Experiment Results

From autoresearch/results/ and docs/SYSTEM_OVERVIEW.md § Results. 29 rounds, 145 companies. Composite formula in Autoresearch Loop.

Key runs

Tag	Composite	Extraction	FPR	Avg time	Notes
baseline	46.7	60%	37.5%	—	Pre-JSON-LD, no Places, generous name validation
jsonld-extraction	65.4	80%	6.3%	—	Round 1; +18.7 over baseline. Largest single gain.
google-places-v2	76.1	80%	23.4%	24.7s	Direct Places API; +29.4 vs baseline, +10.7 vs prior best
jsonld-v2	81.6	90%	21.2%	19.4s	All-time best. Committed pre-FPR-gate (round < 15).
extraction-v7 (prod)	63.7	50%	21.4%	43.8s	Current production config
quick-test (1 co.)	93.8	100%	0%	26.6s	Single well-structured site (More PR AB), not representative

Why production config (63.7) lags jsonld-v2 (81.6)

Source: SYSTEM_OVERVIEW.md § Results.

1. The 10-company extraction-v7 test set includes two timeouts (konsultopia.se, frimedia.se) and one parked domain — drags extraction rate down to 50%.
2. FPR still 21.4% — UI phrases slipping through Name Validation. Fix is batchable (add to Blocklists) but not yet shipped.

Full results-file index

Every *.json file under autoresearch/results/, with one-line provenance. Order: chronological. Columns: Extraction (%), FPR (%), total_contacts. Composite is recomputed per Autoresearch Loop formula when stated; otherwise left blank.

File	Date	n	Extraction	FPR	Total contacts	Note
google-places-v2.json	2026-04-02 15:51	10	80%	23.4%	47	Places API direct; +29.4 over baseline
jsonld-v2.json	2026-04-02 15:59	10	90%	21.2%	52	All-time best composite (81.6)
extraction-v7.json	2026-04-02 18:58	10	50%	21.4%	42	Production config snapshot
active-companies.json	2026-04-02 20:05	10	80%	0%	67	Live DB sample; FPR clean
db-companies.json	2026-04-02 20:26	10	50%	4.2%	24	Smaller DB-sourced sample
stockholm-ab-v3.json	2026-04-02 20:46	10	60%	0%	51	Stockholm AB cohort
email-association.json	2026-04-02 20:56	10	60%	0%	45	Email→name pairing experiment
final-clean.json	2026-04-02 21:03	10	70%	0%	52	Post-cleanup snapshot
uppsala-ab-v2.json	2026-04-02 21:26	10	60%	2.2%	92	Uppsala AB cohort; highest raw contact count
current-test.json	2026-04-06 02:43	5	0%	0%	0	Smoke run, all sources empty — likely env/key misconfig
current-test-3.json	2026-04-06 02:46	3	67%	0%	3	Re-run, partial recovery
fixed-test.json	2026-04-06 02:49	3	100%	3.3%	61	Fix verified; matches quick-test quality on a 3-co set
quick-test.json	2026-04-06 03:18	1	100%	0%	21	Single company (More PR AB), composite 93.8
latest.json	2026-04-06 03:18	1	—	—	—	Symlink-style copy of quick-test.json

New untracked artefacts (post 2026-04-06)

• current-test.json, current-test-3.json, fixed-test.json, quick-test.json — debugging session for the source-execution bug surfaced in early April. The arc reads: current-test (0% extraction, broken) → current-test-3 (67%, partial) → fixed-test (100% extraction, 3.3% FPR) → quick-test (100%, 0% FPR on 1 co.).
• continuous-history.jsonl — append-only log written by the new continuous loop (see Autoresearch Loop). Each line is one company test result. Schema: { org_nr, name, city, domain, domain_time_ms, crawlee:{contacts,emails,phones,time_ms}, firecrawl:{...}, maps:{...}, best_source, total_contacts, timestamp }.

29-round aggregate

• 145 unique companies tested
• 78 (54%) produced ≥ 1 contact
• ~450+ contacts total
• Best single company: 32 contacts (inviatech AB)
• 100% domain accuracy among the 78 productive runs
• 82/82 unit tests passing

Why 40% produce nothing

Parked domains, broken SSL, timeouts, sites with no person-level contact data. A real ceiling — no scraper improvement breaks past it. Fallback data sources blocked by ToS or IP restrictions (see Known Issues).

See also

Autoresearch Loop, Autoresearch Result Types, JSON-LD Extraction, Google Places, Name Validation, Known Issues.