Pratyabhijñā Creative Engine (PCE)
A portable plugin (Cursor + Claude Code + standalone CLI) that operationalises Abhinavagupta's Pratyabhijñā five-śakti generative cascade as typed operators over an active-inference / Bayesian Model Reduction substrate, with a recursive vimarśa self-reflexivity layer.
v0.4 — mechanism study of recursive self-reflexivity layers for LLM creative cognition. Cascade-vs-bare null at the pilot's n; recursive revision pass robustly positive on its own (H8a, g = 0.65, p < 1e-4); learned commit gate beats the v0.3 event gate (H8b, F1 0.65 vs 0.52); proxy scorer disagrees with the Sonnet-4.5 LLM-judge at ρ = 0.0 (H9 — flagged as a metric-design problem, not a refutation). Cost ledger split honestly: Haiku cascade $12.73 / 1,277 calls; Sonnet judge $0.48 / 23 rows; combined pilot spend $13.21. v0.4 frozen at paper/v0.4/.
Live site: https://sharathsphd.github.io/pratyabhijna/ — Astro Pages site that reads benchmarks/results_v0.4/stats.json directly and animates all 9 showcase demos with full cascade traces.
Companion work. Pratyākṣa (direct perception / context-discipline) at https://zenodo.org/records/19680692 and https://sharathsphd.github.io/context-engineering-harness/. PCE is the recognition + creativity counterpart in the same author program.
This repository holds the engine, the dual plugin manifests (plugin/.claude-plugin/plugin.json, plugin/.cursor-plugin/plugin.json), the standalone pce CLI, the Phase 7 mechanism pilot's results and audit, the v0.4 paper, and the Astro v0.4 site.
- Frozen scope: v0.4
docs/SPEC_v0.4.md, docs/PRD_v0.4.md, docs/RELEASE_NOTES_v0.4.md, docs/COMPLETION_PROMISES_v0.4.md.
- ADRs: v0.4
docs/adr/v0.4/ (ADR-001 cit_temperature substrate, ADR-002 learned commit gate, ADR-003 free-energy budget, ADR-004 managed-API fairness lattice, ADR-005 fixed-effects H5, ADR-006 typed Haiku errors, ADR-007 SDK code-path removal).
- TRIZ cards: v0.4
docs/triz/v0.4/.
- Plugin manifests:
plugin/.claude-plugin/plugin.json and plugin/.cursor-plugin/plugin.json (both at v0.4.0).
- Hypotheses: H1.v4–H4.v4 (per-domain), H5.v4 (fixed-effects pool), H6.v4 / H7.v4 (fairness controls), H8a/b/c.v4 (mechanism decomposition), H9.v4 (judge agreement).
Three install paths
# 1. Standalone CLI (works in any shell with `claude` on PATH)
git clone https://github.com/SharathSPhD/pratyabhijna.git
cd pratyabhijna
uv pip install -e . # required: registers the `pce` console script
# and pulls numpy / sentence-transformers
pce smoke # verifies the CLI, OAuth substrate, and cascade module
pce cascade --prompt "Write a haiku about rain on a tin roof" \
--constraint "imagism" --k 4 --seed 4242
# 2. Cursor plugin
cursor --install-plugin .
# 3. Claude Code plugin
claude plugin install https://github.com/SharathSPhD/pratyabhijna
# or, for a local clone:
ln -s "$(pwd)" "$HOME/.claude/plugins/pce"
See docs/RUN_LOCAL.md for the full operator guide, including the precedence chain (defaults → repo pce.toml → user ~/.config/pce/config.toml → env vars → CLI overrides) and example configurations.
Reproduce v0.4 numbers
# Regenerate the figure pack and autoreport against the published stats.json
python -m benchmarks.figures --version v0.4
python -m benchmarks.autoreport --version v0.4 --strict
# Build the paper PDF (tectonic preferred; latexmk also supported)
cd paper && tectonic -X compile main.tex
# Build and serve the Astro site locally
python scripts/prepare_site_data.py
cd docs/site && pnpm install && pnpm build && pnpm preview
The v0.4 pilot ran the same four base arms as v0.3 plus a five-policy commit multiplexer over the cascade arm. The arm matrix is:
| arm | role |
|---|
haiku_bare | architecture-vs-nothing primary control |
haiku_cascade | primary treatment (always-shadow revision; commit per policy) |
haiku_bare_2K_scorer | +K-compute control (H6.v4) |
haiku_generic_revise_2pass | revision-protocol control (H7.v4) |
The commit-policy multiplex over haiku_cascade is always_draft / always_revise / event_gated (the v0.3 policy) / learned_gate (ADR-002) / oracle (analysis upper bound).
Pre-registered hypotheses (v0.4)
SharathSPhD
