Skip to content

reference-orchestration.md

Latest commit

 

History

History
134 lines (88 loc) · 5.51 KB

reference-orchestration.md

File metadata and controls

134 lines (88 loc) · 5.51 KB
Raw

Reference — Orchestration

OmegaClaw exposes two reasoning engines (NAL and PLN) plus direct memory recall. The LLM decides which to use, when to stop, and whether to act on a result. This page catalogues those decision policies. (ONA is a planned third engine — see reference-lib-ona.md for its experimental, not-installed status.)

1. Engine selection (pattern → engine)

The LLM uses heuristic triage: recognize the reasoning shape, then pick the engine.

Situation Pattern Engine
Factual recall, no chains Direct retrieval query / episodes
Known chain A → B → C Deduction NAL |-
Observed effect, seeking cause Abduction NAL |-
Multiple instances → generalization Induction + Revision NAL |-
Property-based categorical inference Modus Ponens PLN |~
Independent evidence to merge Revision NAL or PLN
Real-time temporal sequences Temporal inference (no stock engine — ONA is the planned target but not installed; fall back to NAL with external temporal grounding)
Evidence conflicts Revision NAL / PLN
Novel hypothesis generation Abduction / Induction NAL

Fallback rule: if an engine returns empty, reformulate premises (fix term order or missing middle). If still empty, switch engine.

2. Stopping criteria

Halt conditions the LLM is expected to monitor after each inference hop:

Signal Threshold Action
Confidence floor c < 0.3 Halt. Conclusion is unreliable.
Sufficiency threshold c ≥ 0.6 Actionable for practical decisions.
Diminishing returns New hop reduces confidence more than it adds information Halt.
Resource budget 5 commands per cycle Hard ceiling enforced by the loop.

The per-cycle command ceiling is set in the output-format directive built by getContext in src/loop.metta.

3. Action thresholds (three-tier gating)

Once a truth value is in hand, compare it against these tiers before taking an action:

Tier Gate Meaning
ACT f ≥ 0.6 AND c ≥ 0.5 Actionable — take the step.
HYPOTHESIZE f ≥ 0.3 AND c ≥ 0.2 Worth exploring — gather more evidence.
IGNORE below both Insufficient evidence — do nothing.

The NAL expectation helper is useful for ranking candidates:

exp = c × (f - 0.5) + 0.5

It maps (f, c) to a single scalar in [0, 1] suitable for a priority queue.

4. Conflict resolution

When two results disagree:

  1. Prefer the higher-confidence result if frequencies agree.
  2. If frequencies clash, invoke revision to merge the premises as independent sources.
  3. Respect engine domains — NAL for inheritance-style chains, PLN for property-based inference.

Revision produces a frequency that encodes the disagreement (drifts toward the middle) while confidence grows — the new truth value makes "substantial but conflicting evidence" explicit in the math.

5. The defense stack (four layers)

OmegaClaw assembles these layers to resist noisy or adversarial input:

Layer 1 — Novelty modulation

New claims are discounted by their novelty:

c_new = c × (1 - novelty)

Claims without precedent in long-term memory enter the reasoner at reduced confidence.

Layer 2 — Action thresholds

The ACT / HYPOTHESIZE / IGNORE gates above prevent low-evidence claims from reaching decisions.

Layer 3 — Attention budgeting

A priority queue ranked by NAL expectation (exp = c × (f - 0.5) + 0.5) with hard step limits per cycle. Spend attention on the most promising inferences first.

Layer 4 — Adversarial premise testing

Regression suite: confident lies, direct contradictions, and gradual poisoning are injected into the pipeline. The stack has been validated to downgrade adversarial inputs in all tested cases.

6. Multi-cycle reasoning pattern

Complex questions usually cannot fit in one cycle because the LLM must emit all commands before seeing results. A typical decomposition:

Cycle Purpose
1 Gather information — query memory, search web, read files, fetch external data.
2 Atomize the relevant knowledge; run the first NAL or PLN step.
3 Revision with independent evidence; follow-up inference.
4 Threshold check; decide whether to send an answer.

The LLM is expected to pin state between cycles so the next turn can continue from the same plan. See tutorial-08-reliable-reasoning.md.

7. When automatic orchestration fails

If the LLM picks the wrong engine or formulates bad premises (see reference-failure-modes.md for measured rates), the observable symptoms are:

  • Empty result from |- or |~.
  • Result with suspiciously high confidence after many hops.
  • Same question produces different truth values across runs (high variance).

Standard recovery:

  1. Reformulate premises (fix term order, copula, granularity).
  2. Switch engine if the pattern matches another.
  3. Inject external grounding — see tutorial-07-grounded-reasoning.md.

See also