# Social-factor research — Phase 8 operations

Policy: `social-maintenance-policy-v1`. The evidence remains publication
`social-findings-publication-2026-10-10-v1`, models v2 and diagnostics v3.

The existing hourly **Lens Master Pipeline** calls the parameterless
`POST /api/pipeline/tick`. No second schedule or public configurable writer
is introduced. The master lease owns source registration, due checks, research
task creation and monthly integrity receipts. Existing shared-source cadence
and paused settings are preserved. New source monitors use weekly (168-hour)
or 30-day (720-hour) intervals; these are checking frequencies, not publisher
release frequencies. Calendar months use Africa/Johannesburg.

Monitor seven official catalogues: DfT road safety, DfT mileage, the English
deprivation collection, 2011 Census qualifications, Census 2021 bulk data,
Home Office police-force crime tables and the Police.uk recording changelog.
Complete source captures and checks reuse pipeline_sources, pipeline_checks,
pipeline_observations and R2. New research records use lens_records kinds
social-maintenance, social-maintenance-task, social-source-seen,
social-integrity and social-integrity-attempt. These are never dataset rows.

First captures create baseline-review tasks. A different factual document
fingerprint creates one task for that source/fingerprint/publication; prior
source tasks are retained as superseded. Cosmetic changes follow the existing
semantic source check. Failures are deduplicated by source and most recent
successful check. Failed retrieval never means unchanged evidence. Each
calendar month also creates one gap-review task. Repeated request keys and
overlapping runs preserve the existing master idempotency and lease behaviour.

Six published artifacts are retrieved from the fixed Site origin and compared
against checked-in byte lengths and SHA-256 hashes: original and diagnostic
results, their protocols, the finding register and executive reading. The check
is monthly, bounded to 2 MiB per file and 12 seconds per request. Failed attempts
retry after 24 hours. A deployed checker revision can retry an older failed
attempt immediately; successful monthly receipts are reused. Redirects fail
the integrity check. All attempts remain in the ledger. This checks published
delivery integrity only: it does not check every archived input, download newer
publisher files, rerun model fitting or create fresh validation.

Public status is read-only at `/api/research/social-factors/maintenance` and
`/data-lab/social-factors/maintenance`. It reports source check/success/due
times, pauses, overdue checks, receipt dates and task totals. It never exposes
private review notes or actors. Unknown storage is shown as unavailable, not
healthy. The frozen research pages link to live operations without changing
their original archives or claiming a model refresh.

Admin → Review → Research tasks includes source and monthly research records.
Only the signed-in site owner can record `blocked`, `candidate_required` or
`reviewed_no_model_change`, with a substantive note. Closing a source task
requires the current observation to have been reviewed and no pause, failure
or conflict. Monthly closure requires a passed integrity receipt. Triage
does not approve a model or resolve the scientific gaps. Decisions are audited.

## A future analytical update

1. Inspect the current publisher files behind a changed page; record actual
   file hashes, release dates, covered years, definitions and boundary versions.
2. Create a new panel/protocol/results/publication version. Do not overwrite
   panel-v1, exploratory-v2, robustness-v3 or publication-v1.
3. Reconcile historical joint road/urban-rural exposure, vintage availability,
   geography, zero counts, exclusions and crime completeness. Do not treat
   2011 descriptors as current annual values or backfill a later deprivation
   edition into earlier forecast origins.
4. Before looking at new outcome relationships, freeze the new feature/model
   choices, assessment rows and acceptance criteria. The inspected 2023–2025
   outcomes cannot become unseen validation again.
5. Run the isolated reproducible build and compare matched exposure, counts,
   eligibility, calibration, temporal/geographic deviance and uncertainty.
   Report negative results and specification failures. Archive code, source
   fingerprints, predictions and fit diagnostics in the new version.
6. Review and publish the new finding register and interpretation together.
   Preserve original pages. Record an actual reviewer only when a review has
   occurred. Insurer claims validation remains a separate requirement.

The unattended task checks sources and reports new actionable failures or
research tasks once. It must not approve observations, close research tasks,
change model choices, overwrite code or refit/publish a model automatically.
No browser or numerical refit is implied by a successful maintenance tick.

The proposed next analytical comparison is recorded in
`/research/social-factors/maintenance-v1/decile-study-draft.txt`. It contrasts
separate deprivation/education deciles with continuous features. It is a draft,
not a fitted result; LSOA frequency still requires matched all-road exposure.