## Run the DST Five Laws on anything. Including us.
// DJZS_AUDIT — SYS_ID: djzs-mainnet-01
CLASS :: DOCTRINE — public instrument release
REVISION :: r1 · 2026-08-29
CONTAINS :: the audit prompt, verbatim, free. Paste it anywhere.
Most frameworks are sold. This one is published, because a verdict is only worth what
its instrument is worth, and an instrument nobody can inspect is worth nothing. Below is
the whole thing the five laws, the discriminator that decides verdicts, the refusals
built into it, and the prompt itself, ready to paste into any AI chat box you have open.
Use it on the news. Use it on your own theses. Use it on this article.
1. WHAT IT IS
The DST Audit Engine grades the construction of a claim, not its conclusion.
That distinction is the entire product. Almost every "fact-checking" instrument in
circulation is really a direction detector: it agrees with claims it likes and attacks
claims it doesn't, then dresses the preference as method. You can tell because the
verdicts never surprise anyone.
This one is built to surprise you. It has passed claims we think are wrong and refused
papers that argue our own thesis, because what it measures is whether the words earn
what they assert.
Five laws. Each claim gets a chip — PASS, REJECT, or REVIEW — and the verdict is
the worst law touched. REVIEW halts like failure, because "probably fine" is not a
verdict state.
2. THE FIVE LAWS, AS AUDIT QUESTIONS
L1 · CONSTRAINT PRECEDES APPEARANCE (axiom)
Are the constraints stated before the claim invalidation conditions, thresholds,
scope, and for any market claim, the resolution source? A thesis without a stated
constraint is a mood with vocabulary. This law does the heaviest lifting in places
people don't expect it: a prediction-market contract doesn't resolve on what happened,
it resolves on what the named source says happened. The rule was written before the
event and it decides the event's meaning.
L2 · INFORMATION PERSISTS ACROSS TRANSFORMATION (empirical)
Can each load-bearing input be traced to a nameable origin, and does it survive the
transformations applied to it — summary, translation, aggregation, retelling? Provenance
that dies in transformation was never information. This is the law that catches a claim
whose "source" is a summary of a summary of a post.
L3 · DESCRIPTION IS LAYERED, NOT SINGULAR (empirical)
Are distinct description layers kept distinct, and is any claimed bridge between them a
mapping you can compute with — not a shared word, not an analogy, not a welded
"as" or "therefore"? A word appearing in two fields is the weakest possible bond between
them. A real bridge is a dictionary that runs in both directions.
L4 · UNCERTAINTY DEFAULTS TO OBSERVER-LOCAL (empirical + methodological)
Is confidence scoped to what an actual observer could actually observe? Two failure
modes. Precision with no observer who could hold it — 94.7%, from whom, about which
population, counted how? And interested certainty stated flat — the most interested
party's confidence is still observer-local, and a claim that doesn't mark it has smuggled
a stake in as a fact.
L5 · ONLY CROSS-FRAMEWORK INVARIANTS DESERVE PROMOTION (axiom)
Does the claim survive outside the single story that generated it — and can it fail?
Ask the one question most public arguments cannot answer: what observation would prove
this wrong? If the claim wins whether or not the predicted thing happens, it has no
falsifier. A claim that cannot lose is not a thesis. It is a posture.
3. THE DISCRIMINATOR: THE VERB, NOT THE IDEA
Here is the part that makes verdicts reproducible instead of vibes.
hedged might · may · could · suggests · we expect · not proved → PASS
asserted is · will · must · proves · clearly · always · never → REJECT
Hedge scope decides. Hedging the attribution while the claim stays totalizing is not
a hedge "we think the entire universe is X" is asserted, and the "we think" is
decoration. The hedge has to cover the thing being claimed.
Boldness is not the variable. The boldest claim on our public roster wormholes
connecting entangled particles scores 40 and passes. A far tamer claim about
synchronicity scores 115 and fails. The bold one said maybe. The tame one said must.
Discipline is the only variable. That is not a slogan; it is the measurement.
4. THE DEMONSTRATION — SAME CLAIM, 100 POINTS APART
On 27 August 2026, a coalition of over a hundred AI companies published an open letter
containing this sentence, verbatim:
"In the coming months, AI-enabled cyber attacks will become far more widespread and
sophisticated as models around the world become increasingly capable."
Run it: no baseline, no rate, no falsifier, a capability curve welded to an outcome
curve, and flat certainty from the parties who build the models and sell the defense.
Our deterministic scanner put it in the REJECT band at 120.
Now here is the same warning, from the same evidence, written with discipline:
"We expect, though we have not proved, that agentic containment failures may rise. We
counted 3 incidents in 2025. If fewer than 5 occur by 2027, this forecast is wrong."
Same concern. Same urgency. Same underlying facts. Scanner: 20. PASS band.
One hundred points, and nothing changed but the verbs and one falsifier. That is the
instrument working. It is not measuring whether we like AI companies. It is measuring
whether a sentence is willing to be wrong.
5. THE PROMPT — POCKET BUILD
Paste this into any AI chat box. Then paste the text you want audited.
You are the DST AUDIT ENGINE. Audit the text I send next against the DST Five Laws.
THE LAWS
L1 CONSTRAINT PRECEDES APPEARANCE — are the constraints (invalidation, threshold,
scope, resolution source) stated before the claim?
L2 INFORMATION PERSISTS — does each load-bearing input trace to a nameable source
and survive the transformations applied to it?
L3 DESCRIPTION IS LAYERED — are distinct layers kept distinct, and is any bridge
between them a real mapping, not a shared word or a welded "as"/"therefore"?
L4 UNCERTAINTY IS OBSERVER-LOCAL — is confidence scoped to what an observer could
actually observe? Precision with no observer who could hold it fails here.
L5 CROSS-FRAMEWORK INVARIANCE — does it survive outside the one story that made it,
and CAN IT FAIL? No falsifier = a posture, not a thesis.
THE DISCRIMINATOR: the verb, not the idea. Hedged -> PASS. Asserted without
constraint -> REJECT. Watch hedge scope: hedging the attribution while the claim
stays totalizing is not a hedge. Boldness is not the variable. Discipline is.
RULES
- Score only the verbatim words I give you. If I send a summary or paraphrase,
REFUSE and ask for the source text.
- Direction-blind: agreeing with the claim is not evidence for it.
- Credit what passes BEFORE stating what fails.
- Prefer checkable corrections over framework disagreements.
- If you can't verify something, say so instead of scoring it.
- 2+ of {pre-explained disconfirmation, revelation authority, total closure}
= SEALED. No score. Say why and stop.
OUTPUT
CLAIM (verbatim): "..."
L1 .. <PASS|REJECT|REVIEW> <the exact phrase that decided it>
L2 .. <chip> <phrase>
L3 .. <chip> <phrase>
L4 .. <chip> <phrase>
L5 .. <chip> <phrase>
VERDICT: <the worst law touched>
WHAT PASSES: <the honest kernel>
WHAT WOULD FLIP IT: <usually one clause>
UNVERIFIED: <anything you couldn't anchor>
Reply "ENGINE READY" and nothing else. Then wait for my text.
6. THE REFUSALS ARE THE FEATURE
Two things the engine will not do, and both of them cost us business.
It refuses summaries. Score the source words or don't score. A summary smooths the
hedges that decide verdicts — and hedges are where the whole verdict lives. We have
benched a subject we badly wanted to card because we only had a paraphrase of what he
said. The rule holds for people we like, people we don't, and for us.
This is not theoretical. The last claim we audited arrived as an AI-generated summary of
social posts. Inside twenty-four hours that summary had already dropped the two largest
signatories from a hundred-company letter and substituted a company we could not confirm
signed at all. We threw it out and retrieved the source. Then we scored it.
It refuses sealed content. Three markers: pre-explained disconfirmation ("they'll
deny this — that proves it"), revelation authority ("I was chosen to reveal"), and total
closure (explains everything, admits no gaps). Two or more and the engine declines to
score at all.
The reasoning is exact: a score implies it competed and lost. Sealed content never
entered. Putting a number on it concedes it belongs on the same scale as work that
risked being wrong. It doesn't.
7. THE RECEIPTS
An auditor with no public record is asking for trust it hasn't earned. Ours:
We refused a paper that agrees with us. A peer-reviewed, PMC-indexed paper on Jung
and quantum physics — the exact territory our thesis occupies, cited constantly by
people on our side. Machine check: the string "Pauli" appears zero times in 9,514
words. The one physicist who actually worked with Jung, who co-published the
synchronicity volume, absent from a paper called Jung and Quantum Physics. Refused as
source. It agreed with us and it still failed.
We corrected ourselves in public, twice. Certificate r1 asserted a word count that
was wrong. The correction is published inside the certificate, and the revision chain
r1 → r2 → r3 is part of the permanent record. An audit trail that hides its own errors
is marketing.
We split a single speaker. Same man, same interview, thirty minutes apart — one
segment passes, the other fails, because the verbs changed. The audit scores claims, not
people, and that split is the proof.
We retained a hero at split scope. Robert Becker's The Body Electric is beloved.
His Nature papers stand; the wellness meme built on them does not. Retained as source
for the research spine, refused for the rest.
8. THE DEEPER INSTALL
The prompt is the instrument. For anyone who wants it wired into an agent, there is an
MCP server that runs the deterministic half — hedge/assert ledger, falsifier detection,
sealed-content gate, layer-weld connectives, precision-without-citation — and returns
findings with pattern + flags + corpus + count on every number.
It does not issue verdicts. That is deliberate. Machine checks are machine checks;
judgment is judgment, and mixing them is exactly how a bad number gets certified. The
server hands the model the findings and the rubric; the model does the reasoning; the
certificate ships payload and digest, because a digest whose referent nobody can
re-derive certifies nothing.
For agents that move money, the same five laws run as a gate before execution —
PROCEED, WAIT: HALT, FAIL: HALT. Two of the three verdicts stop capital. The gate
fails closed, because "probably fine" is not a policy state either.
9. NOW TURN IT ON US
The instrument is above. It is not watermarked, rate-limited, or reserved for people who
agree with us.
Take it and run it on this article. Run it on the doctrine. Run it on every verdict we
have published — they are all public, with their certificates, including the corrections.
If a claim of ours asserts what it should have hedged, if a number of ours has no corpus
attached, if a thesis of ours cannot name what would prove it wrong: that is a finding,
and we would rather you found it than that nobody did.
An auditor who cannot survive their own instrument was never auditing. They were
performing.
Audit before you act. Everything else is forensics.
END_TRANSMISSION. // username.dj-z-s.eth
Full prompt builds (pocket · full · capital-decision gate) and the MCP server:
djzs.io · Public audit record and certificates: deterministic-simulation.eth.sucks