Key takeaways
- A rubric has four parts: hard rules, weighted preferences, exceptions and questions.
- Every change is published as a new version, and earlier screening runs keep the version they used.
- Hard rules are checked in code against extracted facts; preferences are scored per criterion with a rationale and citation.
- A missing value is Unknown, never a failure. A mismatch is surfaced for review, never an automatic rejection.
- Funds can keep different rubrics for different strategies or stages.
Planned feature: not in Delsivo yet, and details may change before it ships.
What goes into an investment thesis rubric
| Part | What it is | Example for a fictional fund, Northwind Seed |
|---|---|---|
| Hard rules | Conditions a deal must meet to be in scope | Stage is pre-seed or seed. HQ in Europe or India. Round size $500k to $3M |
| Weighted preferences | Things you like more or less, each scored 0 to 10 and weighted | B2B software (weight 30), repeat founders (20), evidence of paid usage (25), capital efficiency (25) |
| Exceptions | When a rule may be set aside, stated in advance | Pre-revenue is fine if a technical founder holds relevant patents |
| Questions | What to ask when something is unknown | What share of revenue comes from the top three customers? |
Weights are normalised in code, so they always add up to 100 however you enter them. The rubric is the same one the thesis fit report explains on each deal.
How a rubric is applied to a deal
- 1The deck is read and company facts are extracted with slide citations.
- 2Each hard rule is checked in code against those facts. The result is Met, Mismatch or Unknown.
- 3Each preference is scored by the model from 0 to 10, with a short rationale and a citation to the slide, form answer or other source used.
- 4The weighted total is computed in code from the criterion scores, never by the model.
- 5Exceptions that qualify are shown on the deal, so a reviewer sees why a rule was set aside.
- 6Unknown values turn into questions your team can send the founder.
Why Unknown is not the same as Mismatch
A deck that doesn't mention round size hasn't failed your round-size rule. It just hasn't answered it. Treating silence as a no quietly filters out companies whose decks are short, which at pre-seed is most of them. Delsivo keeps the two apart by design: an explicit mismatch and a missing value give different results.
Even a mismatch doesn't reject anyone. It is surfaced for an analyst to review. People move every deal, pass on it or take it forward, and record why.
Versions, strategies and re-runs
- Draft a change, preview it, then publish it as a new version. Only published versions screen deals.
- Keep separate rubrics, for example one for pre-seed and one for seed, or one per sector strategy.
- Re-running a deal against a new version creates a new screening run. The earlier run stays, with the rubric version, the evidence used, and the model and prompt version recorded.
- Which rubric version is published is always a person's decision.
How to write criteria that work
Criteria work best when a reader could check them from evidence. "Strong team" is hard to apply consistently. "At least one founder has shipped a product to paying customers before" can be checked and cited. Write each preference so the rationale has something concrete to point at, and write a question for any criterion founders often leave out of decks.