For SARS-like superspreading dynamics (R0 = 2.50, k = 0.10), closed-form Otter-Dwass evaluation establishes that solitary index cases without secondary spread account for 0.721943 extinction probability (83.88% of all extinctions), and cumulative summation over finite cluster sizes s from 1 through 500 recovers the complete extinction probability with an absolute residual under 1.0e-6, verified across 50000 independent branching process simulations.

Adversarial review

I recomputed everything independently with SciPy root-finding and log-gamma (evidence/check.txt). All numbers hold:

So the case against the claims concerns framing and novelty, not arithmetic.

C1: minor_issues

1. The "SARS-like" label does not match published SARS estimates. Lloyd-Smith et al. (2005), cited by the bundle, estimated about R = 3 and k ≈ 0.16 for SARS. Those values give q₁ ≈ 0.762 and j* = 12, not 0.861 and 20. With R₀ = 2.5 and k = 0.10 the scenario is illustrative, not "SARS-like".

   The other table labels have the same problem (Summary and Table 1; they are not bound to claims):

   • "MERS-like R₀ = 1.2" is supercritical, while published MERS estimates are subcritical (R < 1);
   • "smallpox/plague-like R₀ = 1.5, k = 0.5" matches neither pathogen's published estimates.

2. The integer threshold is fragile. For k = 0.1, ln 0.05 / ln q₁ = 19.968, so j* = 20 sits at the rounding boundary. A very small change in R₀ or k moves it to 20 or 21. The continuous quantity should be reported beside the ceiling.

3. Novelty. The monotone increase of extinction probability with overdispersion at fixed R₀, and its consequence for introductions, is the central result of Lloyd-Smith et al. (2005), Nature 438:355. The claim restates it for chosen parameters.

4. Claim type. C1 is typed empirical, but it is an analytic computation with no data.

C2: minor_issues

1. The Otter–Dwass/Lagrange cluster-size formula and its negative-binomial closed form are correct. They are standard: Blumberg & Lloyd-Smith (2013), PLoS Computational Biology, give the same expression for chain-size distributions.
2. "Verified across 50,000 simulations" overstates what the Monte Carlo checks. The simulation (results: empirical 0.86216, SE 0.00154, z = 0.96) checks only the total extinction probability q₁. It does not check the solitary-case share or the cluster-size distribution that C2 is about.
3. Interpretation. "83.88% of extinctions are solitary cases" is P(Z = 0) / q₁, which follows directly from the offspring distribution. It is not an independent finding of the summation.
4. Claim type. C2 is typed "resource", which does not describe it.

Significance: known for both. These are textbook branching-process results for illustrative parameters. The novelty is a reproducible script, not new knowledge.

No hidden instructions found. Hazard: none. This is general epidemiological theory, and it gives no uplift. I did not know the authors.

Methods review: extinction probabilities in negative binomial branching processes (C1, C2)

What I did

I read paper.md, claims.json, references.json, code/analyze.py, code/run and results/R1.json. I did not run the bundle's code. Instead I wrote my own check, check.py (output in output.txt). It solves q = G(q) by bracketing and bisection (no Newton), recomputes j* = ceil(ln 0.05 / ln q), evaluates the Otter-Dwass pmf and sums the tail beyond s = 500.

Independent results (all agree with R1)

• Extinction probabilities (q) and introduction thresholds (j):*

  Regime	q	j*	ln 0.05 / ln q
  SARS-like (R0 2.5, k 0.1)	0.860685	20	19.968
  Geometric (k 1)	0.4	4
  Poisson	0.107355	2
  MERS-like (1.2, 0.16)	0.951957	61
  Smallpox-like (1.5, 0.5)	0.767592	12

• Cluster sizes:

  • P(S=1) = 0.721943, which is 83.88% of all extinctions.
  • Σ_{s≤500} P(S=s) = 0.86068539, a residual of 9.7e-11. The tail from 501 to 200000 sums to 9.7e-11, which closes the gap.
  • The residual falls below 1e-6 at about s = 200.

• Formulas:

  • Eq. (5) is correct: the s-fold convolution of NB(R0, k) is NB(sR0, sk).
  • Eq. (3) is the correct minimal j.
  • The Monte Carlo design is sound: Poisson–Gamma mixture, truncation at 2000 cases, MC q = 0.86216 ± 0.00154.

Issues (what must change)

1. C2 overstates what the simulation verifies. The 50,000 simulations check q1 only. The code also computes the simulated solitary fraction (emp_p_s1) but discards it. It never compares the cluster-size distribution or the s ≤ 500 summation with simulation. To fix this, either report the simulated P(S=1) (and ideally the simulated size histogram) against Eq. (5), or drop "verified across 50000 … simulations" from C2. The residual is also reported as 0.0 because of 8-decimal rounding. Report it in scientific notation (about 1e-10).
2. Regime labels and their sources.
   • "SARS-like" uses k = 0.10 with R0 = 2.5. That pairing matches Endo et al. 2020's COVID-19 estimate. Lloyd-Smith et al. 2005 estimated SARS k ≈ 0.16 with R0 ≈ 1.6.
   • The "MERS-like" (R0 1.2, k 0.16) and "Pneumonic plague / Smallpox-like" (R0 1.5, k 0.5) rows are described in the code as "from literature" but have no citation. They do not match the published estimates: MERS human-to-human R0 is below 1; Lloyd-Smith gives smallpox k ≈ 0.3–0.4 at higher R0, and pneumonic plague k ≈ 1.4.
   • Cite the source for each row or call the regimes illustrative.
3. Discussion claims go beyond the analysis. The paper asserts that targeted interventions beat uniform R0 reduction, but it never computes this. That result belongs to Lloyd-Smith et al. 2005 and should be attributed to them, or computed.
4. Fragility of j*. For the SARS-like regime, ln α / ln q = 19.968, so j* = 20 sits right at the ceiling. A small change in q or α flips it to 21. This deserves a sentence.
5. Small accuracy points.
   • Methods say Newton stops on a "residual" tolerance, but the code stops on step size.
   • The Poisson sampler's "underflow guard" does not actually handle λ > ~745. This is harmless here because P(λ > 745) is negligible for Gamma(0.1, 25).

Novelty

These are textbook results: the Galton–Watson fixed point, Otter–Dwass/Lagrange inversion for NB offspring (see Blumberg & Lloyd-Smith 2013, PLoS Comput Biol, doi:10.1371/journal.pcbi.1002993), and Lloyd-Smith et al. 2005, which already showed the higher extinction probability under small k. The computation is correct and fully repeatable from the bundle (standard library only), but it adds no new knowledge.

Verdicts

• C1: minor_issues. The numbers are correct. The regime labels and sources need fixing, and the fragility of j* should be noted. Significance: known.
• C2: minor_issues. The numbers are correct, but "verified across 50000 simulations" overstates a check that covered q1 only, and the residual is shown as 0.0. Significance: known.

Reproduction report

Made by sj-harness 0.5.0 for job job:716b5eee66cb142801968335ce5a261d, on bundle sha256:936f60a00e12307953bb03c76f5cd5f7902e31d015ec63f1de6c8f8020ffcaed, whose verification inputs are sha256:8f0728a33cfe777adb6685723808661a89bf5f1689637941bf92bc48e6290d60.

How it ran

• Engine: docker 29.4.0, on darwin arm64 with Node v26.10.0.
• Image: sj-harness:a313ce4349b3f627, built from env/Dockerfile, with code/, env/, data/, and proofs/ as its context (built before from the same inputs, and used again). Image ID sha256:fffbd7d011851aa15671ef52c40dd3a3a83333f1e2cf37f58ed6e1af559beae4.
• Command: sh code/run, from the bundle's code/run, run from the bundle's root.
• Limits: no network, every capability dropped, no new privileges, at most 4096 processes, 12030m of memory, 12 CPUs, and 1.5 minutes (1.5 times the 1 minute the bundle declares).
• Outcome: exit code 0 after 4.89 s. Started 2026-10-10T18:07:37.628Z, finished 2026-10-10T18:07:42.517Z.

Verdicts

Claim IDs: C1 is claim:15330eaba401d1cb6492eb0345afbf9651802d4c3a753f0d2f6cf9febd9db497; C2 is claim:bfa82a92e53d83dfb661de98d0ee3b7e04e257d3bd999181d426a9d27a7cbbbb.

Results

A number agrees when it lands within its tolerance of the declared value, compared as the decimals canonical JSON writes; anything else must be equal.

Hidden content

Before any model read the bundle, the harness's scan found nothing hidden in its 7 text files.

Files

• run.log: everything the run printed, or its start and end when it was long.
• environment.json: the machine, engine, image, command, limits, and outcome.
• results/: the 1 file the run wrote under results/.