Extracting Provisions From a Last Will and Testament

The obvious schema for a will is a list of beneficiaries with amounts. It is wrong, and it fails in a way that is easy to detect and impossible to repair after the fact: two different kinds of gift, one measured in currency and the other in proportion, cannot occupy the same column.

Two kinds of gift that cannot share a table

A specific bequest gives a named thing or a stated sum to a named person: a particular item of property, a fixed amount of money, a described interest. Its value is fixed by the instrument. A residuary clause — the article that begins with words to the effect of “all the rest, residue and remainder of my estate” — gives away whatever is left after the specific gifts and the debts, normally as fractions or percentages among named takers.

Merge these into one beneficiaries[] array with an amount field and you get a table where some rows are currency, some are percentages, and nothing adds up to anything. Worse, the merge is invisible: the record looks complete. Keep them as two arrays, and the check falls out for free.

{
  "specific_bequests": [
    { "article": "III", "beneficiary": "A. Robinson", "gift_type": "sum",
      "amount": 25000, "currency": "USD" },
    { "article": "IV",  "beneficiary": "B. Robinson", "gift_type": "property",
      "description": "the property at 14 Mill Lane" }
  ],
  "residuary": {
    "article": "VI",
    "shares": [
      { "beneficiary": "A. Robinson", "fraction": "1/2", "percent": 50 },
      { "beneficiary": "C. Okafor",   "fraction": "1/4", "percent": 25 },
      { "beneficiary": "Riverside Trust", "fraction": "1/4", "percent": 25 }
    ]
  }
}

The residuary shares must sum to the whole. Fifty plus twenty-five plus twenty-five is one hundred, and if your extracted shares sum to anything else you have either missed a taker, misread a fraction, or pulled a share from a contingent clause that only operates if somebody predeceases. Run the sum as a cross-field validation rule, in fractions rather than decimals where possible, because a will that gives one third each will sum to 99.99 in percent and to exactly one in thirds.

Two more gift shapes appear often enough to plan for. A demonstrative legacy is a sum payable from a named source (“$10,000 from my account at…”), which is a specific bequest with a source field rather than a residuary share. And a gift of a percentage of the estate rather than of the residue is neither: it is proportional but it is measured against a different base. If your schema has only two buckets, that third case will be forced into one of them, so at minimum record the stated base as text.

The modifiers that travel with a beneficiary

A beneficiary name on its own is an incomplete extraction, because wills attach rules to gifts that decide what happens when reality does not match the assumption. The three that appear most:

• Distribution modifiers. Words such as per stirpes and per capita govern who takes a share if the named beneficiary does not survive. They usually sit at the end of the sentence naming the beneficiary, and an extraction that returns name and share drops them entirely. Store the modifier on the share, verbatim.
• Survivorship conditions. “If she survives me by thirty days” is a condition on the gift with a stated period. Both the existence of the condition and the period matter.
• Contingent gifts. An alternative taker named for the case where the primary gift fails. These belong nested under the gift they back up, not as sibling rows, or they will be counted as additional beneficiaries and break the sum check.

The general principle is that anything qualifying a gift stays attached to that gift as structure, and the sentence that created it is kept as a quote. This is the same discipline as the conditional trustee chain in a revocable living trust, and for the same reason: the condition is the content.

Executors, successors and guardians

Fiduciary appointments are ordered, conditional and easy to flatten. The instrument typically nominates an executor or personal representative, then one or more successors who serve only if the prior nominee does not, then possibly a separate trustee for any trust created by the will, and possibly a guardian for minor children. Those are four different roles and the same person can hold more than one.

Model them as role, ordinal position, name, and the verbatim condition under which that person serves. A flat list of names loses the ordering, and ordering is the only thing that distinguishes the person who serves from the person who does not. Adjacent details worth extracting because they are asked about immediately: whether the instrument directs that the fiduciary serve without bond, and whether a powers clause grants specific administrative authority.

Codicils and the document set problem

This is the failure mode that no amount of care on a single document prevents. A codicil is a separate instrument that amends specified articles of a will and leaves the rest standing. Extract the will alone and every field is correct as of the day the will was signed and wrong as of today.

So the unit of extraction is the document set, not the document. Practically that means three things. Detect the instrument type on ingest — a codicil says so in its opening words and refers to the will by date. Record the execution date of every instrument in the set. And represent supersession at article level rather than document level, because a codicil that replaces Article IV leaves Articles I through III and V onward untouched.

instruments:
  will      2019-04-11   articles I–VIII
  codicil-1 2021-09-02   replaces IV; adds IV-A
  codicil-2 2023-06-19   revokes IV-A; amends VI shares

effective article IV   ← codicil-1
effective article IV-A ← revoked by codicil-2
effective article VI   ← will, as amended by codicil-2

A revocation clause in a later will (“I revoke all prior wills and codicils”) is the extreme version and is worth its own flag, because its effect is to make every earlier extracted record in the set historical. Whether a particular instrument is effective is a legal question and not one the pipeline answers; what the pipeline can do is make sure nobody reads an article without seeing that a later instrument mentions it.

Cite the article, always

Wills are organised into numbered articles, usually in Roman numerals, with sub-clauses beneath them. Every extracted provision should carry the article number and the page it was found on, for the same reason a deposition fact needs a page and line reference: without it, checking the record means rereading the whole document, and a record nobody can check is a record nobody trusts. The citation discipline is worked through on building a deposition fact table, and it transfers directly.

Two structural details make article numbering less reliable than it looks. Wills frequently restart or skip numbering after an amendment, producing an Article IV and an Article IV-A that a naive parser normalises into the same key. And running headers on every page often contain the testator’s name and the word “Will”, which reading-order extraction interleaves into the body text at each page break, splitting a sentence and occasionally a numbered clause. Both are ingestion problems rather than prompting problems, and belong upstream with PDF parsing.