GrandNexus v7.2.3-rc
An anchored translator and a reference engine that decide without inventing, and a laboratory to measure whether a loop learns.
Claude translates a text and a question into a small logic language (anchored to the text: every statement quotes
the span it comes from); the reference engine nexus_logic decides (yes / no / unknown / contradictory) with a printable
proof, checked against clingo; the answer is the strict verdict (text only), the background-knowledge verdict is reported apart.
Latest experiments are documented in LOGBOOK.md; generated benchmark outputs are intentionally excluded from version control.
Everything else in this repository (the 2025 cognitive-architecture scaffolding: memories, graph, orchestrator, 15/05 layer)
is an experimental ground around that core, kept and runnable but not demonstrated.
Keywords: cognitive architecture, neuro-symbolic AI, symbolic reasoning, agent memory, continual learning, knowledge representation, explainable AI, grounded LLM translation, proof-carrying decisions, AI safety, and open-source AI research.
This archive is a complete, self-contained snapshot: no notebook chain, no %%writefile, no update to graft.
Quick start (Colab or any Python 3.10+)
unzip GrandNexus_v7.2.3-rc_complete.zip -d /content # -> /content/GrandNexus
pip install -q -r /content/GrandNexus/requirements.txt
import sys; sys.path.insert(0, "/content/GrandNexus")
from grandnexus.boot import boot
gn = boot(llm="auto") # nx pathway (Claude -> nexus_logic) if ANTHROPIC_API_KEY is set, else grammar only
gn.print_report()
tr = gn.modules()["text_reasoner"]
print(tr.answer("All men are mortal. Socrates is a man. Is Socrates mortal?")["answer"])
gn.shutdown()
Modes: laboratory (default, deterministic, nothing runs in the background) or live (boot(start_core=True) / gn.start()).
What is where
| Path | Content |
|---|---|
grandnexus/ |
the package (core, memories, cognition, graph, LLM interface, 15/05 layer) |
grandnexus/boot.py |
assembles and wires the whole system |
grandnexus/nexus_logic/ |
the core: language spec, parser, reference engine (disjunctions, formulas, reductio), clingo export, bridge |
grandnexus/llm_interface/translator_v5.py |
Claude as anchored translator into the nexus_logic language (prompt v5.4), Claude-alone control |
grandnexus/cognition/reasoning/text_reasoner.py |
text -> program/premises -> reference engine -> verdict (grammar, JSON and nx pathways) |
grandnexus/cognition/deliberation.py |
the deliberation experiment (several readings, targeted re-reading; hypothesis rejected, logbook 44) |
grandnexus/llm_interface/llm_translator.py |
the earlier JSON translator (prompt versions v1-v4), kept for comparison |
tools/bench_folio.py, tools/bench_deliberation.py, tools/diff_engines.py |
FOLIO bench (content-sha freezing, trivial controls, --strict), deliberation bench, differential reference / clingo / home engine |
tools/test_bench.py |
bench for hand-written sets: freezes a set, measures 4 modes, never prints item texts |
tools/smoke_test.py, tools/lab_*.py |
import test, labs (controls, layer oracles) |
tests_lab/test_*.py |
offline tests (no API) |
LOGBOOK.md |
logbook: every hypothesis -> result -> decision (49 entries) |
tests_lab/resultats/ |
generated FOLIO outputs (kept out of version control) |
archive/ |
original 15/05 and 29/05 sources (untouched) |
.git |
full history: every repair is one commit (git log --oneline) |
Checks
cd /content/GrandNexus
python tools/smoke_test.py
for t in tests_lab/test_*.py; do python $t; done
python tools/diff_engines.py tests_lab/programmes_nx # reference engine vs clingo (pip install clingo)
python tools/bench_folio.py --split train --seed 2 --n 100 --modes translator,direct --strict # needs ANTHROPIC_API_KEY
Language
Code, tools, tests, logbook and README are in English. The legacy 2025 modules (outside the core) still carry French comments and log messages; they will be translated in a later pass.