tri fpga-batch-rehearsal · the batched wire protocol, rehearsed off-board

26 checks, no board, no RTL: a clean link returns every receipt with no retransmit, 11 loss and fault cases recover, and a lie, a duplicate or a wrong-key answer stops the run. self-test: PASS.

  ok   clean link, two passes: every receipt (220+594 frames), every row, no retransmit, no resync      
  ok   frame economy: SETX {PLANES}xC then RxC DOT6 per pass, >4x fewer than today                      
  ok   one budget across both passes: retry space untouched, node pinned once                           
  ok   the RTL hunting parser agrees with the aligned model on a clean link                             
  ok   recovered, 1 retransmitted, 1 resyncs: 12 bytes lost mid SETX answer                             
  ok   recovered, 1 retransmitted, 0 resyncs: a whole DOT6 answer lost, stream still aligned            
  ok   recovered, 1 retransmitted, 1 resyncs: 16 bytes lost across two DOT6 answers                     
  ok   recovered, 1 retransmitted, 1 resyncs: the last answer of the pass lost                          
  ok   recovered, 3 retransmitted, 2 resyncs: two holes, one in each phase                              
  ok   recovered, 1 retransmitted, 0 resyncs: a flipped tag bit on a DOT6 (fault damage)                
  ok   recovered, 1 retransmitted, 0 resyncs: a swallowed DOT6 answer (fault drop)                      
  ok   recovered, 1 retransmitted, 1 resyncs: an answer under a nonce never issued (fault nonce)        
  ok   recovered, 4 retransmitted, 1 resyncs: one request byte lost (a w byte): the parser eats the next
 frame's magic                                                                                          
  ok   recovered, 1 retransmitted, 0 resyncs: a request's last byte lost: the frame limps, the parser mi
s-answers it                                                                                            
  ok   recovered, 1 retransmitted, 0 resyncs: a request's AA lost: the node never sees that frame       
  ok   a workload with a y-nonce collision exists (search found one)                                    
  ok   the colliding frame (ys[2..5]=[3, 0, 1, 0] spell an issued nonce) settles by tag, not by a false 
status/duplicate stop (1 both-shape frames)                                                             
  ok   still stops the run on a wrong y byte with a valid tag ('lie')                                   
  ok   still stops the run on a cell signing with another key ('exhausted')                             
  ok   an unkeyed cell (status 0x04) stops the run at its first answer, no retry                        
  ok   one DOT6 answer served twice stops the run ('duplicate')                                         
  ok   RAM drift after an honest SETX is never credited: retries burn out and the run stops             
  ok   zero-plane skip: identical rows, the all-zero chunk's 12 DOT6 frames skipped                     
  ok   14 random hole patterns: 6 stopped on the lie, 8 passed clean                                    
  ok   rehearsal pair: honest and corrupt links agree on every row (5 retransmits over 3 holes)         
self-test: PASS                                                                                         
trinity-fpga $

recorded 2026-10-04 04:52 UTCreal 4.7 sshown 4.7 sexit codes 0

$ tri fpga-batch-rehearsal

Staged: the prompt and the typing. Real: every byte the commands printed, at the time they printed it. Any silence longer than 2 s is shown for 2 s, and the title bar says so while it happens. Edited: home directory shown as ~.

GIF Preview card