NK-Ops is an experimental research repository studying how meaning-bearing operators behave under controlled stress conditions. Instead of embeddings or token similarity, it measures operator-level divergence at the verse (ayah) level and aggregates it into author-level response profiles across Qur'an translations.
This repository documents a candidate phenomenon (see reproducibility scope below — this is an observation from our deterministic engine, pending independent validation, not a settled result):
Under increasing semantic stress, translators appear to separate into a small number of stable response profiles rather than drifting randomly.
For each author (Qur'an translation):
- Ayah-level operator response (
SOFT,SILENCE) - Binary divergence signal (
divergent ∈ {0,1}) - Stress parameter ( s ∈ {0.2, 0.5, 0.8} )
- Aggregated divergence rate per stress level
The pipeline is deterministic (no stochastic components): given a fixed engine version and fixed inputs, it returns identical outputs.
- Ayah-level stress experiments — per author and stress level →
by_author/<author>/ayet_results.tsv - Author profile construction — aggregate divergence rates across stress levels →
author-profile.tsv - Stress-response grouping — authors grouped by response-curve shape (not embeddings) →
author-profile-clustered.tsv
- Low stress (s = 0.2): system is globally stable; divergence rates are minimal.
- Mid stress (s = 0.5): early differentiation appears; author behaviors begin to separate.
- High stress (s = 0.8): author response curves separate into a small number (≈3) of groups; the separation tracks response shape rather than lexical choice.
Caveat. These groups come from clustering author response curves, and clustering can impose structure that is not intrinsic. With 46 authors the exact number of groups is not robustly established, and "separation under stress" is a descriptive observation, not a statistically tested phase transition. The whole result also depends on the (withheld) engine's definition of "stress" and "divergence", which has not yet been externally validated. Treat it as a candidate phenomenon pending independent replication.
- Explores an alternative to token/embedding-centric analysis, in which operator activity is the measured signal.
- If the grouping holds up under independent validation, it would offer operator-level, meaning-oriented alignment/interpretability metrics.
We use field/force vocabulary ("meaning under stress", "operators acting on meaning") as framing, not as a physical claim. NK-Ops is one subsystem of a broader NK research program.
To be explicit about reproducibility:
- Included: the operator taxonomy, methodology and documentation, and result artifacts (profiles, tables, figures).
- Not included: the core operator/MSV engine (research IP), and the large source corpus.
As a consequence, the pipeline is not runnable end-to-end from this repository alone. The documented method is deterministic and can be independently re-implemented, but bit-exact reproduction from this repo by itself is not possible. Claims of "reproducibility" in this repository should be read in that light.
This repository documents Phase-2: Author Stress Response Analysis (with Phase-1 operator/MSV framework and Phase-3 decision-gate material). Findings are provisional and pending independent validation. Further phases will extend the operator taxonomy and test cross-text generalization.
The directory operators/ now contains the full public release of the NK-Ops
Turkish operator inventory:
- 206 operators (
operators/TR/ops/**/op.json, imported affixes underoperators/EXP/): operator identity, class, surface forms (allomorphs), POS in/out constraints, phonotactics, morphotactic notes, and semantic policy notes. 179 operators are active; deprecated ones carry their tombstone status in the index. - Taxonomy —
operators/TR/canonical_taxonomy_v3.json(ontological top classes, alias & deprecation policy). - Index —
operators/TR/ops_index.json(op → path, surface forms, status).
What is deliberately excluded: the NK field equations and per-operator
field parameters (field_* fields of the internal schema) are research IP and
are not part of this release. The inventory is the linguistic layer:
identity, form, combinatorics, and policy.
On validation: NK-Ops ships an internal validator, but measurement showed
its discriminative power on multi-candidate analyses is ~0.9% — it ranks
candidates and annotates consistency; it does not reject. Treat MEŞRU
(legitimate) as "morphotactically consistent", not "correct in context".
Context-level selection is a separate, unreleased layer.
Citation: if you use this inventory, please cite the A1 paper — Morphemes as Measurable Operators: An Interpretable Composition Algebra over a Graph-Derived Semantic Core for Turkish (2026), DOI: 10.5281/zenodo.21939027.
License: repository code/documentation under MIT (LICENSE); the operator
inventory data under CC BY 4.0 (operators/LICENSE-DATA).