सिद्ध — back the work, not a product
Siddha is an open research program, not a thing on sale. We are mapping how much of machine intelligence can be grown from data in closed form — no backpropagation, no GPUs — and saying, in public and with numbers, exactly how much cannot. This is the stage where backing matters most: before any product, while the result is being hardened and the walls are still being mapped.
Every number we publish is reproducible from the public benchmark code. Backing funds the next experiment — it does not buy a conclusion.
Four ways to move it forward
Money, machines, data, or a sharp critique — each one advances the science. None of them buys equity or a product, because there isn't one yet.
The work that moves Siddha forward is unglamorous and concrete: more benchmarks, harder datasets, the noisy-data robustness study reviewers asked for. Sponsorship buys researcher months and the compute to run them — no equity, no product, a research program you help carry across its next result.
Sponsor a result →Synthesis is strongest where models must be grown where the data lives — clinics, courts, factories, public offices. Bring a real, labeled corpus and a problem, and we will synthesize against it in the open and publish what we find, win or lose.
Propose a pilot →Synthesis itself runs in seconds on a CPU — but mapping the walls does not. Larger-scale vision runs and 100-million-token bits-per-byte sweeps need real machines. CPU and GPU credits go straight into experiments whose numbers land on the research page.
Offer compute →Our strongest results and our published losses sit at the same size, with the scripts to reproduce them. The most valuable contribution is a sharp critique: re-run a benchmark, break a claim, or point at the wall we have not seen yet.
Argue with us →Support is structured as research funding and partnership. There is no product and no security on offer — only a research program you can help carry across its next result.
Where the support goes
We do not raise against a vision. We raise against the next experiment — and report it back at the same size whether it wins or loses.
Compute and researcher time for harder, larger evaluations — the noisy-data study, larger-scale vision, and the bits-per-byte sweeps that keep the n-gram ceiling honest.
The work of bringing real-world corpora into the open protocol — synthesizing against them where the data lives and publishing the result either way.
Researcher months to extend the architecture, write up results at the size of the wins and the losses, and answer the reviews in public.
What backing does not buy is a finding. If an experiment we fund returns a negative result, it goes on the research page with the wins — that is the discipline that makes the rest worth believing. See the results so far →
Back us directly
Pledge as a patron, sponsor a result, crowdfund a benchmark, or offer compute and in-kind help. Tell us the lane that fits and roughly what you have in mind — there's no minimum, and no equity or product changes hands.
Every message reaches the team directly and gets a personal reply. Backing funds the next experiment — never a conclusion.