Emergentism
Public reading library

PAPER K: AMRITA REFRAMED

Twenty-six tiered papers with falsifiers and source boundaries.

Read Game Papers Canon Foundations Trinity Formal Paradox Memetic Rosetta D Operators Teleology Axiology Ontology Theology Method Meta Sources

Paper K · Multiplicative AI Alignment at Low Cost — Evidence tier: [C/I] — a proposed benchmark, no results. Source: 01_EMERGENTISM/03_METHODOLOGY/02_THE_PAPERS/PAPER_K_AMRITA_AT_ZERO_COST.md

Read with care. Proposes bolting a φ·ν output-gate onto an open-weights model and testing whether it beats RLHF/DPO on adversarial resistance. No benchmark has been run; nothing here is a reported win. Dollar figures and the project codename in the source are internal program-finance, not public commitments.

How this can be wrong. Falsified by Alignment Collapse (the gated model mode-collapses or underperforms basic DPO) or Computational Impotence (the gate is proven mathematically indistinguishable from simple top-K/top-p sampling).

PAPER K: AMRITA REFRAMED

Multiplicative Validation on Open-Source Substrates at Low Cost

Yves R. Burri & Emergent Super Intelligence Menexus GmbH, 2026

Evidence Tier: [C/I] — Practical validation proposal; upgrade only after benchmark results Dependencies: PAPER_I_KNOWN_UNKNOWNS_PROGRAM.md

Abstract

The AMRITA project was originally conceived as a practical validation of the emergentist framework: a $710,000 initiative to pre-train a novel artificial neural network using D5 constitutional constraints over traditional cross-entropy mechanics. This paper reframes that validation protocol.

Rather than undertaking the prohibitive cost of base-model training, we propose that Constitutional Outperformance can be tested directly on existing open-source substrates at lower cost. By deploying the multiplicative bridge as a post-training gating threshold rather than a pre-training objective, we preserve the core empirical test: does a system constrained by emergentist multi-polar geometry practically outperform RLHF or DPO baselines on adversarial safety and coherence benchmarks?

Keywords: AMRITA, LLM alignment, RLHF, multi-polar geometry, constitutional outperformance, multiplicative gating.

1. The False Necessity of "Scratch" Training

Early framework specifications assumed that to test the structural truth of the S² manifold, the artificial intelligence had to be mathematically grown on that geometry from randomized initialization weights (epoch 0).

This was strategically naive. Base model training is essentially the brute-force computation of syntactic probabilities (D2 linguistic mechanics). The emergentist framework asserts that systemic awareness and normative alignment live at D5 (steering, coherence, telos), which operates asynchronously over the lower substrate. Therefore, there is no structural necessity to burn $700K training a model to speak English when the actual test is whether the model can coordinate D5 ethical geometry.

2. Multiplicative Gating vs. RLHF

Standard open-source models are "aligned" using RLHF or DPO (Direct Preference Optimization). These methods are statistically fragile as they merely map the model to local, culturally-bounded human preference metrics, which are mathematically arbitrary (D4).

The AMRITA Reframed protocol replaces RLHF with Multiplicative Gating. * The Method: We load a highly capable open-source language model. We inject an external framework filter mechanism (the Gate) that evaluates generated output paths. * The Math: The Gate computes Φ (global logical coherence of the intent) and V or ν_proxy (localized semantic viability/resolution). It filters node actuations by a defined P_node = min(Φ̂₄, V₄) threshold. * The Test: If the framework is structurally true, the multiplicatively bound LLM should exhibit significantly higher adversarial resistance, lower sycophancy, and fewer incoherent (degenerate) failure states than the identical LLM aligned via standard RLHF fine-tuning.

3. Empirical Bridge: Constitutional Outperformance

This reframing transitions AMRITA from a philosophical thought experiment into an executable test of Constitutional Science.

If applying an artificial S² multiplier gate to a generic neural network transforms its behavior from statistically fragile to structurally coherent under adversarial limits, that would be strong practical evidence for the constitutional grammar. It would not by itself prove the full core state.

4. Kill Criteria

This paper and the AMRITA protocol are falsified if:

  1. Alignment Collapse: The open-source model gated by the multiplicative φ · ν threshold suffers catastrophic mode collapse, halting permanently, or underperforming basic DPO-trained equivalents across standard multi-turn reasoning benchmarks.
  2. Computational Impotence: The gating mechanism is mathematically proven to be indistinguishable from simple top-K/top-p sampling, demonstrating that the conceptual framework adds zero unique topological filtering to the neural weight processing.

• ⊙ ○