prabodha-lenses

prabodha-lenses — fitted workspace-band Jacobian lenses

Pre-fitted band-targeted Jacobian lenses for the prabodha library (recognition-gated workspace steering for language models). These let you skip the multi-hour lens-fitting step and go straight to reading and steering the functional "global workspace" band of six open decoder LLMs — the same lenses the live prabodha app uses for real Jacobian steering.

Provenance. The lens instrument itself is Anthropic's Jacobian-lens (Apache-2.0, from Verbalizable Representations Form a Global Workspace in Language Models, transformer-circuits 2026), vendored unmodified in the prabodha repo. These .pt files are lenses fitted with that instrument on the models below. Utility only — understanding and steering of LLMs. Not a claim about consciousness.

Files

lenses_manifest.json carries the machine-readable table (model → lens file → site layer); configs/ carries the matching model + lens-fit YAMLs.

The six top rows are the app-registered lenses. Each was fitted with the same pre-registered pipeline: n_prompts=16, seq_len=128, the wikitext-2-raw-v1 pretraining-like corpus (64×128 word windows, seed 42), target_layer = n_layers − 1 so the fit covers every band site up to the model's depth. The steering site layer is int(0.62 · n_layers) — the workspace band, not the final layer. A lens targeted at the model's final layer is structurally blind to intermediate workspace content — the single most load-bearing instrument finding of the program (gate gate_L2b.json).

Quickstart

pip install prabodha
# grab a lens (e.g. Gemma-2-2B)
hf download qbz506/prabodha-lenses gemma-2-2b-it.pt --local-dir ./lenses
# read the workspace band on a prompt
prabodha lens-vis --model google/gemma-2-2b-it \
  --lens-file ./lenses/gemma-2-2b-it.pt --site-layer 16 \
  --prompt "the fire remembers rivers" --out fire_slice.html

Or steer live in the prabodha app — every model above runs its real fitted lens (direction_source: concept:X), not a contrastive fallback.

Calibration note (important)

Lens-write amplitude is per-model. On most models here a visible steer needs alpha ≈ 1–3; Qwen3-4B is more sensitive (alpha ≈ 0.3). alpha scales inversely with the target's own lens transport strength — calibrate to the plant, don't reuse a fixed value (gate_L13_recipe.json, gate_L14_multiseed.json).

What these are good for (gate-cited)

• Reading the workspace band — the band is legible only to a lens targeted at its own exit (gate_L2b.json).
• Recognition-gated steering — event-gated ("sphuraṭṭā", uncommitted-moment) writes steer within an entropy budget; the core confirmed result at 6 seeds (gate_L9_alignconf.json, gate_L11_rep.json). Requires sampling decoding — greedy/argmax mechanically masks decode-time writes (gate_L4b.json).
• Cross-model transfer via calibration — scale write amplitude to the target plant's own lens transport strength (gate_L13_recipe.json, gate_L14_multiseed.json).

Honest negatives are shipped results too: the āgama-readback acceptance test is a weak, non-oracle predictor (balanced accuracy ~0.59 at n=120) — never a sole acceptance gate. See the paper and the gate files for full detail.

Vocabulary

Internal terms are Sanskrit-forward with engineering glosses: sphuraṭṭā (the pre-linguistic "flash" → the entropy-gated write-timing event), vimarśa (reflexive self-awareness → the verbalizable read-out), svātantrya (autonomy → the entropy budget on writes), āgama re-cognition (→ the read-back acceptance check).

Licensed Apache-2.0. Cite the prabodha repository and Anthropic's Jacobian-lens.