Portable Data Architecture×Agentic build research

Make platform choice reversible.
Then prove the contract can carry it.

Contractone page
Blind builds16 readings of it
Rulingsowner decides
Bindingsany platform
1 row
changed by swapping Postgres for DuckDB
1,470 rows
changed by reading the same contract another way
60% → 0%
build disagreement after 15 rulings (Koha port)
01The paper

The contract is the product. Platforms are adapters.

Consumers bind to the contract §2–4, §6.2

business · policy · service · evidence · economics

Exit cost lives in the consumers §9.3

40 consumers × £15k = £600k exposure · £50k/yr overhead · paper's worked example

The experiment §10.1

1Baseline on binding 1
2Swap the binding, consumers unchanged
3Change a consumer
4Negative tests
5Evolve the contract
6Service tests
7Exit drill: rebuild on a third platform

Where the paper asks for evidence §3.3 · §4.2 · §10.1

“The differences define the portability ceiling, which is the most valuable finding.”

Consumers come to depend on behaviour the contract never declared.

AGENTIC COVERAGE: UNTESTED
02A small run of §10.1Measured

Swap the engine: 1 row. Read the contract another way: 1,470.

The paper's own §4.1 product, Customer Transaction History. Six outputs, 300 synthetic customers, 4,754 transactions, Postgres 16 and DuckDB. Each square is one output row.

1of 1,448 rowsPostgres → DuckDB, SQL unchanged
Customer 142's average: 292.78 vs 292.77. DuckDB averages decimals in floating point.
1,470of 1,963 rowsThe same contract, read 16 defensible ways
75% of rows depend on at least one decision the contract leaves open.
unchanged changed by the platform changed by a reading of the contract
Self-noise check first: binding 1 against itself, 0 rows differ. Readings enumerated by hand; ranges show the largest alternative.
03Three instruments · R. Houhou

Turn the paper's untested steps into evidence.

1

Build it more than once

Agent non-determinism as the sensor. Where blind builds split, the contract is silent. Where one stands alone, that build has a bug.

§4.2§5.3§5.4§10.1Measured on code
all agree: settled split: a question for the owner one alone: a bug
60.3%
of 12,008 inputs disputed by 8 builds, 2 labs
15
questions behind every split
15
rulings by a stand-in owner
0%
disputed across 7 fresh builds
Koha CalcFine, 82 lines of Perl. 2 builds from 2 labs surfaced 7 of 17 known gaps; 4 builds from one lab, 3.
2

Read what the estate already decided

Mine PRs, reviews, incidents and query logs for implied terms and the dependencies consumers never declared.

§3.3§7.2§9.3Hypothesis
merged & rejected PRsreview commentsincident fixesconsumer query logstransformation code
→
Implied contract termse.g. every consumer filters to posted rows
Hidden consumer dependenciese.g. reports that rely on row order
Questions nobody has asked yetsent to the owner as rulings
Nearest evidence · 3,000 replayed sessions, old app vs blind rebuild
different response
128
different database
2,048
Most behaviour lives below the interface people look at. Examples above are illustrative.
3

Hot-swappable standards

A catalogue of how engines differ on the same SQL, and an agent that grows a new binding until it reconciles with the live one.

§2.3§9.3§10.1 test 7Partly measured
BehaviourPostgresDuckDBSQLiteSnowflakeOracleSpark
reference measured difference set by server config: 149 days moved to catalogue
29,935 / 29,935
replayed requests matched by a blind Flask → Node rebuild, plus every database state, after 3 rulings and 1 fix
04Every difference gets one ruling

Three outcomes, all already in the paper.

Contract term
Reading A
Reading B
output rows move between the two readings
Proposed class shown for each case. Rulings belong to the owner and governance (§5.4).
05The stunt

Delete the binding. Grow it back from the contract alone.

If agents that never saw the product can rebuild it on another platform and it reconciles with the live one, the contract really is the product.

Live bindingkept sealed as the reference
Contract + casebookthe only input
Blind agents, 3 labsno source, no existing SQL
Third platformblank environment
Replay against livereference queries + generated inputs
What grows backis what the contract really carries
What does notis the portability ceiling, listed row by row
hoursmodel costrows reconciledconsumer changes · target 0open questions raised
06§10.1, extended

Same seven tests. Three new steps. Test 7 goes blind.

NEWMulti-build8 blind builds of binding 1
NEWRulingseach becomes a check
1Baseline+ self-noise first
2Swapagent grows binding 2
NEWConsumer miningpredict who breaks
3Change a consumeras written
4Negative testsas written
5Evolveas written
6Serviceas written
7Exit drillblind, from contract only
new steppaper's test, sharpenedas written
Two weeks

Solo, synthetic

  • ProductCustomer Transaction History (this run)
  • Stepsmulti-build, rulings, swap, blind exit drill
  • Ownerstand-in
  • Costmodel spend, est. tens to low hundreds of £ (not measured)
Together

One real product, sandbox

  • Productparty & legal-entity hierarchy (App. B)
  • Stepsall ten
  • Ownera real one, ruling
  • Bonusthe casebook feeds §10.1's optional AI-consumer test
New for App. D → open questions per contract contract completeness: % a blind rebuild reproduces exit-drill time
07What is shown, what is not
MeasuredA careful one-page data contract leaves most output rows open1,470 of 1,963 rows · synthetic · readings by hand
MeasuredA close engine swap with unchanged SQL changes little1 of 1,448 · Postgres → DuckDB only
MeasuredServer config can change meaning with no contract change149 of 214 daily totals under a UTC server
MeasuredIndependent builds surface spec gaps; rulings make fresh builds converge60.3% → 0% · one 82-line function · 5 of 7 off-grid probes still split
MeasuredAn agent can rebuild a running system blind against a sealed reference29,935 / 29,935 · 365-line web app
MeasuredPeople cannot judge a specification by reading it3 of 10 written laws weaker than intended
HypothesisMulti-build finds data-contract gaps a careful reader missesharness written · not yet run
HypothesisEstate mining recovers implied terms and hidden consumersnot run
HypothesisAgent-built rebinds lower the cost of keeping the option openneeds the blind exit drill
UnknownReal owners rule quickly and consistentlyevery ruling so far by a stand-in