Recursive Self-Improving Meta-Reasoning Architecture for Expert Work
Sanscritic is an AI Neolab building metareasoning systems and expert foundational models — AI whose reasoning is legible, addressable, and forkable — like formulas in a spreadsheet — made for work where the reasoning is the product.
Expert Examples - Biotech attorneys · Robotics safety & compliance engineers · CLO portfolio analysts · Medical writers · And more
LLMs are trained for one good answer. Expert work is nuanced
Where LLMs do really well
- Math, code, and the physical sciences.
- Reinforcement Learning with verifiable rewards on massive datasets.
- Always exactly one correct answer.
- A single, sequential chain of thought.
What expert work is
Experts only trust reasoning they can inspect, modify and branch. That is how hallucinations are caught and corrected.
Amplification, not automation.
The industry has built superbly for one paradigm - and lags in the other. Automation and amplification have different demands.
Automation
e.g. an agent extracting information from a PDF and entering it into a legal billing system
- Reasoning is a means to an end - consumed by the agent to perform the action accurately
- The output self-verifies — code compiles, tests pass.
- Opaque chain-of-thought is acceptable.
- Measured in tasks completed.
Amplification
e.g. a biotech attorney reasoning across molecular biology, patent law, licensing, FTO, FDA regulation
- Reasoning is the product - consumed by a human expert
- No test suite says the answer is right — the expert verifies by inspecting the reasoning.
- The trace is a first-class artifact.
- Measured in hypotheses examined per hour.
Forkable, like a spreadsheet.
The spreadsheet amplified accountants by making the reasoning — the formulas — visible, editable one cell at a time. Sanscritic is that layer for expert judgment: every trace is a sheet of cells you can inspect, edit, and fork.
Trace A — original
1
Frame
Patent claim scope
2
Evidence
Licensing terms, §4.2
3
Assume
Claim 3 is enforceable
4
Infer
Freedom-to-operate blocked in the EU
5
Conclude
Redesign required
Trace A′ — forked at step 3
1
Frame
Patent claim scope
2
Evidence
Licensing terms, §4.2
3
Assume
Claim 3 is invalid — prior art, 2019
4
Infer
Freedom-to-operate clear in the EU
5
Conclude
Proceed to file
Swap one assumption in one cell; everything downstream recomputes. The expert reads exactly what changed — and why.
01
Legible
See which frame was applied where — not a wall of prose.
02
Addressable
Point at step 7 and say: here is where it went wrong.
03
Forkable
Branch at a decision node, swap an assumption, rerun the what-if.
If you don't own your AI's reasoning, you rent it — and pay in expertise.
Reasoning is implicit
Baked into the weights by the lab's training choices. You can't see it, audit it, or change it.
Your only lever is fragile
Prompting — until a paraphrase or a silent model update erases your carefully engineered behavior.
Your knowledge leaks upstream
Every prompt, correction, and workaround feeds the frontier lab — training the model your competitors rent next.
Sanscritic inverts the ownership: experts curate their own reasoners. The control stays with them, and the knowledge never leaves.
Reasoning as a structured debate.
Sanskrit epistemology spent multiple millennia formalizing how rigorous, debate-based reasoning should proceed. It defines canonical lenses — each specifying probes (what to ask) and process (how the argument must move).
That's the last Sanskrit you'll hear from us. From here on: plain English.
130
Canonical lenses
The metareasoner detaches reasoning from the language model.
The metareasoner
01 Probes, not prompts
A neural network sits on top of a tuned LLM, probing it from multiple perspectives custom-built for the expert task.
02 Created by AI
The metareasoner itself is generated by AI — one per expert domain.
03 No data or rewards needed
Building the expert foundational model requires no datasets and no reward models — no RLVR, no GRPO.
Expert foundational models
01 Trained on reasoning trajectories
Fewer than 50–100 questions and their epistemic traces — no answer datasets, no reward models.
02 Deploys on a single GPU
Small enough to run where the expert works — on-prem, inside the firewall.
03 Debates internally
Trained to debate in latent space — the multi-perspective argument happens inside the model's representations, not in token-by-token prose.
Experts tune with words, not reward functions.
1
AI drafts an evaluation rubric
For the task, in the expert's own vocabulary.
2
The expert edits it in plain text
Like marking up a junior colleague's memo.
3
The rubric steers the debate
Lens selection and adjudication follow it at inference time.
4
The next trace improves
And the expert can read exactly what changed.
No gradient updates. Hours, not months.
The Reasoning Workbench: a Cursor-like engine for expert work.
01
MetaReasoner
Probes, branches, verifies.
02
Expert Foundational Model
Tuned per domain.
03
Long Context Handling
Full case files, not snippets.
04
Long Horizon Execution
Multi-step work, end to end.
05
Textual Feedback Improvement
Improves from expert notes — no retraining.
One engine · Any expert domain