Freeze the batch
Lock exact IDs, the development branch, agent ownership and authenticated access.
Open research log · July 2026
One working day supervising nearly five months of research.
A real AI workflow investigated 1,697 messy car boot sale records, navigated thousands of pages and changed a live production directory. This is the evidence, the failures and the public transcript.
68,862 public events · four disclosure checks passed
2,447searches
767domains
303images inspected
15research agents
The problem
There is no single source of truth.
Car boot sales are published across organiser websites, Facebook posts, council calendars, flyers and local directories. Some have no website. Others leave old pages online after moving or closing.
A scraper can collect words. It cannot reliably decide whether a sale runs this Sunday, whether 6am is for sellers or buyers, or whether an old venue has been replaced. That requires search, navigation, comparison and judgement.
Messy sourcesstructured evidenceuseful visitor copy
The publication gate
Finding a matching name was only the beginning.
A record became an active public page only when the evidence could support a real decision by a visitor.
Verification skill · production contract
No self-approvalThe complete gate used for every listing
Evidence
Find and establish the record
Prove identity
Match organiser, venue, address and postcode against the actual source.
Decision
Determine what is true now
Decide lifecycle
Classify active, closed, moved, replaced, duplicate, malformed, unrelated or uncertain.
Enrich every fact
Check schedule, dates, times, fees, category, amenities and stale values independently.
Publication
Build a useful visitor page
Pass the copy gate
Write natural About and What-to-Expect copy without research mechanics or invented atmosphere.
Finish the image gate
Approve an exact inspected image, or record a source-specific terminal rejection.
Release
Prove the reviewed state
Validate in development
Read back exact-ID writes; test routes, relationships and sentence-level copy locally.
Promote once
Use one conflict-checked transaction, then prove parity, images and deployed pages.
A gate failsResearch, correct, merge or classify honestly Re-enter at the failed gate
All eight gates passRelease the reviewed batch to production
The database, decision logs, events, images and public pages must all describe the same reviewed state.
Any gate failsStop, return or classify honestly
Research again, correct the batch, merge a duplicate, or retain the listing as uncertain or closed. A plausible page is not a pass.
What the workflow refused to publish
1,200 records did not become an independent active page.
This includes hard rejections, uncertain records and evidence merged into stronger listings. The categories explain more than one rejection percentage ever could.
432
Historical or circular evidence
Old listings, expired posts and websites repeating the same unsupported claim proved that an event once existed—not that it operates now.
285
Closed, moved or replaced
The original event had ended or its organiser now directed visitors to a different venue.
204
Still uncertain
The evidence conflicted or remained incomplete, so the record was preserved without presenting it as confidently active.
121
Merged into a stronger listing
The information was useful, but it described an event already represented by a better-supported canonical page.
81
Outside the directory scope
The record was not a relevant car boot, market or related public sale.
77
Duplicate or malformed
The source could not identify a distinct, usable event without inventing missing facts.
“An old page proves that an event existed. It does not prove that someone should drive there this weekend.
Human challenges that changed the loop
The workflow improved because its conclusions were challenged.
These were not cosmetic corrections. Each challenge became a reusable rule for every venue that followed.
Existence is not current operation
Finding an old page was initially treated as proof. The human challenge changed the rule: a historic mention only proves that something happened once.
Process change
Require current evidence linking organiser, venue and schedule before assigning active status.
A schedule is more than a weekday
All-year sales, seasonal runs, selected dates, weather policies, buyer entry and seller arrival were being collapsed into one field.
Process change
Verify each practical fact separately and surface conflicts instead of silently choosing one source.
Research notes are not visitor copy
Technically accurate descriptions still read like database reports and source commentary.
Process change
Add a copy gate that rewrites verified facts into concise, human language without losing caveats.
Images cannot be an afterthought
Photo research fell behind venue verification and generic pictures could pass without representing the event.
Process change
Make image review a terminal gate with venue relevance, source traceability and explicit approval.
Production is not a research workspace
Progressive edits against the live database made mistakes harder to isolate and review.
Process change
Move work to a development branch, preserve per-venue decisions and promote one validated set at the end.
Parallelism needs a contract
Early subagents stopped when they proved existence, missing times, costs, amenities, relationships and images.
Process change
Use bounded batches and structured research packets, while one main workflow owns consistency and the final gate.
Was the AI worth it?
It changed where the human time went.
The human still defined “verified”, challenged weak decisions and approved production. The repetitive research and transformation happened at machine scale.
See the resulting directoryEstimated manual equivalent460 hoursAbout 4.5–5 calendar months at five productive hours per weekday
vs
AI-assisted experiment61h 20mAcross six days, with approximately one working day of human supervision
API-equivalent token cost~£1,320
UK contractor equivalent£40k–£50k
Inspect the evidence
Search the public session transcript.
Switch between the readable 2,575-message conversation and all 68,862 retained events—including tool calls, searches, edits, agent turns and usage checkpoints.
Loading the public conversation…