# Q8 Artin zeros to height 100: isolation vs Turing count (10 Nov 2026, PT)

**Result [project — interval proof + literal Turing count].** Every positive zero of each of the three Q8 Artin L-functions with 0<t≤100 now has its own certified Hardy sign-change bracket of width 1e-12: 179 (2.2304.8t5.b.a), 180 (2.2304.8t5.a.a), 224 (2.42025.8t5.a.a). Counts equal the Turing-averaged N₊(100) in each case (b.a: Booker Thm 4.5 count, 179 plus exact central zero); 0 failed seeds, 0 overlaps. A sign change proves ≥1 zero per bracket; disjoint brackets equal to the certified total proves exactly one each. Not an RH claim.

## Diagnosis of the earlier run (185 vs 224, 39 failed) [project]
- **Not** a Turing-averaging error, **not** a window error, **not** missed sign changes or wrong seeds: PARI seeds numbered exactly the Turing count (179/180/224), the seed lists agree with Turing; the failures were all isolation failures, and only for t ≳ 83.5 (b.a: 83.51–99.86, a.a: 84.90–99.99, 42025: 84.56–99.40).
- **Cause: precision.** The kernel computes the completed function Λ(1/2+it), which decays like e^{−πt/2}-ish (value ≈1e-56 at t≈85, ≈1e-62 at 95). The theta sums cancel enormously; at 768 bits the Arb ball radius was ≈1e-59, so near t≳83 the ball contained 0 and no sign could be certified, at every bracket width. (Checked: value 1.5e-56 ± 1e-59 at 85.06, 3e-62 ± 1e-59 at 95.1.) The rigorous quadrature/tail bounds (~1e-102) were never the problem.
- **Fix:** ctx.prec 768→1536 (`q8_artin_isolates_100_v2.py`; results `q8-artin-isolates-100-v2-results.json`, log `q8-artin-isolates-100-v2-run.log`). All brackets then isolate at width 1e-12 with no fallback widths. Runtime 134 s, 134 s, 1092 s. Old v1 files kept as the failing record.
- Figure: figures/q8-artin-isolates-100.svg.

## Item 2: Gelfond [known argument + project check]
Already closed earlier today (open-questions): the merged 2/3/5 gap law does not need Gelfond's theorem; equidistribution of {k log q/log p} at times kπ/log p follows from unique factorisation (log q/log p irrational). Gelfond would be needed only for t↦t log p/π at integer times. Re-reading the argument found nothing to change.

## Item 3: independent coefficient audit [project — exact comparison]
`q8_frobenius_audit.py` / `q8-frobenius-audit-results.json`: for every unramified prime p ≤ 20000, a_p was predicted from the residue degree f of a prime above p in PARI's maximal-order decomposition (f=1→2, f=2→−2, f=4→0, from Q8 Frobenius class) and compared with lfunan(lfunartin) to 20000. 2260 unramified primes per field, **0 mismatches** for all three fields (ramified: 2,3 for the 2304 fields; 5,41 for 42025, all with a_p=0, these are not independently checked here). Composite coefficients follow by Euler product from prime values, so this covers the 20000-coefficient prefix used in the v2 run, up to the ramified/prime-power local factors and trusting that PARI's polynomial/nfinit is correct. Previously audited prefixes were 2000/5000.
