deepakdsoni/antahkarana-base

Antaḥkaraṇa-base

A backbone-agnostic continual-learning "mind" — not a frozen checkpoint, a runtime that lets a model keep learning without forgetting, notice the novel, know what it doesn't know, and consolidate in sleep.

Antaḥkaraṇa-base is the foundational ("base") size of the Antaḥkaraṇa line. It is a four-organ continual-learning core (modelled on the Vedic antaḥkaraṇa: manas / buddhi / ahaṃkāra / citta) that wraps any backbone via a thin adapter and adds: anti-forgetting consolidation, dark-knowledge replay, calibrated abstention, energy-based novelty/OOD detection, adaptive plasticity, and sleep consolidation. It is validated, with seeds and adaptive evaluation, on three different modalities with one unchanged runtime.

• Author: Deepak Soni
• License: Apache-2.0 (language flagship freezes mistralai/Mistral-7B-v0.1, Apache-2.0)
• Base of the language flagship: frozen Mistral-7B (7.25 B params) + 13.6 M trainable adapter (0.19 %)
• Status: research; all six validation gates (G0–G5) green
• Install: pip install antahkarana · PyPI: https://pypi.org/project/antahkarana/ — train it on your data in ~10 lines (a thin adapter + your data stream)

---

What it is (and isn't)

• It is a continual-learning base / middleware: the organs + a generic training/inference loop. A new domain plugs in by implementing one BackboneAdapter (5 methods); the mind is identical for all.
• It is not a single set of weights. The "model" is the architecture + the small trainable adapters per domain. Think "continual-learning OS you drop onto a backbone," not a pretrained snapshot.

Use it for ANY domain

The core is backbone- and modality-agnostic. It already runs unchanged on language (a 7B LLM, generative), vision (a ViT classifier), and security (tabular intrusion detection) — three completely different modalities. Any new domain — coding, time-series, audio, tabular, multi-modal, robotics, … — plugs in the same way: implement one BackboneAdapter over your encoder, hand the core a stream of tasks, and you get continual learning, novelty/OOD detection, calibrated abstention, and sleep consolidation for free. The four organs don't change; only the thin adapter does.

Architecture — the four organs (the Vedic abstraction is the API; SOTA algorithms are the engine)

Organ	Role	Faculties / algorithms
manas (intake)	novelty / OOD	avidyā — max-softmax or energy-OOD (Liu et al. 2020)
buddhi (decision)	calibrate & adapt	pramāṇa abstention (risk–coverage) · viveka selection · guṇa adaptive plasticity (rajas/sattva/tamas × domain base-lr)
ahaṃkāra (identity)	self / boundary	label-free task-boundary detection
citta (memory)	retain/consolidate	saṃskāra (Fisher consolidation) · vijñāna-smṛti replay (naive or DER++, Buzzega et al. 2020) · nidrā sleep

Validated results (seeded; from the gated build G0–G5)

Modality	Backbone	Core result
Language	frozen Mistral-7B + LoRA (generative)	reproduces the validated v3.0 system; DER++ replay drives forgetting → ~0
Vision	frozen ViT-B/16 + adapter (Split-CIFAR-100)	acc 0.925 ± .003, faculties vs naive: ~+9 pts, ~6× less forgetting (3-seed)
Security	tabular IDS encoder (NSL-KDD)	forgetting 0.069 ± .021 vs naive 0.152; zero-day AUROC 0.949 ± .005 (3-seed); calibrated triage

Adaptive / honest security eval: zero-day AUROC for every held-out attack family = dos 0.79 / probe 0.90 / r2l 0.78 / u2r 0.96 (no cherry-pick); an evasion curve (blend attacks toward normal) degrades 0.95 → 0.25 — the detector is a strong novelty signal, not an unbeatable defense.

Validation gates (G0–G5) — how we know it works, and that it's not cherry-picked

The whole thing was built as a gated pipeline: each gate had to go green before the next, so nothing rests on faith and a refactor can't silently break the science.

Gate	What it proves
G0 — golden lock	freeze the known-good baselines; every later gate diffs against them (regression safety net)
G1 — faculties (20/20)	each organ/algorithm is correct in isolation
G2 — integration (12/12)	the four organs compose into one working loop on a trivial backbone
G3-language	the core reproduces the validated 7B continual-learning system
G3.5	a SOTA organ (DER++) measurably improves the real 7B (forgetting → ~0)
G3-vision	the same core runs a ViT classifier — cross-modal proof
G4 — security	the same core runs tabular intrusion detection (3rd modality): no-forgetting + zero-day + triage
G5 — honesty	seeds + std, adaptive evasion eval, per-organ ablations, archived negatives, stated limits

All six green. Because G3-vision and G4-security run the identical core that passed G3-language, "works on any domain" is demonstrated, not asserted — and G4 even surfaced and fixed a general core bug that only a third modality could expose. See GATES_STATUS.md and G5_HONESTY.md in this repo.

How to use

from antahkarana.core import Antahkarana
from antahkarana.domains.language import LanguageBackbone, build_language_stream

bb = LanguageBackbone()                    # frozen Mistral-7B + LoRA
stream = build_language_stream()           # a sequence of tasks/skills
mind = Antahkarana(bb, samskara=True, replay=True, sleep=True,
                   replay_strategy="der",  # DER++ dark-knowledge replay (SOTA)
                   avidya_strategy="energy")
result = mind.run(stream)                  # learns continually; returns matrix, forgetting, risk_coverage, ...

A new domain = one BackboneAdapter subclass implementing named_trainable / train_step / eval_task / predict_with_confidence / confidence_pairs (+ optional logits for energy-OOD, capture/der_loss for DER++). See domains/{language,vision,security,mock}.py.

Per-organ attribution (ablations)

No single organ wins everywhere — it's a toolkit:

• Vision / class-incremental: saṃskāra dominates.
• Security: replay is the main anti-forgetting lever (saṃskāra ~neutral; sleep small/occasionally −).
• Language (7B): saṃskāra → +vijñāna-smṛti → DER++ each lower forgetting.

Limitations (honest)

1. Effect sizes are domain-dependent (large in vision/security, modest in 7B-generative).
2. Generic, not specialized — lands near tuned domain SOTA (e.g. vision 0.925 vs a specialized 0.947), not always at it.
3. Per-domain hyperparameters (base-lr, epochs, memory fraction, guṇa bands) are set, not auto-tuned.
4. The zero-day detector is evadable under full adaptive disguise.
5. Scale: 3–4 seeds; NSL-KDD is a dated IDS benchmark; in-RAM memory, single-GPU loop; research-grade, no formal guarantees.

What didn't work (archived, not hidden)

• Co-training at 7B (shared frozen backbone → correlated views → no gain).
• Self-directed curriculum ("tapas") in the fully-supervised regime (no headroom).
• Order-biased replay memory (train[:frac]) — a real bug the security modality exposed and fixed.

Citation

@software{soni_antahkarana_base_2026,
  author = {Deepak Soni},
  title  = {Antaḥkaraṇa-base: a backbone-agnostic continual-learning mind (manas/buddhi/ahaṃkāra/citta)},
  year   = {2026},
  note   = {Validated across language, vision, and security; gates G0–G5.}
}

The antaḥkaraṇa framing is a design philosophy and explanatory map — every organ resolves to a real, citable algorithm. Not a claim of supernatural capability.