Counting protocol

Rules version: 1.0, 17 September 2026 (R1–R11). The rules are not edited: changes come out as a separate version 2.0 beside this one, and counts under version 1.0 stay as they are.

Every figure on the examples page is a Fermi estimate made before the rules were written down. Here it is the other way round: the rules come first, and the example is computed under them afterwards. The order is fixed in the repository history: the commit with the rules precedes the commit with the calculation. Page history on GitHub.

1. Rules

Written before the calculation and not changed after it. If a rule turned out badly, that is recorded as a result of the test, and the number is not fitted.

R1. Inputs. Only external, checkable quantities: documents, statistics, publications. Every number with a reference. The site author's own estimate is not an input.

R2. What is counted. One channel of emission, declared in advance. Side channels, such as fame, stories about the person and people inspired by them, are not counted unless declared before the calculation.

R3. Minutes M. A saved life is the whole remaining life of the person saved; for a newborn, 1 Vt = 2.5 × 10⁷ minutes. A text read is reading time × number of readers. A speech heard is duration × listeners.

R4. Horizon. The account runs from the start of emission to the moment of calculation. The future is not counted, except for lives already begun: those are counted to their end.

R5. Degree G. Fixed categories; the nearest one is chosen, interpolation is forbidden.

R6. Share. List every link in the chain without which the result would have been different. For each link, two numbers: p, the probability the result would have happened without it; d, the fraction of the horizon lost to delay if it had. Raw weight w = (1 − p) + p × d. Share of a link = w / Σw. This is a chain approximation of the Shapley value: reproducible and efficient.

R7. Sign S. By result. Search for documented harm from the same channel: lawsuits, product recalls, medical complications, recognized victims. If found, S = (benefit − harm) / (benefit + harm) in minutes. If the search was done and found nothing, S = +1, stating where the search was made. S = +1 is not assigned without a search.

R8. Coherence Q. The fraction of emissions whose sign by intention matched the sign by result. For one channel with one documented intention: 1 if they matched, 0 if not.

R9. Error. Every input is an interval. The result is computed at the extreme values and reported as an interval and an order of magnitude. No point figure is reported.

R10. Judgment points. Every place where a rule does not determine the number uniquely is marked ◆ and given an interval with an explanation. Their count is also a result of the test.

R11. No fitting. After the result is obtained, rules, categories and intervals are not changed. A discrepancy with previously published figures is recorded, not removed.

2. Declared example

James Harrison (1936–2025), Australian donor of plasma with anti-D antibodies. Chosen because he has exactly one channel of emission, a deed, and the inputs exist in documents: the number of donations, the number of doses, the epidemiology of Rh disease.

By R2 only the plasma channel is counted. His fame, and donors who came because of his story, are not counted: that is a second channel, and its rules are not declared here.

Inputs to be found before the calculation:

1. Number of donations and years.
2. Number of anti-D doses made with his plasma, and the year the program began.
3. Number of births in Australia during the program.
4. Mortality from hemolytic disease of the newborn before and after prophylaxis.
5. The share of the fall in mortality attributable to anti-D rather than to intrauterine transfusion and neonatal care.
6. Size of the anti-D donor pool in Australia.
7. History of the discovery of anti-D: how many independent groups.
8. Documented harm from anti-D.

3. Inputs

Input 2 does not enter the calculation: a dose is not a life. It is listed because the earlier estimate grew out of it.

4. Calculation

Minutes (R3, R4). Deaths averted by the whole Australian program over 58 years: 15.3 × 10⁶ × (46 − 1.6) / 10⁵ = 6,800, of which a fraction f is attributable to anti-D.

◆ f = 0.5–0.9. Sensitization fell 80–90%, but part of the fall in mortality came from transfusion and neonatal care. The 46 → 1.6 series is British; carrying it over to Australia is also a judgment, the Rh distribution of the populations being similar.

Deaths averted: 3,400–6,100. Severe disability averted (kernicterus in survivors) ◆ 0.5–1 per death, R5 category "disability", G = 0.5. Life-equivalents: 4,250–9,150. Each is a newborn and by R4 counts to the end: 1 Vt.

Impact of the whole program: (4.3–9.2) × 10³ Vt, that is (1–2) × 10¹¹ peirces.

Share (R6). Links in the chain from discovery to the dose in the midwife's hand:

For Harrison, p splits into two numbers. ◆ s is his share of the pool's antibody: one of 100–200 donors, but 51 years and a high titre; 1–5%. ◆ r is the fraction of his supply the program could not have replaced: if donors can be immunized to order, the shortfall closes within a couple of years out of 51, r ≈ 0.04; if supply is scarce (Australia imports immunoglobulins, the pool is small), every donation is marginal, r = 1. Interval r = 0.04–1.

Harrison's share = w / Σw: from 0.0004 / 1.51 to 0.05 / 0.67, that is 0.03–7%.

Result. I = (1–2) × 10¹¹ × (0.0003–0.075) = 3 × 10⁷ – 2 × 10¹⁰ peirces:

I(Harrison) = 1–700 Vt, order of magnitude 10¹ Vt.

Sign (R7). A web search was made for adverse events and contamination via anti-D. It surfaced the Irish and East German hepatitis C outbreaks, each from a single infected donor (NEJM, PubMed); both are other pools, other countries. No Australian case surfaced; Lifeblood and National Blood Authority reports were not read for harm. S = +1.

Coherence (R8). The intention is documented: after transfusions at fourteen he pledged to become a donor, and donation was his only channel. The sign of intention and the sign of result matched. Q = 1.

Judgment points: ten. Carrying the British mortality series over to Australia is not marked separately: it is folded into the interval of f.

5. What the test showed

The earlier figure was wrong by one to four orders of magnitude. The examples page gave Harrison 5–30 kVt. Under the rules he gets 1–700 Vt. Two errors: "2.4 million" is doses and mothers, not babies saved; the whole program saved thousands, not hundreds of thousands. And the 3–6% share was taken by plasma volume, not by replaceability: the donor pool is replenished by immunizing volunteers, and without any one donor most doses would still exist.

The interval spans three orders of magnitude, and half of it is one number. Of ten judgment points, r, the replaceability of a donor in a replenishable pool, contributes a factor of 25 on its own; the other nine together contribute about 25 more, the largest of them, s, his share of the pool's antibody, being 5. This is exactly the warning on the mathematics page: v(S) is not measured but estimated. The rules are reproducible: a second person with the same inputs gets the same interval. Not the same number.

What holds. Even at the lower bound Harrison is above the typical person (0.4 Vt); even at the upper bound he is an order of magnitude below Trump (10 kVt). The whole Australian anti-D program is (4–9) × 10³ Vt, ten thousand biographies. Which row stands above which holds, the absolute figures do not, as the manifesto promises.

Which rules worked and which did not. R5 with fixed categories worked: no dispute about the degree arose. R7 worked unexpectedly: without the mandatory search for harm the sign would have been assigned automatically, and the product does have harm on record, if in other countries. R6 fell short: the rule does not say how to count the replaceability of one member of a replenishable pool, and that produced the widest interval. A rule is needed, but by R11 it is not added here.

The other tables on the site have not been re-run under these rules. Harrison's row on the examples page has been corrected with a link here; the other rows are the earlier Fermi estimates, and after this test they should be read as orders of magnitude, not figures.

6. The second test: a blind recount

On 19 September 2026 two counters, Claude and GPT, received the same rules R1–R11 without the Harrison example, two declared channels, and counted separately without seeing each other's numbers. The cases were ones that do not appear on the site. Claude's count was written down before GPT's answer; its inputs were taken from memory, without references, which is weaker than R1. The assignments, both answers and Claude's count are in the repository history (in Russian). Here the order is not fixed by the repository history: the assignment was committed before the run, while Claude's count and GPT's answers landed in the same commit.

The first run failed. GPT read the rules literally: p, d, remaining life and the number of readers are not external verifiable quantities (R1), so it gave each of them its whole mathematical domain and arrived at "from zero to infinity". R10 requires an interval with an explanation but nowhere says that the interval must be substantive; until now this was visible only from the example. For the second run a clarification was added to the assignment: trivial intervals are forbidden, every judgment point gets an argument, the chain is limited to the links named in the sources. The rules themselves were not changed.

The order of magnitude agreed in both cases, and the intervals overlap almost entirely. Both counters chose the same degree both times (1 and 0.2): R5 with its fixed categories worked again. The lists of links nearly coincided. Both found the same for the sign: documented separation trauma in the children, and charges of manipulativeness against the book.

The agreement is partly luck: for Winton the differences in M and in the share cancelled each other. Behind them are places where the rules are silent:

• R3. It does not say that a Vita is counted in waking minutes: GPT counted 24 hours a day. It does not say how to count remaining life, or whether a saved life should be multiplied by the probability of death without the rescue: Claude multiplied, GPT did not. It does not say how to get from copies sold to people who read.
• R6. It is unclear against which horizon the delay d is measured when the window of opportunity is closing: nine months before the war, or eighty-seven years up to today. It is unclear what counts as the result for a text: the reading of this book, or the skill the reader would have taken from another. It does not say where to cut the chain.
• R7. A single review found forces the sign to be computed to six nines after the decimal point. A threshold is needed below which harm counts as not found.
• R10. It does not define how to count judgment points: GPT has 15 and 12, Claude five each, because it did not count every pair of p and d.

Caveats. Two cases are few. Both counters are language models with similar habits, and their agreement is weaker than the agreement of two people. By R11 the rules are not corrected here: the list above is groundwork for a second version of the protocol.

7. Sources

• Australian Red Cross Lifeblood, James Harrison's story: 1,173 donations, more than 3 million doses since 1967.
• CBS News, James Harrison, credited with saving 2.4 million babies, dies at 88.
• Australian Bureau of Statistics, Births, Australia, 2024; yearly series 1967–2024 via Demographics of Australia.
• GLOWM, Anti-RhD immunoglobulin for the prevention of HDFN: mortality 46 → 1.6 per 100,000 births.
• PMC, Forty years of anti-D immunoprophylaxis; BioDrugs, The role of immunoglobulins in neonatal Rhesus haemolytic disease: contribution of prophylaxis and of other methods.
• ABC News, There are only 100 anti-D plasma donors in Australia; Lifeblood, Pregnancy, anti-D and plasma.
• Lasker Foundation, Vaccine for preventing Rh incompatibility in newborns; Ronald Finn.
• NEJM, Clinical outcomes after hepatitis C infection from contaminated anti-D immune globulin; PubMed, Contaminated anti-D immunoglobulin in the GDR.
• ABC News, Australia's reliance on imported blood plasma: immunoglobulin imports.