Editorial gates & governance¶
Every brief and every Market Analysis Report passes a chain of gates before it ships, and anything that reaches an on-air channel passes one more. The gates are the enforcement of a single principle: the prose may only assert what the tools' data supports. This page documents each gate and the channel-wide inducement line, and cross-links the published policies they enforce.
Why gates, not trust¶
The deterministic layers of the pipeline — the direction signal, the scorer — produce facts. The LLM layers — the brief, the MAR synthesis — produce prose about those facts, and an LLM writing prose can drift: quote a level that isn't in the data, flip the sign on a read, smuggle in a catalyst it can't see. The gates exist because the writing step has a hallucination surface and the publishing step can't. Each gate narrows what the prose is allowed to claim back down to what the data actually says.
Brief gates¶
Tape-fidelity (challenge / defend / verdict)¶
After a brief is written and its structured data extracted, a tape-fidelity gate fact-checks the prose against that structured data. It runs as a challenge / defend / verdict cycle rather than a batch edit, because batch editing by a reviewer LLM produced far more false positives than real findings — the reviewer has the same hallucination surface as the writer.
Instead, each round: a reviewer raises the single most problematic claim and names the section it's in; the writer's default position is to defend, and the writer has tool access — it re-calls the same structured data reads (levels, patterns, derivatives, order book, developments) and cites the exact values that back the claim, served from the original cycle's cache so the defence is checked against the same snapshot the editor sees. The writer only revises if the tools contradict the claim. An editor then accepts or rejects on the evidence trail. The truth hierarchy is explicit: structured data (the deterministic facts) outranks the research summary, which outranks editorial discipline — and editorial commentary (catalyst speculation, institutional intent, sentiment narrative) is out of scope at the brief layer entirely. If two rounds pass without acceptance, a safety floor strips the unverifiable section before the brief ships — an incomplete brief beats one carrying a claim the writer couldn't defend.
Signal coherence (deterministic)¶
Independently, a deterministic check validates that the brief respects the direction signal: the brief's direction must equal the signal's direction, and the brief's conviction must fall inside the band the signal produced. A non-neutral call must carry a numeric entry, target, and invalidation; a neutral call must not carry an entry. These are hard structural checks — the brief cannot disagree with the locked signal.
MAR gates¶
The Market Analysis Report passes three gates in sequence, the first two deterministic and exhaustive, the third semantic.
Grounding gate (deterministic)¶
Every number of magnitude ≥ 10 in an observation or tension statement must approximately match one of that item's cited metrics — precision-aware, so a value legitimately rounded to its printed decimals passes while an invented value fails. An ungrounded number in a statement drops the whole item; an ungrounded number in a valence pole (the opportunity/threat line) nulls just that pole. The gate checks numbers everywhere, so an invented level can't ride in on a valence line. It fails honest — ungrounded items are dropped and logged, not patched.
Significance gate (deterministic)¶
The synthesis may only narrate what the deterministic scorer flagged. Every observation must concern a ticker the scorer raised a candidate for (a ticker-less cross-sectional observation matches the cross-sectional candidate); every tension must share a number or ticker with a tension candidate. Items with no scorer support are invented significance and are dropped — all of them, exhaustively, not just the weakest.
Adversarial challenge (semantic)¶
What a rule can't check — semantic consistency — a final challenge round handles, and it handles every problem at once rather than the single worst, because the MAR is a public surface. A reviewer flags every item with a direction_contradiction (a ticker framed against its locked direction, including in its valence), an overreaching_valence (a personalised call or imperative instead of impersonal opportunity/threat), or a non_contradictory_tension (poles that don't truly oppose). The writer then fixes or argues each flagged item using only the scorer packs — never introducing new facts, always respecting locked directions and keeping valence impersonal — and the reviewer accepts or rejects each independently. Anything still unresolved after the rounds run out is dropped. The whole exchange is written to a diagnostic audit alongside the MAR — what the scorer flagged, what the LLM wrote raw, what each gate dropped — so a human can confirm the model only narrated what it was allowed to.
The inducement gate — decision-support, not inducement¶
The gates above keep the analysis grounded. One more gate keeps it non-inducing, and it applies channel-wide to anything that goes on air.
The publication's editorial stance is decision-support: equip the audience's judgment — the read, the level that invalidates it, what to watch next, the risk — in a conditional, observational register. What must never ship is inducement: imperative directives to transact, and outcome-promises. That register is what turns non-advice market commentary into a (restricted) financial promotion.
It's worth being precise about the line. The regulated "advice" line is a personal recommendation — advice tailored to a specific person's circumstances, objectives, and risk tolerance. A public feed structurally cannot cross that line, because it has no audience-of-one to personalise to. So the line the publication actually steers by is inducement, which a public feed very much can cross. The inducement gate is the enforcement of that floor.
The gate is built for high precision: it matches inducement constructions, not analytic vocabulary. The decision-support register's own words pass cleanly — support, resistance, level, break, breakout, holds above, short bias / long bias, risk, invalidation, 10x leverage. What it catches is the imperative-to-transact (buy now, load up, don't miss, ape in, take your profits) and the outcome-promise (to the moon, going parabolic, guaranteed, can't lose, next 100x). A hit fails the gate and forces a rewrite into the conditional register: the read holds above X; lose Y and the balance has broken — never buy here, never going to pump, never a promised return.
This is also why every MCP response carries a disclosure field, and why the analytical fields are framed as watchpoints and structural inflections rather than triggers. The MCP surface is the same content under the same floor — not personalised advice, not a trade-signal service, not a recommendation surface.
Published standards¶
The gates are the mechanism; the published policies are the standard. They are the public, reviewer-facing statement of what the pipeline above enforces:
- Editorial standards — the editorial floor the brief and MAR gates implement.
- AI disclosure — how AI is used in producing the analysis, consistent with the deterministic-versus-LLM split documented across this section.
- Corrections — how errors are handled after publication, the complement to the gates that catch them before it.
For what the outputs commit to and the canonical disclosure text, see editorial scope. For how to read the two fields the gates most tightly constrain, see conviction & direction.