Hard rules, preferences and exceptions

A rubric has four parts. Hard rules and exceptions are checked in code, with no AI model involved. Preferences are scored by the AI model from the deck, with a citation for each score, and the arithmetic runs in code. Questions collect what to ask the founder. You edit all four on a rubric page in Settings → Thesis and rubrics. See Build your thesis rubric for drafts and publishing.

Hard rules#

Click Add rule and choose what to check:

Each rule has a Name, the Check, its value, and an optional Question when this is unknown. Without your own question, a standard one is used, such as "How much is the company raising in this round, and in which currency?"

Where the values come from#

Rules read the deal's facts. When a fact has more than one value, the most trusted wins: a value your team set beats a founder correction, which beats what the deck says. The fit report names the source of each value: Deck, Founder correction, Your team or Application form.

Stage, sector and geography#

Pick the Accepted values as chips. Delsivo recognises common wording for each value, for example "pre seed" or "angel" for Pre-seed, and "SaaS" for B2B software. Cities count for their country and region, so a company in Pune counts as India and South Asia.

For sector and geography you can add Other words that count, separated by commas, such as "insurtech, payroll" or "Nairobi, Kenya".

• A value the rule recognises and accepts is Met.
• A value it recognises outside your list is Mismatch.
• A value it can't place, or no value at all, is Unknown, with a question.

Round size and currency#

Set a Minimum, a Maximum or both, and the Currency: any currency, with the most used ones first. Type to search by code or name. Leave one side empty for "at least" or "at most". Amounts show in short form under the box, such as USD 1M or INR 4.5 Cr.

If the round is stated in another currency, Delsivo converts it at approximate reference rates and says so, for example "INR 20 Cr (about USD 2.4M at reference rates) is within USD 1M to USD 4M." If the deck gives an amount but no currency, the rule is Unknown and asks which currency. Reference rates cover the common currencies (US dollar, Euro, British pound, Indian rupee, Singapore dollar, UAE dirham, Canadian dollar, Australian dollar, Swiss franc, Japanese yen, Hong Kong dollar and Chinese yuan). When the round and the rule are in different currencies and one of them has no reference rate, the rule is Unknown and says so; check it by hand. Because the rates are approximate, check rounds close to your limits by hand and override if needed.

A field is stated#

Met when the deal has a value for that fact. Otherwise it is Unknown, with a question. It is never a mismatch.

Unknown is never a failure#

A missing or unreadable value is Unknown, shown with a dashed outline and never in red. It doesn't count against the deal: it becomes a question for the founder and an item under Diligence readiness. A Mismatch is shown for review and is not part of the thesis fit score. Neither one moves the deal or rejects it.

Exceptions#

An exception lets a stage, sector or geography rule accept something extra when a condition is true, for example "Allow pre-seed when the team has a prior exit".

1. In Exceptions, click Add exception. You need a stage, sector or geography rule first.
2. Name it under Exception.
3. Choose the rule it Relaxes, and the values it should Also accept.
4. Under Look in, choose Anywhere in the deck or one fact, such as The team fact.
5. Enter the words to look for under For any of these words, separated by commas, such as "prior exit, acquired by, sold to".

When the rule would be a mismatch only because of a value the exception allows:

Preferences#

Preferences are what make a deal more attractive, such as early revenue traction or founder and market fit.

1. Click Add preference.
2. Name the Preference.
3. Under What good looks like, describe it specifically. The AI model reads this next to the deck.
4. Set the Weight from 0 to 10 with the slider.

Next to each weight you see its share of the fit, for example "8 of 10 · 28% of the fit". Weight 0 turns a preference off.

When a deal is screened, the model scores each preference from 0 to 10 with a short rationale and quotes from the slides. A quote that isn't on the slide it names is dropped. If nothing in the deck supports a score, the preference is Not scored instead of getting a low score.

The thesis fit is the weighted average of the scored preferences, from 0 to 100. Preferences that couldn't be scored are left out and the remaining weights are rescaled. With no preferences, there is no thesis fit score; hard rules still run.

Questions#

Questions are what you want to ask the founder. Add one with Add question, then choose When:

• Always ask: listed on every fit report.
• When revenue is unknown (or another fact): raised only when that fact is missing.

The fit report's Questions for the founder combines your questions with the ones raised by Unknown rules and possible exceptions. Repeated questions are listed once.

Limits#

A rubric can hold up to 30 hard rules, 20 preferences, 20 exceptions and 30 questions. To publish, it needs at least one hard rule or one preference with a weight above 0.

Frequently asked questions#

Our deal says "Seed" but the rule shows Unknown. Why?#

The rule found text it couldn't place, or the fact is missing. Open the rule on the Fit tab to see the value found. If you know the answer, override it with the right value and a reason. See Override an assessment.

Does a founder correction change the result?#

Yes, on the next screening. Founder corrections win over the deck. Run the screening again from the Fit tab to use them.

Can the AI model change a hard rule result?#

No. Hard rules and exceptions are checked in code. The model only scores preferences.