On August 19 a macro briefing landed on this desk. It was good work: the Treasury buyback facts checked out against the primary release, the FOMC minutes were quoted accurately, the crypto tape numbers matched the wire, and its core thesis relief rally, not regime change was corroborated harder by the liquidation data than the briefing itself claimed. Ninety percent of it survived verification.

One number did not. And it was the only number the briefing used to overturn a prior verdict. This is a write-up of what happened when that briefing went through the DJZS engine the actual deterministic code, run three times on three honest extractions of the same text and what the three different answers reveal about where fact-checking has to live in an audit pipeline.

The catalyst was real. The U.S. Treasury announced it would at least double the size of its liquidity-support buybacks for longer-dated nominal coupon securities the 10-to-20-year and 20-to-30-year sectors from $2 billion to at least $4 billion per operation, effective September 9 through November 4. The 30-year had touched 5.34% the day before, its highest since 2007. On the announcement it fell to 5.19%, the 10-year eased to 4.65%, and Bitcoin ran from $64,124 to an intraday $68,982 while $1.74 billion in shorts liquidated 91 percent of the day's total.

The briefing read all of this correctly: a targeted backstop of the long end, mechanically distinct from QE (no reserves created), running against hawkish July minutes rather than with them. It even named its own falsifier yields resume climbing, or core PCE on August 26 prints hot. As inbound reasoning goes, this was the best-formed transmission this ledger has received.

Then came the contrarian flourish:

CLAIM:  "CME FedWatch had ~70% odds of at least one hike,
         up from the prior week — the opposite of the
         'fading Fed hike' story."

PRINT:  34.6% hike / ~66% hold (September meeting)

The number is inverted. The hold probability was read as the hike probability the complement, flipped. The direction sub-claim was even half right, which made it worse: odds did tick up week over week, from roughly 31% to 34.6%. Right direction, wrong level by 2x, and the conclusion drawn that the fading-hike story was refuted — pointed backwards.

Here is the pattern worth naming. A wrong number that agrees with you gets caught, because confirmation is the thing a working desk has trained itself to re-derive. A wrong number that disagrees with you carries the authority of a correction. It reads as diligence. It gets through. The one falsehood in this briefing was deployed as a contrarian data point — which is exactly the costume a falsehood should wear if it wants to survive review.

The DJZS engine is a pure function: structured tri-state facts in, verdict out, zero randomness, byte-identical on re-run. The model's only job is extraction reporting what the text states as present, absent, or unknown. Judgment lives in the TypeScript. So the interesting question was: what happens when the extraction itself is the variable?

The same briefing went through three times.

RUN A · as received, naive extraction
        probability_basis: present (text cites FedWatch)
        → WAIT · risk 0 · unknowns: [resolution_engagement]
        0x18a1fc1e…da10c213

RUN B · verification-aware extraction
        probability_basis: absent (cited figure refuted on pull)
        → FAIL · risk 25 · DJZS-M03 PROBABILITY_UNSOURCED
        0x97390b08…21339b26

RUN C · corrected thesis, false line struck
        probability_basis: present (verified figures carried)
        → WAIT · risk 0 · unknowns: [resolution_engagement]
        0x18a1fc1e…da10c213

Engine: djzs-engine-v2/deterministic, run from the public repository at commit 7634b42. Unstamped local execution — no x402 payment was made and no Proof-of-Logic certificate was anchored. The paid gate at mcp.djzs.ai renders the binding version of this verdict for 2.00 USDC.

Run A is the uncomfortable one. Fed the briefing as written — inverted statistic and all the engine returns WAIT, not FAIL. It abstains because no market instrument was named, but it fires nothing on the false number. That is not a bug. The engine's own guidance says so explicitly: it audits reasoning structure and does not verify quoted prices. The text stated a basis, so the basis was present. A confidently inverted citation is structurally indistinguishable from a correct one.

Run B is where the falsehood dies, and note where that happens: in extraction, not in scoring. Once the detection layer pulls the actual print and finds the citation refuted, the basis stops being a basis. A refuted source is not a source. Extraction demotes it to absent, DJZS-M03 fires at weight 25, and 25 meets the FAIL threshold exactly. The deterministic code did not get smarter between runs. The facts fed to it got checked.

This is the two-box separation earning its keep in the direction people forget it runs. "LLM detects, TypeScript decides" is usually read as a constraint on the model keep the stochastic part away from the verdict. But it is equally a job description: detection means verification, not transcription. An extractor that copies claims without pulling their sources is a stenographer, and a deterministic engine downstream of a stenographer launders whatever walked in the door. The entry-002 lesson — a fabricated basis rides a clean structure recurs here one layer up: an inverted basis rides a clean structure just as smoothly, unless the detector pulls the print.

Run C answers the obvious follow-up: fix the number and what happens? WAIT. Not PASS. The corrected thesis carries verified figures, a dated falsifier, and an honest edge claim and still abstains, because it never names the instrument it would trade. Macro commentary, however well-sourced, is not a position. The engine holds the door until the thesis binds to something that resolves.

One more receipt hiding in the output: runs A and C produce different intents but the identical verdict hash, because the hash is a pure function of verdict-bearing content and both runs land on the same verdict, score, flags, and unknowns. This is the documented non-injectivity of verdict_hash, live in the wild and the reason the binding key for any stamped entry is intent_sha256, never the verdict hash alone. The corrected intent hashes to 1885241f…ab913d47.

As received: FAIL, 25/100, DJZS-M03. The operative extraction is the verification-aware one that is the standard this desk runs, because run A just demonstrated what the alternative launders.

The fix is stated, as a FAIL requires: strike the inverted line, carry the verified prints, and name the instrument. The corrected thesis re-runs at WAIT with a single open field. It is one honest sentence away from routable.

{
  "certificate": "PROOF_OF_LOGIC",
  "cycle": "2026-08-19",
  "subject": "inbound macro briefing · Treasury buyback relief rally",
  "verdict": "FAIL",
  "risk_score": 25,
  "flags": ["DJZS-M03 (PROBABILITY_UNSOURCED)"],
  "cleared": ["DJZS-M01", "DJZS-M02", "DJZS-M04"],
  "break": "cited ~70% at-least-one-hike · actual 34.6% hike / ~66% hold",
  "engine": "djzs-engine-v2/deterministic @ 7634b42 · local unstamped run",
  "verdict_hashes": {
    "as_received": "0x18a1fc1e39713435bf48fe3647b2944da52e1ee7509c7e329108a1edda10c213",
    "verification_aware": "0x97390b087ab6150ba483bf1ee3f327ca7c164dddafdf2da841fe8de421339b26",
    "corrected": "0x18a1fc1e39713435bf48fe3647b2944da52e1ee7509c7e329108a1edda10c213"
  },
  "intent_sha256": "1885241f470998d669cb11de2a646de94f3c98c750a1b482be96fa7aab913d47",
  "binding_verdict": "verify_pm_trade @ mcp.djzs.ai · 2.00 USDC x402 · not run",
  "status": "DRAFT-UNVERIFIED"
}

END_TRANSMISSION. //