pipelines/DEEP_READ_PILOT_2026-07-21.md
The first real run of the deep book-processing contract: 37 Opus readers each mining one
source cover-to-cover against live canon, producing a per-book EXR extraction record, then
ten per-shelf fresh critics with fixers. Provenance: wf_c97f3dee-711 (50 agents, zero errors).
Records: docs/book_extractions/EXR_<source_id>.json.
The pilot proved the read-each-book system end to end and surfaced its own first lessons.
All ten shelves returned GO. The care guardrails held on their own: readers refused to make
real or revered figures (Jesus, Pilate, Aquinas, Descartes, Spengler) into entity candidates
per §17.1, flagged the ebu gogo as elevated-care cryptid (genuine living tradition, framed
authorially, never resolved into "proof"), and kept the vril firewall.
Richest books: BALI Sekala Niskala (62 rows) · the two Dunsany volumes (42 + 39, the fairy-literary
reception layer) · the Keo land-perspective ethnography (33) · the Batak power-of-writing study (33).
mounting signs; the player traces the signs back to a forgotten promise and helps the heir
honor it. A self-contained mystery-into-restoration arc, drawn from Allerton's Wae Rebo
fieldwork — the book the region page marked pending on MUST-ACQUIRE.
across Nage/Lio/Manggarai villages; the contradiction between them IS the puzzle, logged
feature-by-feature, never resolved into proof (the Homo floresiensis echo authorial-only).
neutral village to weigh two clans' rival oral histories on a counterweight bar and deliver
a standing-shifting ruling. From the Keo tuka-timba tradition.
Wall," "The Shade-Stalker's Weakness": the real biology of the world's largest lizard
(midday shade-retreat, refusal to swim, thermal limits) becomes readable tactical rules for
a guardian-test the player passes without a kill.
a whole class of reading-the-oracle puzzle minigames (the datu's oracle table, the
auspicious-day layer, the fowl's-gut omen) grounded in real manuscript practice.
1. RECORD-SHAPE NORMALIZATION (the shelf critics' one MAJOR, per shelf). Readers used a
lighter record shape than contract §2 on 7 of 10 shelves — every seed carries seed_kind +
description + proposed_layer + citations, but only the 3 fixer-touched shelves (~82 seeds)
carry the full routing set (significance_lenses, region_ref/chapter_ref FKs, quest_type).
FIX for v0.2: pin the §2 field set in the reader prompt as a hard schema (or run a cheap
normalization pass before ratification). The records are proposal-tier, so this blocks
nothing now; it is the gate the merge step enforces before any row enters a ratification lane.
PARTLY DONE: the flores fixer already authored the freeze — docs/proposals/EXR_RECORD_SCHEMA_v0.2.md
(the reconciled record shape, subordinate to contract §2). The remaining work is running the
deterministic re-serializer across the ~7 lighter-shape shelves' records, a merge-step task.
2. MISFILE CAUGHT: SRC_00015 is not a Flores source — it is an Åbo Akademi welcome address
(abstract Western philosophy, zero Flores content), misfiled on the flores-indonesia shelf.
Re-shelved to cross_cutting this pass. A region-scoping-correctness catch — exactly the
failure class the noise gate exists for.
3. SIX ZERO-YIELD RECORDS: SRC_00231/00487/00491/00492/00493/00494 (the "admin-Journal-manager"
Scripta/Donner artifacts) returned empty across all classes — thin front-matter/admin pages.
Flagged as re-triage candidates (extraction_status left not_deep_read? — set to
deep_read_extracted with an empty-yield note; a future pass confirms they are administrative,
not thin-but-real).
4. REGION-PAGE CITATION FIX the readers caught: flores_island.md §3/§20 cites SRC_00177 as
"...reduced Range"; the actual paper is "...reduced DISPERSAL" (Jessop et al. 2018). Queued
for the next region-page pass.
The contract, the tier stack, the EXR home, the lifecycle column, and the §8 wiring are all
live. The pilot confirms the deep read produces rich, cited, care-clean, game-surface-named
harvest. The one gate before the region-by-region rollout is the v0.2 record-shape pin so every
record enters the merge step well-formed. That is the next director step, not a Josh decision.