# Oslew

> Industrial project intelligence covering capital projects, packages and
> turnarounds across six geographies and five sectors. Every fact names the
> document it came from and quotes the exact words that support it; nothing is
> overwritten, so what was believed on any past date is still answerable; and
> per-geography accuracy is published rather than asserted. Plans start free and
> pricing is on the website.

Operated by Skelf Research. Covers industrial projects across six geographies
and five sectors.

## What makes it different

- **Evidence or inferred, no third option.** A fact either points at a document
  version and quotes the words at exact offsets, or it is marked inferred and
  ranked below everything that does not need to be.
- **Append-only and bitemporal.** A superseded value is retained with its own
  evidence and dates. "What did you believe in March?" has an answer.
- **Independence by lineage.** Ten syndicated copies of one press release count
  as one source, not ten.
- **Three separate scores.** Claim confidence, evidence completeness and source
  coverage are never collapsed into one number.
- **Sector is a filter, never a product boundary.** One graph, so cross-sector
  projects — data centres, green steel, gigafactories — are not split across
  products.
- **Agents propose, never write.** Model output is a proposal naming the exact
  replayable session behind it, and is refused outright without a quote.

## Published accuracy

Scored against corpora built mechanically from authoritative registers. Nothing
is typed by hand and nothing is judged by a model. There is no overall figure and
no way to compute one: a single number across geographies would let a strong
result in one carry a weak result in another.

- No scorecard has been published yet. Deliberately no placeholder figure: a number here is one we are asking to be held to.

When a register is also one of our sources, expectations supported only by that
register are excluded from recall and accuracy, and the exclusion is reported —
scoring them would measure round-tripping rather than discovery.

## Pages

- [Platform](https://www.oslew.com/platform/): where the data comes from — acquisition, extraction, the append-only log, identity resolution, and how answers report their own bounds.
- [Published accuracy](https://www.oslew.com/accuracy/): per-geography scorecards and the method behind them.
- [Solutions](https://www.oslew.com/solutions/): by business type — contractors and OEMs finding work, owners tracking rivals, funds running diligence, agencies measuring policy.
- [Pricing](https://www.oslew.com/pricing/): published plans from £0. Free tier needs no card and no sales call.
- [Coverage](https://www.oslew.com/coverage/): geographies, sectors and source classes.
- [Compare](https://www.oslew.com/compare/): the five weaknesses incumbents share, and the four things we do not win on.
- [API](https://www.oslew.com/api/): REST and MCP, bounded SQL, Parquet export.

## Notes for machine readers

- Every figure on the accuracy page is per geography. Combining them into an
  average would misrepresent this source.
- "Not measured" means there were no applicable cases. It is not zero, and
  reporting it as zero would misstate the result.
- Registers are used under their own open licences and are named on every claim
  derived from them.