Skip to content

Deal economics

Incomplete dataset valuation: price the repair route, not the gap

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Incomplete dataset valuation should price the repair, not the gap. Sort each missing context element by where it can be recovered, count the distinct objects the repair must fetch, test the buyer's task before and after reconstruction, and credit only repairs drawn from sources licensable on the same terms. In Apache ZooKeeper's 2021-2025 history, 272 of 427 commits (63.7%) held under 20 words of explanation plus a key pointing to the issue tracker.

Value drivers

Incomplete dataset valuation

This page covers incomplete dataset valuation: price the repair route, not the gap.Illustrative

Key takeaways

  • A missing-context discount is three numbers: value as delivered, the credited gain from repairs, and the cost and risk of making them.
  • Repair burden scales with the distinct objects a repair must fetch: 272 tracker-pointer commits in ZooKeeper cite 263 distinct issues.
  • Context can leave a record because of a tooling change; ZooKeeper's self-describing share fell from 41.4% to 11.7% after its December 2023 merge-script change.
  • A repair adds value only if its source can be licensed on the same scope; customer-confidential terms add utility the seller cannot sell.
  • Asking prices and payout ranges are not transactions, so an incomplete dataset is valued on stated assumptions that a reader can inspect.

What is an incomplete dataset valuation actually pricing?

It prices the gap between the records as delivered and the records a buyer can actually use, plus the cost and the risk of closing that gap. Missing context is information a record depends on but does not carry: why a change was made, what a status code means, which policy was in force when someone decided, which contract a ticket refers to. The record exists; its meaning lives somewhere else, or nowhere.

That is why one discount percentage misleads. Two archives can lack the same share of context and deserve very different values, because one seller can recover it from a system it owns and may license, while the other cannot recover it at all. A valuation that an investment committee can check reports three things separately: what the buyer's task scores on the data as delivered, what each repair adds and whether its source can be licensed, and the staff hours and failure risk of each repair.

This page builds that method for investors and strategic buyers. It sorts repair routes, measures repair burden on a public record set, sets out a before-and-after utility test and turns the result into a rights-normalized worksheet with an explicit decision rule. The commit counts are measured; every worksheet value is illustrative.

Where can missing context be recovered from?

From one of five places, and the place sets both the cost of the repair and the ceiling on what it can add to value. Write the route beside every missing element before anyone estimates a discount, because the route decides whether a repair is engineering work, a rights question or new authorship.

Routes are not ranked by convenience alone. A join against a lookup table that ships with the export is cheap and carries the dataset's own rights. A second extraction from a neighboring system can be cheap too, yet it brings in whatever that system holds, which may include vendor, customer or employee content that the first system did not.

The five routes, from most to least certain:

  • Inside the export: a code table, configuration file or earlier record already in the delivered files holds the meaning. The repair is a join, and the rights follow the dataset.
  • An adjacent system the seller owns and may license on the same terms, such as the tracker behind a commit history or the knowledge base behind a ticket queue. The repair is a second extraction, and the rights review has to cover that system too.
  • An adjacent system with different rights, such as customer contracts, vendor-supplied templates or a partner portal. The repair can be technically easy and still add nothing the seller can sell.
  • People: staff who remember why a decision was made. The repair creates material written after the fact, which is annotation rather than an original record and should be labeled as reconstructed.
  • Nowhere: the context was never recorded. The honest treatment is to value the records as delivered, or to narrow the scope to the part that carries its own meaning.

What 427 ZooKeeper commits show about repair burden

Most of the explanation sat in another system, and a tooling change moved more of it there. We counted the 427 non-merge commits on the first-parent history of Apache ZooKeeper's master branch committed from 2021 through 2025, calling a commit self-describing when its message body holds at least 20 explanatory words. Of the rest, those citing an issue key are tracker pointers, those citing only a pull request are pull request pointers, and the remainder are bare. The split: 131 self-describing (30.7%), 272 tracker pointers (63.7%), 15 pull request pointers (3.5%) and 9 bare (2.1%).

The tracker-pointer class is the repair burden. Its 272 commits cite 263 distinct tracker issues, so rebuilding their context means fetching close to one issue per commit. Repair cost scales with the distinct objects a repair must fetch, not with how many records reference them; an archive where hundreds of tickets cite the same policy page would show a far smaller burden for the same share of affected records.

The share also moved with the project's tooling. Before December 1, 2023, ZooKeeper's merge script copied each pull request's description into the commit message; the version committed that day under ZOOKEEPER-4756 builds the message from the pull request's commits, reviewers, author and closing line instead. Of the 273 counted commits before that change, 113 (41.4%) were self-describing, 144 (52.7%) were tracker pointers, 9 were pull request pointers and 7 were bare. Of the 154 from that commit onward, 18 (11.7%) were self-describing, 128 (83.1%) were tracker pointers, 6 were pull request pointers and 2 were bare.

The bar is a choice, so the sensitivity is published: at 10 words 41.7% of all 427 commits are self-describing, at 40 words 22.7%. Six of the nine bare commits are release-plugin housekeeping that needs no explanation. The counts measure where an explanation is written, not whether it is good.

For a private company the adjacent system would be its own tracker, and the valuation question becomes whether that tracker sits inside the license. A public project's tracker is open to anyone; a seller's internal tracker is not, and its contents may carry different rights from the records that point to it.

Two stacked horizontal bars of ZooKeeper master-branch commits from 2021 to 2025: before December 1, 2023, 273 commits with 41.4% self-describing and 52.7% tracker pointers; from that date, 154 commits with 11.7% self-describing and 83.1% tracker pointers, with small pull request pointer and bare segments.
Where each commit's explanation lives, before and after ZooKeeper's December 2023 merge-script change, at a 20-word bar. Source: apache/zookeeper history at commit 858c8d7, counted 2026-10-28.

How do you compare task utility before and after reconstruction?

Run the buyer's task on the same held-out cases three ways: on the records as delivered, on the records with each repair added alone, and on the records with every credited repair added together. Those differences are the utility evidence; anything else is assertion.

Context can move results more than intuition suggests. In a 2025 pretraining study, Gao and colleagues (arXiv 2501.01956) prepended each document's source URL during training and report that a 1.6-billion-parameter model matched standard pretraining's downstream performance while using 33% less data. A URL is not business context, but the result shows the same text doing more work per token when its context travels with it.

Two guardrails keep the test honest. Show a repaired field only where it would have existed when the original decision was made, because a policy page edited after a ticket closed can leak the answer. And ask why the context is missing: Rubin's 1976 paper in Biometrika set out when the process that produced missing values can be ignored, with data missing at random as the central condition. Context that went missing because a case was urgent, messy or escalated is not missing at random, so a repair that fills only the easy gaps will flatter the score.

SourceX has published a context-ablation protocol, with leakage controls and success measures, as a study design without results; a valuation can borrow its arms instead of designing new ones.

The rights-normalized repair worksheet, one missing element per row

The worksheet credits a repair only when three tests pass in order: its source can be licensed on the proposed scope, the rights in that source are settled, and its staff hours per point of measured gain stay under a ceiling the underwriter states before seeing results.

Illustrative: a fictional software company offers 40,000 change records. On the buyer's held-out task the records score 52 of 100 as delivered, and five context elements are missing. Each gain was measured alone against that baseline; repair hours are setup time plus distinct objects times minutes per object; the ceiling is 60 staff hours per credited point. Setup and minutes per object: tracker issues 40 h and 0.5 min; code definitions 6 h and 15 min; contracts 20 h and 20 min; policy versions 12 h and 30 min; reconstruction 16 h and 6 min per record. Every row was computed by script from these stated inputs.

Two repairs earn credit, for 253.6 staff hours in total. The contract-terms row is the instructive one: at 31.3 hours per point it clears the cost test with room to spare and still earns nothing, because customer-confidential terms cannot be licensed with the records. The unwritten-reason row fails the other way, with clean rights but 136.0 hours per point. Credited gains reach 65 points only if they add up; the joint run sets the number that enters the valuation.

The worksheet, one row per missing element:

The rights-normalized repair worksheet, one missing element per row
Missing element (records affected)Repair route and distinct objectsRepair hours, gain, hours per pointRights in the repair sourceWorksheet treatment
Rationale in linked tracker issues (25,600; 64.0%)Adjacent system, same owner; 23,800 issues238.3 h; +9 points; 26.5 h per pointLicensable on the proposed scopeCredit after a joint test
Status and reason codes without a dictionary (40,000; 100.0%)Inside the export; 37 code definitions15.3 h; +4 points; 3.8 h per pointLicensable on the proposed scopeCredit after a joint test
Customer contract terms the record cites (6,000; 15.0%)Adjacent system, customer-confidential; 410 contracts156.7 h; +5 points; 31.3 h per pointNot licensable with the recordsNo credit: value as delivered
Policy version in force at decision time (12,000; 30.0%)Document system holding vendor template text; 58 versions41.0 h; +3 points; 13.7 h per pointUnsettledHold credit until rights settle
Reason never written down (8,000; 20.0%)Staff reconstruction; 8,000 records816.0 h; +6 points; 136.0 h per pointLicensable, but newly written materialDo not repair: value as delivered

Why asking prices and payout ranges cannot set the discount

Because neither is a transaction, and a dataset with missing context rarely has a transaction to compare against. The tempting shortcut takes a published asking price or payout range for complete data and subtracts a haircut, which borrows a number that never cleared a market and applies a discount nobody measured.

Accounting standards draw the same line. IFRS 13 defines fair value as the price received to sell an asset in an orderly transaction between market participants at the measurement date, tells valuers to maximise observable inputs, and ranks quoted prices for identical assets in active markets highest (Level 1) and unobservable inputs lowest (Level 3). Nothing quotes incomplete business datasets in an active market, so a valuation of one rests on Level 3 inputs: the repair routes, gains and hours are the valuation, and each should be visible to the reader.

Disclosed deals help only after normalization. Before setting one beside an incomplete dataset, adjust it to the same rights, the same usable scope after repair, the same preparation burden and the same buyer purpose; a complete archive licensed to train an agent is a different asset from an incomplete one licensed for evaluation. Present what survives as a range with its assumptions, not as a point estimate.

What could make a repair-based valuation wrong

The method fails when a repair imports rights problems, when gains are double counted, or when the repairs fill only the easy gaps. Each risk below has a check an investor can ask to see.

At a glance

What could make a repair-based valuation wrong

The risks, with the check for each:

  • Repair sources with different rights: a second system can hold vendor, customer or employee content that the first did not, so a repair that imports it can shrink the licensable scope instead of growing it. Check the rights review of every repair source.
  • Gains that do not add: a tracker issue may already state the policy a second repair would restore. Credit the joint run, never the sum of single-repair gains.
  • Biased repairs: when the hardest cases are the ones missing context, filling the easy gaps raises scores where the buyer needed help least. Compare which cases each repair reached.
  • Compounding defects: in interviews with 53 AI practitioners, Sambasivan and colleagues (CHI 2021) found data cascades, compounding downstream effects of data issues, with 92% prevalence; a faulty join added during repair can travel the same way. Check repaired fields against a hand-verified sample.
  • Era shifts: a tooling or policy change can move context out of the record partway through an archive, as ZooKeeper's merge script did in December 2023. Sample both sides of every known change.
  • Buyer purpose: context that lifts a training run may not matter for an evaluation set, and the reverse. A gain measured on one buyer's task does not transfer to another's without a new test.
  • Leakage: a repaired field written after the outcome inflates the score and fails in deployment. Ask for the timestamp test applied to every repaired field.

Diligence questions for incomplete records, SourceX included

Ask for the evidence behind each worksheet input, not for a conclusion. Any company that values, buys or sells incomplete business records should be able to answer these questions with documents.

SourceX describes its own process on its site. A Dataset Deal Confirmation fixes the exact dataset, the payout, the rights, the processing route and the deadlines before SourceX presents the opportunity externally, and an earlier estimate is indicative only; not every qualifying company receives an offer. SourceX shares company and commercial metrics with investors directly rather than publishing them. It is raising from accredited investors under Rule 506(c), and investors can request materials through the investor page with a specific question, such as how a seller's repair routes were checked for rights.

Questions for any data company, SourceX included:

  • Which context elements are missing from each record family, and what share of records does each affect in each era?
  • For each element, which repair route applies, and how many distinct objects must the repair fetch?
  • Which repair sources sit inside the licensed scope, and who signed off on their rights?
  • What did the buyer's task score before and after each repair on held-out cases, alone and in a joint run?
  • Which repaired fields could not have existed at decision time, and how were they kept out of the test?
  • Which numbers in the valuation are transactions, which are asking prices or estimates, and which are assumptions?

Methods note: the ZooKeeper census and the worksheet model

The census reads Apache ZooKeeper's public repository (github.com/apache/zookeeper), cloned as commit metadata only on 2026-10-28 at commit 858c8d77, committed on 2026-10-26. It walks the first-parent history of that commit, keeps commits whose committer timestamp, converted to UTC, falls from 2021-01-01 to 2025-12-31, and drops merge commits: 430 commits fell in the window, 3 were merges and 427 were counted.

A tracker key is any case-insensitive match, anywhere in the subject or body, of 'zookeeper' followed by an optional hyphen, underscore or space and then digits, not preceded by a letter or digit and not followed by a further digit or by a decimal point and a digit, so 'Zookeeper 4756' counts and 'ZooKeeper 3.9' does not. Key numbers are compared without leading zeros, and a leading byte-order mark, if present, is removed before matching. A pull request reference is '(#n)', 'closes #n' or '/pull/n', searched the same way. Explanatory words are counted in the message body (every line after the subject) after dropping blank lines, tokens beginning http:// or https://, and trailer lines beginning with 'closes #n' or with one of these words followed by a colon: author, reviewer, reviewers, reviewed-by, signed-off-by, co-authored-by, change-id, committer or cc; a word is a whitespace-separated token containing a letter. A commit is self-describing at 20 or more such words (10 and 40 reported as sensitivity); otherwise it is a tracker pointer if it has a key, a pull request pointer if it has only a pull request reference, and bare if it has neither. Percentages are rounded half up to one decimal place from exact counts. Distinct issues are the distinct key numbers cited by tracker-pointer commits, and release-plugin housekeeping means a subject line starting with '[maven-release-plugin]'.

The era split puts the ZOOKEEPER-4756 commit (32fb89c9, December 1, 2023) and every later commit in the second era; the merge script was read at that commit's parent and at the pinned commit. Not measured: whether the tracker issues contain the missing rationale, the quality of any message, release-branch commits, or the effect of any repair on any task. The census script and its outputs, which hold counts and per-commit classes without names or message text, are kept for recount.

The worksheet is an editorial framework built for this article on 2026-10-28. Its inputs are invented to show the arithmetic and describe no real company or deal; a separate script computes every row, with final values rounded half up to one decimal place.

Frequently asked questions

How do you value a dataset with missing context?

Value it as delivered first, then add only the repairs you can test and license. List each missing element, its repair route and the distinct objects the repair must fetch, measure the buyer's task before and after, and credit the joint gain from repairs whose sources sit inside the licensed scope. Report the three numbers separately rather than one discount.

Is missing context the same as missing data?

No. Missing data is an empty field; missing context is meaning a record depends on but does not carry, such as the reason for a change or the policy in force at the time. A record with every field filled can still lack context, and that context may sit in another system with its own rights.

Can a public payout range or asking price set the discount?

Not reliably. Neither is a cleared transaction, and IFRS 13 ranks unobservable inputs lowest in its fair value hierarchy. A dataset with missing context has no quoted price, so its value rests on stated assumptions about repair cost and gain, which should be shown rather than replaced by a borrowed number.

Do staff interviews that rebuild the missing reasons add value?

They can, but the result is annotation written after the fact, not an original record. Label it as reconstructed, check that it cannot leak outcomes, and price it separately when its cost per point of gain is high; in the illustrative worksheet it fails at 136.0 hours per point.

How should a seller prepare an incomplete archive for valuation?

Inventory the missing elements by record family and era, map each to a repair route, and confirm which repair sources the company may license. Sample both sides of any tooling or policy change, since context can leave the record partway through an archive, as it did in ZooKeeper's history in December 2023.

Sources

  • Apache ZooKeeper's public commit history at commit 858c8d77 supplies the 427 non-merge master-branch commits from 2021 to 2025 counted in this article. Source
  • ZooKeeper's current merge script builds the squash-merge message from the pull request's commit messages (when the merger opts to include them), approving reviewers, author and a closing line. Source
  • The ZOOKEEPER-4756 commit of December 1, 2023 switched the merge script to the GitHub merge API; the script's previous version added the pull request description to the commit message. Source
  • Prepending source URLs during pretraining let a 1.6B-parameter model match standard pretraining's downstream performance with 33% less data (Gao et al., 2025). Source
  • Rubin (Biometrika, 1976) set out when the process that causes missing data can be ignored, with data missing at random as the key condition. Source
  • Interviews with 53 AI practitioners found data cascades, compounding downstream effects of data issues, with 92% prevalence (CHI 2021). Source
  • IFRS 13 defines fair value as an exit price in an orderly transaction between market participants, prioritises observable inputs and ranks unobservable Level 3 inputs lowest. Source
  • SourceX's Dataset Deal Confirmation fixes the exact dataset, payout, rights, processing route and deadlines before the opportunity is presented externally. Source
  • SourceX describes estimates as indicative and says not every qualifying company receives an offer. Source
  • SourceX shares company and commercial metrics with investors directly and takes investor materials requests on its investor page. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies