Bidlo production coverage audit October 5 2026

Read-only structural audit of 70 shared collections, 397 fields and 6,333,963 active shared documents. The results identify nonfinite quantities, an impossible deadline, broken shared graph references, a repeatedly failing classification job and weak provenance. Source expectations and lifecycle coverage limit how much missing data can be classified as a customer-facing defect.

Findings and actions

Project presence and source coverage

The project denominator is 81,770. These counts show presence, not missing-data defect rates. They retain canceled, completed and unknown-status projects. They exclude soft-deleted documents and tenant/user overlays.

Compared with the framework baseline, project population increased by 131. Source-array and Bid Date presence each increased by 131; file, item and subcontractor project counts were unchanged. This is a comparison of two live observations, not a reconciled regression analysis.

All remaining sources and every source × agency × bid month × lifecycle × ingestion month cell are retained in project-cohorts.csv. Bid month uses the production date parser in UTC for grouping; it does not validate source timezone semantics. Ingestion month uses document created_at and is kept separate. Alias reconciliation and parser/run cohorts remain unclassified. Unknown references remain explicit buckets.

Lifecycle and eligibility

52,217 projects (63.9%) have unknown or unresolved lifecycle. Counts follow the current Project Status tag; lifecycle accuracy is unverified. The source capability contracts, extraction SLAs, archive adapter, required/optional rules and attribution grace periods are not established in this audit. Eligible denominators, customer-use-case usable counts and usable rates remain unknown; absent optional or derived fields are not labeled defects.

Items quantities units and prices

All 119 format failures are literal NaN, so they fail finite usable-quantity requirements. 120 items have no quantity row. Zero and negative quantities are review categories rather than automatic defects; adjustments and source semantics may justify them. Source basis, unit compatibility and schedule row-count reconciliation were not verified.

1,228,800 items have at least one valid Unit target; 4,912 do not. 739 items have multiple Unit targets and 2,601 have multiple Scope targets. The field metadata does not declare these single-valued, so these are compatibility review candidates rather than confirmed cardinality defects.

1,210,260 active shared items have a valid Project link. Of the remaining 23,452, 22,351 have no active shared Project edge and 1,101 link to tenant-owned projects. These cannot all be labeled failed extraction: review unattached staging rows and ownership semantics separately.

Prices are profiled in the field matrix for Bid Amount, Bid Items Price, Project Items Unit Price, estimates and subcontract prices. A parsed amount alone does not establish currency, price basis, bidder/version, award state or source accuracy. The 11 Engineers Estimate format failures include ranges that require separate range fields or a validated adapter, rather than an invented single amount.

Graph references and duplicate candidates

Relationship checks cover direct and declared reverse fields, soft-deleted endpoints, expected collection and shared scope. The main matrix records at least one valid target separately from presence of any edge. Graph totals in graph-integrity.csv also include active edges attached to deleted or tenant documents; they are technical observations and must not be counted as customer-visible defects.

69,731 legacy File → Project edges use the Project Files field. 69,723 have active shared endpoints and all 69,723 duplicate normal forward counterparts across 17,009 projects. They add zero project coverage. Validate storage adapters before cleanup; the declared-direction field census remains unchanged.

The evidence exports contain 3,933 affected relationship IDs for the selected five high-impact fields and 1266 affected tag-value document/field references. Other graph failures and empty fields are retained in the complete matrix.

UUID/tag adapters found unresolved representations in Country, Set Asides, Delivery Method and legacy Planholder/Subcontractor fields. Country includes 4,419 plain labels. These are adapter review candidates, not proof that the underlying country or entity is wrong. Valid UUID references to inactive, nonshared or wrong-collection targets are reported separately.

Exact source-tag + Project ID groups produced 2,806 candidate groups involving 7,044 projects. This is not a deduplicated canonical identity: distinct contract versions, reused identifiers, mirrors and addenda must be reconciled. Cross-source mirrors and fuzzy title duplicates were not assessed.

Provenance files and pipeline

The seven-day value-change profile contains 5,542,490 shared base changes. It profiles changes rather than all historical facts or current values. The query uses database now() minus seven days; observations were taken during this audit, without one transaction-wide snapshot.

Only 509,657 changes (9.2%) have a run ID. Seven distinct nonblank IDs match neither scraper_runs.run_id nor scraper_runs.id. This establishes an unresolved namespace/mapping gap, not seven failed ingestion runs. TxDOT and FDOT recent change rows have no run IDs; no historical backfill was inferred.

File metadata exists for 134,617 of 175,311 active shared files; only one metadata row has a nonblank SHA-256 and positive byte size. 100,186 files have page records spanning 4,912,827 page rows, with no nonpositive/null page numbers. Page and metadata presence do not establish item-to-page/table/row lineage, download success or external document capture.

The seven-day run snapshot has 1,084 runs: 87 failed, 995 successful and 2 abandoned. 572 runs report zero rows. Zero-row success is not classified as a failure without a source discovery manifest. project_classification_sync accounts for 79 failures; the root cause is unconfirmed by the sanitized error-pattern checks.

There are 129 schedule records and 121 effectively enabled records. Some last-success timestamps are old or absent. Manual, annual and backfill entries cannot be assessed with a blanket seven-day threshold; cadence, aliases, scheduler execution and source SLAs need reconciliation. Schedule observations are exported without asserting outages from age alone.

Framework checks that need additional evidence

Field coverage matrix

All 397 fields are included, including newly introduced, empty, hidden and computed fields. 16 collections have no shared active documents. Presence rates use the observed collection population; zero populations have no rate. 65 fields in populated collections have less than 5% presence. This is a review list, not a completeness score.

Field IDs are stable keys. Scalars exclude owner/user overlays and deleted rows. Relationship presence uses declared-direction edges, while validated target presence requires active shared endpoints in the referenced collection. One valid endpoint does not guarantee that every edge is valid. All observed scalar row counts equal distinct document counts, so no multiple scalar rows were observed in the scoped document/field pairs. Semantic conflicts were not evaluated.

Download the full field matrix CSV

Method and reproducibility

Queries ran against the postgres-prod read-only connector on October 5, 2026, approximately 06:29–06:59 America/Chicago. Counts are batched live observations and are not a single atomic snapshot. All existing shared collection records were included because the collection schema has no deletion/archive state; visibility is not treated as lifecycle. Deleted fields/documents/values/relations and tenant/user overlays were excluded from the coverage matrix.

20 owner-null documents belong to tenant-owned collections and are reported as scope review observations; they are excluded from the 70-collection denominator. There were no owner-null active documents with missing collection IDs. This is not a tenant data audit.

A broad scalar scan and broad graph scan were cancelled after exceeding the intended batch budget, and Soil Area was split into field batches after its first attempt was cancelled. Subsequent collection and field batches completed. The connector enforces read-only transactions and has a 10-minute server timeout. A SET LOCAL prelude returned an unsupported multi-result response; subsequent long batches used audit-side cancellation at roughly 45 seconds instead. No indexes, functions or production tables were created.

Numeric validation accepts signed decimal/scientific notation and rejects NaN/Infinity; it does not validate sign semantics or arbitrarily large exponent usability. Dates use the existing pure parse_stored_date_value function, accepting epoch/ISO/supported legacy dates; source timezone accuracy is unknown. Location checks use the existing pure safe_geojson_point parser. All 1,643,601 populated location values parse as Points; Point parsing does not prove physical worksite accuracy or geographic consistency.

Exports contain counts, UUIDs, source/agency labels and selected nonsensitive numeric/date evidence. No personal contacts, credentials, signed URLs, full extraction content or raw error logs are exported. Every check outcome is an observation, a confirmed structural/format failure, a review candidate or an explicitly unverified check. Proposed remediation has not been executed.

Audit files

• findings.csv
• entity-findings.csv
• entity-findings.jsonl
• field-coverage.csv
• project-cohorts.csv
• source-summary.csv
• quantity-findings.csv
• money-and-date-findings.csv
• relationship-findings.csv
• reference-findings.csv
• assignment-candidates.csv
• duplicate-candidates.csv
• graph-integrity.csv
• pipeline-runs.csv
• schedule-observations.csv
• provenance-by-source.csv
• production-audit-results.json
• reproducible-queries.json
• registry.json
• audit-manifest.json