Impossible Papers
Multi-agent research engine that takes an impossible hypothesis, stress-tests it against real academic literature, and produces papers that read like hard science.
The Problem
There are ideas that sit at the edge of science — hypotheses too speculative for a journal submission but too structured for a blog post. Consciousness as a somatic field. Luminiferous ether revival. FTL propulsion from first principles. Will-driven motion.
The trick is that these ideas can be explored with full scientific rigor even when the premise is false. A good impossible paper commits to internal consistency: it derives consequences, cites a real intellectual lineage, shows its math, and states what would falsify it. The reader's pleasure comes from watching a false premise handled with completely straight-faced discipline.
Writing one by hand takes weeks of literature review, formal derivation, and careful self-editing. Impossible Papers automates the scaffolding so the human can focus on the idea.
The Refinement Loop
Impossible Papers is not a paper generator. It's a hypothesis refinement engine. A raw idea goes in fuzzy and comes out battle-tested against real literature.
Step 1 — Formalize: An LLM takes your shower thought, conversation transcript, or one-line hypothesis and decomposes it into a structured thesis with typed, dependency-linked claims. Foundational axioms first, derived claims next, speculative leaps last.
Step 2 — Literature: Semantic Scholar and OpenAlex are queried for real academic papers. Each source is mapped against each claim — supporting, challenging, or irrelevant. The agent identifies structural gaps: where does the argument break without a source? Only then are synthetic references generated to fill those gaps.
Step 3 — Reassess: The hypothesis is stress-tested. Which claims have real support? Which are hanging by invented bridges? Did the literature reveal a better version of the idea? The system returns a verdict — solid, revise, pivot, or abandon — and routes back to whichever step needs work.
Literature That Earns Its Place
The literature agent doesn't decorate — it diagnoses. Real papers from Semantic Scholar and OpenAlex are searched, mapped against each claim, and assessed for strength. The agent builds an argument map: which sources support which claims, which challenge them, and where the lineage breaks.
Invented references are structural, not cosmetic. They exist only
because gap analysis proved the argument can't stand without them.
A synthetic "Moretti 2008" paper fills a specific hole — the missing
link between IIT's information metric and field theory — published
in a real journal, with a plausible abstract and realistic authors.
Internally flagged real: false. In the published paper,
seamless.
Checkpoint System
# Every step writes a JSON checkpoint.
# The pipeline can restart from any point.
hypothesis.json ← structured thesis + claims
literature.json ← real sources, argument map, gaps, synthetics
reassessment.json ← verdict, strength ratings, revisions
# Resume from any checkpoint with a different model:
impossible-papers run --from runs/run_006/hypothesis.json \
--step literature
Model-Agnostic Per Step
Every agent's LLM is configured independently via YAML. The formalization step might run on Qwen 7B locally through Ollama. The reassessment step might use Claude Sonnet for stronger reasoning. The literature mapping might use a 70B model on a HuggingFace endpoint spun up on demand.
Swap providers without touching code. Add a new provider by
implementing one method: complete(system, user, temperature).
The config router handles the rest.
Human in the Loop
The pipeline pauses after every step. Each checkpoint is a deliverable for human review. Is the thesis framing right? Are these the claims you want to make? Are the real sources the right ones? Are the gaps correctly identified?
Any step can be a restart point. Revise the seed, re-run formalization with a different model, expand the literature search with new queries. The system remembers its state; the human steers the direction.
Built on the Caria Stack
Impossible Papers lives in the caria monorepo. Its academic search layer — Semantic Scholar, OpenAlex — is the same infrastructure planned for Bookshelf's discovery features. Its embedding pipeline will share LaBSE with Monsieur.lib's esoteric catalog of 161,000 items.
For hypotheses that touch Western esotericism, consciousness, or occult-adjacent domains, Monsieur.lib provides a unique source catalog that no academic API covers. The literature agent can pull from both peer-reviewed science and the hermetic tradition.
The First Hypothesis
Level 1 consciousness is somatic, not cortical. The body receives the full unfiltered data stream — all sensory, proprioceptive, interoceptive input. The brain runs in parallel on a filtered subset. What we call "gut feeling" is the soma having integrated information the brain hasn't received yet. AI is structurally incapable of L1 consciousness because it only has the brain-analog.
Singer & Damasio (2025) argue that interoception constitutes the foundational substrate of consciousness, grounded in core biology rather than high-level cognition. The real literature supports the hypothesis. Impossible Papers found it.
Roadmap
V1 (current): Hypothesis refinement loop — formalize, literature review, reassess. Checkpoint-restart. YAML model routing. CLI interface.
V2: Formalization agent (equations, toy models), visualization agent (figures, phase plots), editor agent (consistency audit), falsifiability sub-loop. Flask renderer with MathJax.
V3: SIA-style meta-learning across runs. Full-text RAG over key papers. Monsieur.lib integration for esoteric sources. Web UI for checkpoint review.