01 · Sources
Where the data comes from
Every project page on Puna pulls from three ingestion paths; all of them are auditable.
- Public corpus. Operator filings and regulator disclosures (SEC EDGAR, ASX, HKEX/Shenzhen, TSX-V, AIM), US Geological Survey and Department of Energy disclosures, government mining-authority and lithium-strategy announcements (Argentina, Chile, Bolivia, and provincial), EU Critical Raw Materials Act designations, provincial cadastres (Argentina, Chile), technical reports (NI 43-101 / JORC / S-K 1300), and project announcements.
- Web ingests. Continuous monitoring of news, ministry announcements, and operator communications. Sources are tagged & dated; nothing accepted without a citation.
- Community contributions. Files dropped into the public corpus by users that know something we don't. Tagged by contributor type (known subscriber / web-ingest / anonymous). Reviewed by AI before inclusion.
02 · Pipeline
How the data becomes a project page
Three stages, run on every project in the dataset.
Stage 1
Ingest + structure
Raw documents are parsed, KPI values are extracted with the exact source sentence preserved, and stored with a chain back to the source document. Values that can't be found in the corpus stay null, and they render unsourced rather than fabricated.
Stage 2
Pre-compute synthesis
Puna runs a monthly synthesis pass (project summary, peer comparison, investment thesis) using a large language model anchored to the corpus. Each output cites the documents it's built on, so a reader can trace any claim back to its source.
Stage 3
On-demand research
Agent Research, available on every project page, runs a fresh brief on request: synthesising the public corpus into a tailored read of what stands out, adding in the private files from the user. Each brief cites the documents it draws from, and adapts to the user's context.
03 · Provenance
Every number traces to a source
Every KPI on a project page (grade, resource, planned production, capex, opex, chemistry) either points back to the document it came from, or renders subdued and unsourced. There should be no third state.
- Hover any value. A small marker on each sourced cell opens a tooltip with the document title, publisher, and the exact sentence the value was extracted from.
- Click through. The tooltip links to the original document on the publisher's site — the NI 43-101, the JORC technical report, the regulator filing, the company release.
- One number, many sources. Where multiple documents corroborate the same value, the strongest source shows in the tooltip and the rest sit in an expander. Where two documents disagree, the page surfaces the disagreement rather than silently picking a winner.
- Subdued means honest. When the corpus doesn't carry a value, the field renders dim and labelled "not in Puna dataset". We should never present a guess.
04 · Honesty
What Puna does NOT do
The most interesting line from any AI product, is what it refuses to do. Ours is specific.
Company Strategy
What Puna does: Puna connects institutional investors, acquirers & royalty companies to project owners. We do this by making lithium projects more understandable for investors, and easier to develop for miners.
What Puna does not do: We don't own mining tenements or land positions. We don't drill, build plants, or provide mining services. This creates independence, which allows us to serve mine operators and capital allocators without a conflict on either side.
Data Strategy
- We do not invent data. If the published corpus does not hold a resource estimate, the field on the page reads "— not in Puna dataset," not a guess.
- We do not run ads. There are no sponsored slots on any Puna surface. The Monday brief is not monetised by advertisers.
- We do not sell user data. Email addresses given to us for magic-link access or the Monday brief are not shared, traded, or rented.
05 · Cadence
How often data refreshes
- News + operator communications: monitored daily by the intel specialist. New items surface against each user's watchlist as they happen.
- Project metadata: operator, stage, grade, resource, planned production, and ownership history, is re-checked at least monthly against the latest public corpus pass.
- Pre-computed synthesis: summary, peer comparison, investment thesis is regenerated monthly.
- Recent intel + community contributions: surfaced continuously as the ingest happens.
- Agent Research briefs: generated on request, using the corpus as of the moment of the click.
- Monday brief: one email a week, on what moved in lithium relative to the watchlist of the user. No filler.
06 · Control
Your data, your control
Puna is built on contributions to a shared public corpus. But you keep meaningful control over what you put in, what's private to you, and how Puna treats you on your own surfaces.
- Private data. If you upload files into "Your Vault" or into "Your Uploads" this data is private by default, and not accessible to others unless you approve it to be. It is accessible to your Puna agent when you "toggle Puna ON" so that it can serve you via its analysis.
- Making files public. A file enters the public corpus if you upload it when you are signed out (anonymous), if you upload it directly to a project's page (even when you are signed in) or if you push it from private to public. Once you make a file public you can't make it private again.
- Easy to delete: Your private contributions have a delete button. One click removes it. No support ticket, no email request. It's gone the moment you click.
- Personalisation off-switch. The "Puna ON" toggle controls whether Puna uses your private uploads to personalise your own queries. Off means your adviser still works; it just stops reading your context.
07 · Coverage
Being honest about gaps
Our coverage is global by intent but uneven by reality. Argentina, Chile, and the broader Lithium Triangle are well-covered because we have been operating there for years and because the public corpus is deep. Other geographies (early-stage African pegmatite, brine claims, certain Tibetan plateau projects, smaller private operators) are thinner.
If you see a gap on a project you know, drop a file into the public corpus on that project's page.
We know we can't be perfect. But what we can do is represent the truth, as best as we can see it.
08 · Principles
Intelligence doesn’t solve everything
- Commercial alignment matters more than intelligence. In the long run, conflicts of interest don't work.
- Present the truth, as best as we can see it. No marketing gloss. The truth can only be what it is.
- Name the gaps where they are. Honesty about absence is worth more than a confident guess.
- Intelligence informs. Users decide. The judgment stays with the user. Its not our job to act for them.