Concept study · $BTAI · v0.4 draft
An intelligence
layer for Bitcoin
native data.
BitcoinAI is a design study for a research stack that reads public chain and market data, reasons over it with an ensemble of models, and returns an answer with its sources attached. It holds no keys, moves no funds, and predicts no prices.
Design targets
- Fixed concept supply
- 21,000,000
- p95 inference budget
- 41 ms
- Answers carrying a source trail
- 100%
- User keys ever held
- 0
Targets for the reference design, not measurements of a live service.
01 — Thesis
Bitcoin is already legible to machines. It is barely legible to people.
Every block is public, every transaction is verifiable, and none of it arrives in a shape a person can reason about quickly. The gap is not more data — it is interpretation that can be checked. That is the job this concept assigns to a model: compress the chain into an explanation, then hand over the receipts so the explanation can be dismantled.
The stack is deliberately narrow. It answers questions about state that has already been settled, it names its uncertainty, and it stops at the point where a human has to make a decision.
-
i
Legibility over volume
One paragraph a reader can verify beats a dashboard nobody audits.
-
ii
Provenance by default
Every claim carries the block height, endpoint, and window it was drawn from.
-
iii
Stated uncertainty
Confidence is a labelled field, not a tone of voice. Thin data says so.
-
iv
Restraint as a feature
Non-custodial, no automated trading, no forecast dressed up as analysis.
02 — Systems
Three subsystems, each with a job it can be held to.
Ingestion and feature building
Blocks, mempool state, and public venue quotes land in an append-only feature store. Nothing is overwritten, so any answer can be replayed against the exact inputs that produced it.
- Sources in the reference design
- 7
- Feature window
- 144 blocks
- Store shape
- append-only
- Backfill depth
- 18 months
Retrieval, ensemble, and disagreement
A small ensemble reads the retrieved features alongside protocol documentation. When members disagree beyond a threshold, the spread is surfaced in the answer instead of being averaged away.
- Ensemble members
- 3
- Disagreement threshold
- 0.18
- Retrieved passages per query
- 12
- Answers abstained on, target
- 4.6%
Explanation with a source trail
The output is prose a person can read, followed by the block heights, endpoints, and feature versions behind each line. A verifier pass re-checks every figure against the store before the answer is released.
- Verifier pass
- mandatory
- Citations per claim
- ≥ 1
- Replay window
- 90 days
- Export
- JSON + Markdown
03 — Architecture
One path from block to answer, with a gate before the exit.
-
01
Ingest
Nodes and public venues stream in. Raw payloads are hashed on arrival.
-
02
Features
Derived series are versioned and written once, never edited in place.
-
03
Ensemble
Models read features plus retrieved protocol text and draft an answer.
-
04
Verify
Figures are re-derived from the store. A mismatch blocks the response.
-
05
Deliver
Prose, confidence, and the full source trail leave together or not at all.
04 — Limits
What this concept will not do.
Written down first, so the scope cannot quietly drift later.
- No price prediction. The stack describes settled state. It does not forecast markets.
- No custody. No keys, no deposits, no withdrawals. There is nothing to take.
- No autonomous trading. Agents may draft a proposal; a person executes or nothing happens.
- No unsourced claims. If the source trail is missing, the answer is withheld.
- No financial advice. Research output, read at the reader's own risk.
05 — Network
A fixed supply, allocated to the work that has to happen.
21,000,000 $BTAI
Illustrative allocation for the concept. No token has been issued, and nothing on this page is an offer to sell one.
| Allocation | Share |
|---|---|
| Research & model development | 30% |
| Contributor rewards | 20% |
| Ecosystem & partnerships | 15% |
| Liquidity | 15% |
| Core team, locked | 10% |
| Advisors | 5% |
| Distribution & grants | 5% |
06 — Roadmap
From a written specification to a stack that can be audited.
-
Phase 01
Genesis
Specification, threat model, and brand system. Feature schema drafted in public.
-
Phase 02
Activation
Ingestion prototype, first ensemble, verifier pass wired in front of every answer.
-
Phase 03
Expansion
Replay tooling, published model cards, external review of the source trail format.
-
Phase 04
Evolution
Open evaluation set, community-run verifiers, and abstention rates reported openly.
Status