Whitepaper · The Source-of-Truth Kernel

Integrity,
enforced at the floor.

Data your agents can't lie to.

Today's defenses live in application code, off to the side of the data path. A new writer ships without the check; a refactor quietly drops it. Move enforcement to the one doorway every write must cross, and integrity stops being a courtesy. It turns into a property of the system.

100%

rows fingerprinted

0

false rejections

197

corrections captured

1

line to enforce

 Scroll ↓Cuộn ↓ 

01The trust gap

A wrong number means two very different things.

It's 3 a.m. A feed hiccups and writes one corrupted price. Just one bar, off by a single decimal. Nobody is watching. By morning that one row has rippled through every model, signal, and dashboard that read it.

When a human glances at that number, something itches. They stop. When an autonomous agent reads it, there is no itch. It trusts the row, acts instantly, then defends the bad number with a fluent paragraph of reasoning.

A human reading a bad number raises an eyebrow. An agent acts on it. Confidently, fluently, at machine speed.

02State of the art

How the market defends data today, and why it still leaks.

Every defense is real and useful. They also share one fatal property: they live in application code, off to the side of the data path, never across it.

Code-side validation

Collectors & gateways

A new writer ships without the check. A refactor quietly drops it. The validation is a courtesy the writer extends, never a condition of writing.

LLM guardrails

Prompts & filters

Guardrails police the model's behavior. None of them can repair a fact that was already wrong in the source table.

Contracts & ledgers

Schemas & hashes

Most data-contract tooling describes expectations without enforcing them at write time. Side-car ledgers prove a copy, not the row the agent actually reads.

03The pattern

Stop asking for integrity. Enforce it.

Old model · trust by hope

The writer is responsible for writing good data. Validation is optional. Any bypass, whether a bug, a shortcut, or some forgotten wiring, lands bad data and nobody notices.

New model · trust by proof

The database is the choke point and refuses to be lied to. Provenance, rules, and history are conditions of being allowed to write at all.

Enforcement belongs at the single doorway every writer must walk through. Put it upstream and it stays optional. Put it at the floor and it becomes a property of the system.

Cron job

tick · XAUUSD

AI agent

signal write

Backfill

EOD repair · D1

Human console

manual fix

Rogue writer

unregistered source

SoT KERNEL

single trigger · on write

1Provenance

2Obligations

3Tamper-evidence

Source of truth

✓ committed row
run_id  a7f3c2e9
source  OANDA ✓
sha256  7c1e…b40
sealed  14:32:07 UTC

 ✓ registered → sealed & committedđã đăng ký → niêm phong & ghi nhận ✕ unregistered → rejected before it landschưa đăng ký → bị từ chối trước khi hạ cánh 

 Fig. 01 · One enforced doorway. Cron jobs, agents, backfills and humans all funnel through the kernel's three gates into the source of truth. An unregistered writer is turned away at the floor. Hình 01 · Một cánh cửa được thực thi. Cron job, tác nhân, backfill và con người đều dồn qua ba cổng của kernel để vào nguồn sự thật. Writer chưa đăng ký bị chặn ngay tại nền. 

04The architecture

One choke point. Three guarantees.

A thin layer of enforcement living inside the database, on the exact tables that matter. It answers the three questions an agent, or an auditor, might ask of any row.

01

Provenance: where did this come from?

A canonical registry of approved sources. Every row declares its origin and carries a run-correlation ID, so any number can be traced back to the exact process that produced it.

02

Obligations: does it obey the rules?

A declarative rule registry, enforced by a single trigger on every write. Each rule is a row, not a code branch: soft logs, hard rejects. Promotion is a one-line update. Reversible, and no deploy.

03

Tamper-evidence: can change be proven?

Every row carries a cryptographic fingerprint over its economic content, in a generated column the writer cannot set. Prior versions are captured into an append-only, immutable history. "It used to say X, and here is when it changed" becomes a query.

How a write earns its placeCách một lượt ghi giành chỗ đứng

INSERT /
UPDATE

1 row

01

Provenance

source ∈ registry?

+ run_id stamped

✕ unregistered → reject

02

Obligations

soft·loghard·reject

✕ hard fail → reject

03

Tamper-evidence

sha256 → gen column

prior version → history

✓ sealed & logged

✓

Committed

provenance attached

 Fig. 02 · A single trigger. A write clears all three gates, or it never lands. Hình 02 · Một trigger duy nhất. Một lượt ghi vượt qua cả ba cổng, hoặc không bao giờ hạ cánh. 

Identity travels with the row

Enforcement reads "who am I" from the row itself, never from session state. Poolers reuse connections across transactions, so provenance has to be carried by the data, not by the connection.

When a value changes, nothing gets erasedKhi một giá trị thay đổi, không gì bị xóa bỏ

Current row · xauusd_bars

ts     2026-03-29 · D1
close  2,362.40 → 2,362.55
fp     sha256 7c1e…b40

generated column · writer can't setcột sinh tự động · writer không đặt được

UPDATE received · price correctionNhận UPDATE · sửa giá

prior version
captured firstphiên bản cũ
được lưu trước

before change landstrước khi thay đổi xảy ra

Append-only history UPDATE / DELETE ✕

v3 · now2,362.559f3a…c12
v2 · 14:322,362.407c1e…b40
v1 · 09:152,361.10a93f…2dc

 Fig. 03 · A correction never overwrites the past. The prior version is sealed into an append-only history before the new value lands, and the trail itself rejects UPDATE and DELETE. Hình 03 · Một lần sửa không bao giờ ghi đè quá khứ. Phiên bản cũ được niêm phong vào lịch sử chỉ ghi thêm trước khi giá trị mới xuất hiện, và bản thân nhật ký từ chối UPDATE và DELETE. 

05  The doctrineHọc thuyết

Integrity is not correctness.

The kernel proves a number is what it claims to be and hasn't been silently altered. It does not promise the number is right. A perfectly registered, untampered source can still send a wrong price.

The machine verifies the data.
The human owns the decision.

06In production

What changed in practice.

Before → after

Bad sources became impossible, not improbable.

A misconfigured writer no longer quietly poisons the table. It gets rejected at the write, where the failure is cheap to spot.

Before → after

"Did this change?" became a query.

The old answer was archaeology. Now one query returns every prior version and the moment it was superseded. Corrections became visible instead of feared.

Before → after

Agents got something to cite.

Read paths surface each row's source and fingerprint. "The data is real" became a fact an agent can point to, not an assumption.

Before → after

The blast radius of a bad deploy shrank.

Enforcement lives in the database, so a regression in a writer cannot disable it. The write path runs as a low-privilege role that can't drop a table or forge history.

07The evidence

Proof, not promises.

Validated end to end against a live deployment. We tested it on a throwaway copy first, both before and after we went fully enforcing.

100%

rows carry a verifiable fingerprint

0

rejections on legitimate traffic

197

real corrections, captured & queryable

1

line to flip shadow → enforce

✓ Verified

The fingerprint holds independently.

An external recomputation in a separate language reproduces the database's hash byte for byte. Verification doesn't depend on trusting the database that stores it.

✓ Verified

The history can't be rewritten.

The audit trail rejects updates and deletes outright. You can append, but you can never edit. Even the privileged path can't quietly revise the record.

08Honest scope

What we deliberately didn't do.

By design

Integrity, not correctness

A registered source can still be wrong. We prove origin and tamper-evidence. A human still owns whether it's right.

On our schedule

Not every rule is hard yet

Source enforcement is hard. Price-sanity rules only log for now. Promotion is one line, flipped whenever we choose.

Scoped

We didn't boil the ocean

We hardened the highest-value tables first. Others have different shapes and earn their own design later, instead of a rushed one now.

Nước chảy đá mòn.

The takeaway

Build the trust into the floor.

You can't prompt your way out of bad data, and you can't validate your way out of it from code any deploy can regress. A source of truth that can't be lied to, because the lie is rejected before it lands.

Get in touch →

Built on plain PostgreSQL. No new infrastructure.

Integrity,enforced at the floor.

Tính toàn vẹn,thực thi tại nền dữ liệu.

A wrong number means two very different things.

Một con số sai mang hai ý nghĩa rất khác nhau.

How the market defends data today, and why it still leaks.

Cách thị trường bảo vệ dữ liệu hôm nay, và vì sao nó vẫn rò rỉ.

Collectors & gateways

Collector & gateway

Prompts & filters

Prompt & bộ lọc

Schemas & hashes

Schema & hash

Stop asking for integrity. Enforce it.

Đừng cầu xin sự toàn vẹn. Hãy thực thi nó.

One choke point. Three guarantees.

Một điểm nghẽn. Ba bảo chứng.

Provenance: where did this come from?

Nguồn gốc: dữ liệu này từ đâu ra?

Obligations: does it obey the rules?

Nghĩa vụ: nó có tuân luật không?

Tamper-evidence: can change be proven?

Chống giả mạo: thay đổi có chứng minh được không?

Integrity is not correctness.

Toàn vẹn không phải là đúng.

What changed in practice.

Điều gì đã thay đổi trong thực tế.

Bad sources became impossible, not improbable.

Nguồn xấu thành bất khả, không còn là khó xảy ra.

"Did this change?" became a query.

"Cái này có đổi không?" thành một câu truy vấn.

Agents got something to cite.

Tác nhân có thứ để trích dẫn.

The blast radius of a bad deploy shrank.

Bán kính thiệt hại của một bản deploy hỏng thu nhỏ.

Proof, not promises.

Bằng chứng, không phải lời hứa.

The fingerprint holds independently.

Vân tay đứng vững độc lập.

The history can't be rewritten.

Lịch sử không thể bị viết lại.

What we deliberately didn't do.

Những điều chúng tôi cố ý không làm.

Integrity, not correctness

Toàn vẹn, không phải đúng

Not every rule is hard yet

Chưa phải luật nào cũng hard

We didn't boil the ocean

Chúng tôi không ôm đồm tất cả

Build the trust into the floor.

Hãy xây niềm tin vào chính nền móng.

Integrity,
enforced at the floor.

Tính toàn vẹn,
thực thi tại nền dữ liệu.