Methodology

BSAHI — Bitcoin Resource Accounting. How every number in this program is obtained, graded, checked, and reproduced. Version 1.0 · 2026-09-17 · Author: Prateek Poswal (Independent Researcher, ORCID 0009-0005-2139-1877)

1. Principle

A number may only be presented at the strength its evidence supports. This program therefore separates three layers and assigns every dataset a grade, and it treats a published correction as a normal output rather than a failure.

2. The three layers

The authoritative classification of every published number is the Evidence Matrix. Nothing may be called Observed if it is Reconstructed or Modelled.

3. Grades

• A — direct primary measurement (node RPC, on-chain fact), dated.
• B — frozen primary capture with a recorded unit and date (B* = with a documented gap).
• C — model using documented assumptions; the direction is the claim, the level is not.
• D — reconstruction by interpolation with no continuous source.

4. Instruments

5. Data sources

• First-party node — a Bitcoin Core node (pruned): gettxoutsetinfo, getblockheader, getblockchaininfo.
• Public network endpoints — mempool.space (fees), blockstream esplora (block summaries), blockchain.info (historical daily series).
• Frozen captures — every external series is captured once, stored under captured-data/, and read from disk thereafter, so a result can be recomputed offline and cannot silently drift.

6. Units and conventions

• Hashrate series are TH/s unless labelled EH/s; derived EH/s = TH/s ÷ 10⁶.
• Byte quantities are bytes; GB is decimal (10⁹).
• Fees are satoshis in captures, BTC in daily series, USD only where a price is applied and stated.
• Block space is sat/vB; block weight is weight units (WU) with the 4 MWU ceiling.
• Every dataset that carries a unit records it in a unit field; the integrity audit asserts derived units agree with raw ones.

7. Standing integrity checks

tools/research/integrity_audit.py runs on every change and exits non-zero on failure:

1. Heights/dates — every (height, timestamp) pair is monotonic; a height's chain date is kept distinct from its capture time (they differ by years while a node is syncing).
2. Anchor consistency — known on-chain anchors match their heights.
3. Units — derived fields agree with raw ones (hashrate, power, GB decimal).
4. Provenance — each dataset records schema + generated_at + a source.
5. Layers — each research dataset is classified Observed/Reconstructed/Modelled.

8. Reproducibility

• SCCR: three independent implementations (JavaScript, Python, C) agree to < 1e-6 on a frozen capture; research/reproduce/cross_check.sh prints the verdict; reproduce_sccr.ipynb walks through it.
• Sensitivity/CI: sccr_sensitivity.py is deterministic (fixed seed).
• Datasets: data/datasets.json is the versioned index with layer and grade.

9. Corrections policy

When a value is wrong, we publish the correction with its cause and regenerate every dependent value. Precedent: a 10× time-horizon error in our own model was found, corrected, documented, and every value regenerated — the conclusion survived, and the correction is part of the record.

10. Known limitations

• UTXO/state is Observed at one height only; the 2013-2026 series remains Reconstructed (grade D) until a node reaches the tip and the series accrues.
• Electricity price, ASIC efficiency, node throughput and operator wage remain assumptions (grade C/D).
• Per-block samples are a spot-check, not a survey.
• Field Core Web Vitals and search-console data are not yet instrumented.

12. Population and observability

Bitcoin's node population cannot be counted. Every attempt measures a partial view with its own bias, so this program publishes the views and their mechanisms, not a single confident number.

The role axis first. Miners (a few dozen pools) produce blocks and receive fees; validating nodes bear the storage and validation cost and receive nothing. BSAHI's N is the validating set. The two must never be merged.

Four quantities, never conflated:

Four instruments (first-party):

The direction of the error is stated, not hidden. Non-listening nodes are excluded from the index, so the reported network-wide storage externality (L_net) is a lower bound and the reported SCCR an upper bound; the baseline is conservative. The unpublicised-node band (N = 50K/80K/100K) is published alongside every SCCR figure.

Exact requests only. getnodeaddresses 32000 was a request ceiling that truncated the set; the instruments now use getnodeaddresses 0 (all known addresses). The first exact measurement returned more addresses than the ceiling had — which is why the ceiling figure was retired.

See Measuring the verification population and the Evidence Matrix.

11. Citing

Poswal, P. (2026). Bitcoin Resource Accounting. Bitcoin Sahi. ORCID 0009-0005-2139-1877. Data and code: CC BY 4.0.

← All research · ← Back to Learn