Databricks + Jev PDF Lab

Research prototypes for lowering PDF extraction cost and latency without relaxing completeness, exact values, provenance, or Databricks Precision Mode.

Current result: no quality-equivalent, end-to-end Jev payoff demonstrated. The repository publishes reusable code, synthetic tests, aggregate findings, and failures—not a production router or an accepted hotel-facts dataset.

What is here

• Metadata-only Jev routing, compact shared-state requests, strict response validation, and a single-attempt transport that retains ambiguous outcomes.
• Bounded concurrent request scheduling with durable round markers, no automatic replay, and explicit question-to-page binding independent of response order.
• Exact-output native TextPage reuse, page/region assembly, provenance checks, and scope-bound cache decisions.
• A selective-parsing rehearsal preserving physical pages, native geometry, managed elements, cross-page context, and unresolved review states.
• Reversible table-wrapper compaction with exact round-trip checks and original-content fallback; this is not a validated extractor input contract.
• Citation interval unions that preserve uncited gaps, and annotation-preserving compaction with synthetic regression tests; neither proves semantic support.
• Source-bound inline-price recovery into separate review-required shadow outputs, preserving exact literals, existing fields, and inherited citation IDs.
• Schema-wide field inventory that distinguishes missing, null, empty, and populated fields, with literal citation diagnostics but no semantic acceptance.
• Reversible execution of supplied, reviewed field selections with source/schema hash guards and existing citations; no automatic semantic discovery or acceptance.
• Field-only composition of retained cited responses, preserving untouched fields, exact input text, citation namespaces, annotation origins, and inverse restoration.
• Source-only table-scope candidates with exact spans and explicit abstention; layout hypotheses remain separate from semantic ownership and acceptance.
• Chained projection lineage that preserves the exact parent derivation and inverse restoration, without relabelling edits as service-generated evidence.
• Exact row-level repair excerpts with retained parent context, origin labels, and explicit omissions; supplied selections are not automated semantic routing.
• Parent-conditioned composition of retained donor markers, with exact source wording, existing citation spans, origin labels, and reversible lineage.
• Exact span-recipe compilation for SQL-native evidence reordering, with complete character accounting and explicit repeated ranges.
• A reversible, reviewed symbol-to-legend bridge adding only references to existing citations; supplied semantic mappings remain unaccepted judgments.
• BBOX-bound physical-page derivation and corpus-wide literal diagnostics, including output-identical cached-regex and literal-search alternatives.
• No-dispatch verifier request budgeting with full-context sharing, explicit field/owner bindings, and exact serialized reconstruction across four layouts.
• Matched single-question versus shared-state verifier scheduling with complete preflight allowlists, bounded attempts, fail-stop ledgers, and separate determinate/ambiguous reference accounting.
• Deterministic reconciliation of observed usage: signed billing corrections, decimal list pricing, unique job-run attribution, and explicit unresolved shared-warehouse charges. See billing methodology.
• The frozen hotel-facts schema and instructions; Precision v2.1 and citations must remain enabled in any extraction experiment.
• Results, including negative outcomes, doc-router source analysis, reproduction boundaries, and a public journal.

Run offline

Python 3.10+; the included tests need only the standard library:

python3 -m unittest discover -s tests -p 'test_*.py'
python3 examples/synthetic_rehearsal.py

These commands use fabricated inputs and stubbed transports. They do not read PDFs, access Databricks, or call Jev. The actual document experiments were performed only in the approved Databricks workspace, not on the local machine or GitHub Actions. Calling the transport functions directly would make a live request and requires separate credentials and data-boundary approval.

Findings at a glance

Strategy	Observed result	Important limit
Reuse native TextPages	0.316798 s saved out of 2.647231 s; 0.803929 s out of 2.892520 s	Two alternative native-only workloads across 236 pages, not pipeline savings
Compact Jev metadata requests	32.48% fewer input tokens; $0.005799402 → $0.003915828	26/236 recommendations changed; not equivalent-policy evidence
Four concurrent Jev requests	Median router wall time 15.674 → 4.366 s (72.14% lower), four paired trials	Same input-token cost; recommendation variability persists; not end-to-end savings
Smaller extraction groups	Four room records recovered; class-name context recovered	Still missing dedicated class records; extraction 19.826 → 42.791 s
Bounded Jev verifier	3/5 error signals flagged; 0/3 false alarms	Four correlated, previously inspected cases; not calibrated acceptance
Selective-parse rehearsal	50 cases / 36 unique bundles; all 236 pages retained	Saved full-parser outputs; no new parsing or proved work avoided
Reversible table compaction	Complete serialized envelopes 680,282 → 654,812 bytes (3.74% smaller)	Exact parser-output preservation, not measured inference savings
Synthetic compact-input probe	After checker review, compact STRING passes 10/10 narrow item checks	Post-hoc diagnostic; real source quality is separate
Full-evidence menu compaction	4.91% fewer input bytes; 112.352 → 117.254 s	One structured supplement price omitted; neither arm accepted
Source-bound price recovery	One shadow addition; compact amount coverage 55/56 → 56/56; no new inference	About 13 ms recovery kernel, 46.124 s enclosing job; incomplete quality review, no demonstrated savings
Schema-wide quality audit	All 33 schema paths inventoried; 43 unsupported basis defaults remain in compact and shadow outputs	Additional diagnostic work, not repair, semantic acceptance, or savings
Reviewed field projection	49 supplied corrections across two shadows; about 17 ms execution per arm	Post-hoc assistant selections; 33.834 s enclosing job plus unmeasured review/preparation; neither full output accepted
Targeted repair extraction	Enriched input recovers seven distinct legend entries; neither arm recovers conditional preparation properties	36.356 s extraction for both inputs together, plus preparation/evaluation/review; both rejected
Cited field-only composition	Seven-entry legend copied into two shadows; all other response fields unchanged	16–22 ms kernels, 25.397 s enclosing job, plus prerequisite repair/review; no new inference or accepted savings
Structural table-scope control	356 tables accounted for; one candidate matches eight reviewed target links	20 unsupported tables; 0.111 s scan kernel / 29.361 s job; no field repair, semantic acceptance, or savings
Conditional field projection	Two reviewed relations / four parent applications recovered without new inference	Six visual relations remain missing; 28.7 ms kernel / 32.489 s job plus prerequisites and unmeasured review; full output rejected
Row-level repair extraction	Five of seven targeted preparation-marker relations appear across two unpriced preparation records	Explicit parent linkage still missing; third case misses both markers and adds unsupported price bases; 44.302 s added extraction, no full-output merge
Conditional donor composition	Four reviewed visual relations / eight parent applications added without new inference	Two relations / four applications remain missing; 23 ms kernel / 27.553 s job plus prerequisites and unmeasured review; full output rejected
Adjacent evidence with SQL preparation	Both remaining preparation markers recovered in a separate record	No explicit parent applications; two unsupported price bases; 16.735 s extraction / 33.657 s summed statements, not end-to-end latency
Reviewed symbol-to-legend bridge	Two more relations / four parent applications added; known conditional inventory reaches 8/8 relations and 16/16 applications	3.3 ms bridge + 22.1 ms composition inside 37.211 s job, with prior inference/review additional; full output still rejected
BBOX physical-page derivation	640 page arrays changed; structural checks 1/20 → 19/20	Location consistency, not semantic support; 0.205 s kernels / 35.852 s job
Exact-output literal lookup	Median diagnostic kernel 24.566562 → 1.712976 s (93.03% lower)	180 timed outputs unchanged; diagnostic stage only, no end-to-end saving
Shared full-context verifier preparation	Repeated24 → shared24: 299,729,035 → 18,057,439 UTF-8 bytes (93.98% lower), same 148 prepared requests	No requests sent; token cost, inference latency and judgment equivalence unmeasured
Matched shared-state Jev verification	48-claim stage 17.59–17.71 → 1.67–1.91 s; 91.23% fewer reported input tokens	Shared batches match 44/45 determinate development labels versus 45/45 for singles; not equal-quality or end-to-end savings

Token costs use observed usage and a retained published rate. They are not invoices or full pipeline costs. See the results document for stage boundaries, job startup, diagnostic denominators, and quality failures.

Findings extend through E47, including the E29 checker correction, negative real-document compaction result, bounded router scheduling improvement, and reviewed source-bound corrections and field-only composition. The broader audit found field and scope defects; the latest corrections do not resolve all of them or establish payoff. The known conditional inventory is closed, not the whole menu or ten-document corpus. Source transcription, repeated symbols, full field/citation review, independent holdouts, and matched accepted-output economics remain open. E47 now measures hosted verifier overhead, not just serialized bytes. Its quality regression prevents promotion as an equal-quality replacement; two rounds on one inspected document do not establish independent generalization. The offline suite contains 662 synthetic tests; it does not reproduce private document accuracy or historical timing measurements.

Safety and scope

Typed output is not truth. Structural validity is not semantic correctness. Missing evidence must remain unknown, and review is not acceptance. Never infer allergen safety from an absent symbol or missing extraction field.

This is an allowlisted public snapshot. It excludes PDFs, rendered pages, source excerpts, extracted records, hosted request/response logs, credentials, workspace identifiers, source URLs, and private run receipts. The confidential evidence remains outside this repository; the public snapshot cannot independently reproduce the historical accuracy or timing claims. No source-bearing Jev payload is included or authorized by publishing this code.

See PUBLICATION.md for adaptations and known prototype limits. No open-source license has been selected; public visibility alone is not a license.