{"page": "Burrough/Burrough Mystery", "failed_blocks": {"28": 0.0784313725490196, "80": 0.021505376344086023}}

================================================================================
BLOCKED PAGE — diagnosis (investigated 2026-07-01)
================================================================================
This page was NOT combined because 2 blocks failed the ASR quality gate after all
retries. gen_lang.py only checks this file's EXISTENCE (skips the page so the run
can finish); the notes below are for the human doing the fix.

FAILURE CLASS: RUNAWAY-SHORT (both blocks) — VoxCPM generates 3-8s of audio for
a ~75-90s block, then stops. The model "skips to the end" of the block, speaks only
the closing colophon/citation, and terminates. Audio IS produced (no crash) but is
truncated to the last 1-2 sentences. Seed re-rolling and timestep escalation (8→10→12)
all produce identical short output — this is structural, not a sampling fluke.

--------------------------------------------------------------------------------
block 28  (1525 chars)  best_sim = 0.078   ->  RUNAWAY-SHORT (last sentence only)
--------------------------------------------------------------------------------
Block ends with the Hebrews 4:12 citation:
    "...lebih tajam daripada pedang bermata dua mana pun, menembus sampai
     memisahkan jiwa dan roh, serta sendi dan sumsum, heb 4:12."
VoxCPM generated ONLY that closing sentence (~7-8s) across all 3 retries (RTF 0.92).
The "heb 4:12" end-of-verse citation at the tail apparently acts as an end-of-text
marker — model attention collapses to the last fragment and halts early.

Observed: gen 6.8s → audio 7.4s (should be ~75s). ASR heard:
    "Lebih tajam dari pada pedang bermatah dua mana pun, menambu sampai memisakan
     Jua dan Roh, sertasen d"  [= garbled reading of the closing citation only]

NOTE: Nearly identical text appears in Burrough%20Mystery block 28 with the same
failure. Two duplicate clean_text files produced two separate chunk directories.

--------------------------------------------------------------------------------
block 80  (1785 chars)  best_sim = 0.022   ->  RUNAWAY-SHORT (closing date only)
--------------------------------------------------------------------------------
Penultimate block of the book (84 total). Two compounding triggers:
(a) Mid-block untranslated English sentence:
    "the word of god, in the city of london, the fourteenth of the twelfth
     month, 1662 , di mana puisi ini ada dalam tulisan tangan..."
    IDN voice model cannot handle English text — triggers degenerate output.
(b) Block ends with a short signature+date colophon:
    "...edward burrough, london, bulan kesembilan 1658."
    Model attends only to this closing line and generates ~3.7s of near-silence.

Observed: gen 6.3s → audio 3.7s (RTF 1.72). ASR heard fragments like
"Aragat wad puro, London, bulan kesambilan..." = garbled closing date line only.
All 3 retries identical.

================================================================================
HOW TO UNBLOCK
================================================================================
Block 28 — SPLIT at sentence boundary before the "heb 4:12" citation:
  1. In output/idn/blocks/Burrough/Burrough\ Mystery.json, split block 28 so the
     citation becomes its own short block (removes the end-anchor trigger).
  2. Delete the stale .ok markers for block 28, remove this .blocked file, regenerate:
       rm "output/idn/combined_mp3/Burrough/Burrough Mystery.blocked"
       ./run_noneng_cpp.sh idn --no-chunk --only "Burrough Mystery"

Block 80 — CLEAN UP English contamination first:
  1. In clean_text/Burrough/Burrough\ Mystery_IDN.txt, translate or remove the
     embedded English sentence ("the word of god, in the city of london...").
  2. Re-chunk this page and regenerate block 80.

NOTE: "Burrough Mystery" and "Burrough%20Mystery" are duplicate source files.
Deduplicating them would halve the work here.