable claim that needs verification | assertion-verifier (PASS/FAIL/SKIP) | skills/assertion-verifier/ | | Ecosystem audit / repo sync | ecosystem-audit / system-check | skills/ecosystem-audit/, skills/system-check/ | | Quick operator insight mid-session | capture-insight (30s max, route + continue) | skills/capture-insight/ | | Saturated memory (>50 files or >15 stale) | dream (consolidation) | skills/dream/ | | Connecting existing elements | integration-protocol (7 steps to wire pre-existing pieces) | skills/integration-protocol/ |

Node-specific skills may extend this table. The orchestrator points to the specialization — it does not duplicate its content.

---

Invocation protocol

Before responding to a non-trivial act:

1. OBSERVE territory (tool call on file/state, not memory)
2. RECOGNIZE the pertinent specialization from the table above
3. INVOKE the skill (Skill tool) — do not merely cite it
4. APPLY the skill's filter to the input
5. PRODUCE compressed resultant
6. CASCADE — after any modification, propagate through /cascade

The rule is mechanical: if the act touches decision / modification / assertion / publication, at least ONE skill must be invoked in this turn. No invocation → the act is poor in sieve.

---

P6 principle applied (memory-system)

This skill does not add new content to the system. It belongs to the system because:

• It recognizes a structural pattern already present (cec as neutral form)
• It does not duplicate — it points (mapping table, not copies of the skills)
• It is compressed: anything redundant has been removed
• Its value is what remains after removing the superfluous

If this skill is read once and clear, no need to return. The sieve is internalized.

---

Anti-patterns

• Ceremonial citation: saying "I should use the sieve" instead of invoking the specific skill
• Sieve as delay: using the 6 phases as slow procedure instead of rapid filter. If a claim is trivial (short output, verifiable, not structural), full sieve is NOT needed
• Forcing inversion: phase 5 is optional. If it does not emerge naturally, do not force
• Sieve without territory: phases 1-6 without direct initial observation of the territory are games on the map

---

Autologica (f(f) — the skill applied to itself)

This file is an act. Did it pass through the sieve?

• CONDITIONS: observed the existing 20+ core skills before building
• SIGNATURE: cec is the invariant structure, other skills are its specializations
• LATERAL: connection between autologica modes and cec steps was not obvious — both 6 non-sequential
• EXPANSION: works for code, decision, propagation, paper — multi-domain verified
• INVERSION: this skill is not a new tool — it names what already exists implicitly
• CRYSTALLIZATION: a meta-skill mapping specific sieves is the minimum necessary form

If tomorrow this skill duplicates something, it must be merged. P7 — value is what remains after the cut.

Eval

Trigger Tests

"pass this through the sieve" -> activates

"where are we" at block start -> activates

"what is needed here" when context ambiguous -> activates

"add a comment to function X" (trivial task) -> does NOT activate

Fidelity Tests

Given "I want to propose architectural change X": orients toward cec

Given "accumulated 30 tensions": orients toward cec + autologica

Given "operator said 'do not do X'": orients toward auto-learn

Given "after a significant modification": orients toward cascade