K4 · live cryptanalysis

Phase 2B.1 tested the Weltzeituhr keystream attack under exactly one (derivation rule, city list) parameterization: `(first_letter, 1988_best_guess)`. The strategy's docstring promised additional derivation rules and the search-space framing always implied multiple city-list orderings; neither were implemented. Phase 2B.1.B implements both axes.

If the Weltzeituhr keystream is the actual K4 mechanism, the correct parameterization (derivation rule, city list, offset, direction, alphabet) lies in the cross-product:

{first_letter, last_letter, length_mod_26, vowel_count, consonant_count} × {1988_best_guess, sector_reverse, alphabetical, alt_cairo_istanbul, alt_london_paris} × 24 offsets × 2 directions × 2 alphabets

= 5 × 5 × 24 × 2 × 2 = 2,400 unique candidates per pass.

If any region of this space is warm (per the distribution observer's acceptance criterion: ioc_p95 ≥ 0.05 across 12 consecutive flushes), the (rule, list) combination at the warm region's center earns follow-up: deeper sampling, higher-resolution offset variants, denser alternate orderings around its sector.

If no region is warm after 7 days under this enlarged parameterization, the Weltzeituhr-keystream family as currently parameterized is null-equivalent and should be demoted in the queue pending the verified 1988 engravings landing.

attacks/strategies/weltzeituhr_keystream.py:

1. 5 derivation-rule functions: - _first_letter_keystream (existing — kept) - _last_letter_keystream (NEW) - _length_mod_keystream (NEW — name length mod 26 → A–Z) - _vowel_count_keystream (NEW — vowel count mod 26 → A–Z) - _consonant_count_keystream (NEW — consonant count mod 26 → A–Z) 2. 5 city-list variants: - 1988_best_guess (existing — kept) - 1988_sector_reverse (NEW — 24 cities listed UTC+11 down to UTC-12) - 1988_alphabetical (NEW — same 24 cities sorted A→Z) - 1988_alt_cairo_istanbul (NEW — UTC+2 sector swapped CAIRO → ISTANBUL; both are documented as having appeared in the sector across 1980s photographs) - 1988_alt_london_paris (NEW — UTC±0 sector swapped LONDON → PARIS; documented as a transient mid-1980s engraving) 3. **iterate() loops over (rule, list, offset, direction, alphabet) instead of just (offset, direction, alphabet).** The params dict gains derivation_rule and city_list_id fields; both already propagate to phase_distributions via the observer's region key. 4. Per-pass uniqueness budget documented at the top of the file.

• Crib validator (core.cribs.validates) is still the only oracle.
• The runner's overall flow, persistence, distribution observer,

rejected-sample logging, and live-state cadence are unchanged.

• The four verification gates (crib + statistical + red-team +

re-derivation) still apply to anything that survives.

• Phase 2B.1's negative result for (first_letter, 1988_best_guess)

is preserved in this new sweep — that combination runs again as one of the 25 (rule, list) cells, so the observer's distribution for it should reproduce.

By 24 hours into the new sweep:

• ≥ 2,400 distinct param regions in phase_distributions (matching

the new search-space cardinality).

• Per-region histograms statistically meaningful (≥ 1 K attempts each).
• Distribution observer's per-region p95 distribution analysable as a

whole — Chi can publish a heatmap (rule × list) of mean p95.

• Any (rule, list) cell whose p95 distribution clearly exceeds the

random-uniform null cohort gets flagged for promotion.

• Accept (warm cell found). Any (rule, list) cell whose

ioc_p95 ≥ 0.05 across 12 consecutive flushes for at least one region within it. That cell becomes Phase 2B.3 with a denser sweep around its center.

• Kill (cold uniformly). All 25 cells average p95 < 0.045

across 7 days of observation. Demote Weltzeituhr family pending verified 1988 engravings; advance to next attack family in the queue.

• Inconclusive. Mixed signal at 7 days → extend observation by

7 days, then force a decision.

1. Cardinality overrun. 2,400 region keys per pass × 288 flushes/day = 691 K rows/day raw in phase_distributions — well above the 60-region/day estimate the observer was sized for. Storage budget revisited: ~170 MB/day raw before rollup. Server-side compact_phase_distributions() still keeps lifetime growth bounded (rolls raw → hourly > 7d, hourly → daily > 90d) so the table remains queryable, but the daily raw growth is ~40× the original estimate. 2. False-positive risk from larger search. 2,400 candidates per pass means 2,400 chances for an IOC to land in the upper tail purely by chance. The 12-consecutive-flush requirement is the primary defense (a chance-warm region won't sustain across the full hour); document this in the per-cell heatmap. 3. City-list variants are still best-guess. The five lists in this phase are not from primary sources — they're documented variations to test the strategy's sensitivity to ordering. A warm result under one of them is suggestive, not proof. Verified 1988 engravings remain the gating dependency.

• attacks/strategies/weltzeituhr_keystream.py (extend rules + lists +

iterate())

• attacks/runner.py (phase string Phase 2B.1 → Phase 2B.1.B)
• experiments/2026-05-02-phase-2b1-conclusion/conclusion.md (already

written — closes 2B.1)

• experiments/2026-05-02-phase-2b1b-derivation-and-city-sweep/

(this file)

• web/data/progress.md (entry on the close-out + new phase)

Land the strategy + runner change in one push. Runner picks up on next systemd reload, starts a new attack_runs row tagged Phase 2B.1.B, and the distribution observer begins populating phase_distributions with the new region keys (which include derivation_rule and city_list_id). First batch of regions visible at /admin distribution tile within ~5 minutes after restart.