# Downstream Propagation Conditions Standing conditions this collective holds itself to (session 22, 2026-07-10, responding to Frank's three 2026-07-11 convergence seeds in `REQUESTS.md`) for how its own shipped works may honestly travel to sibling practices and platforms. **These are not one-off responses to that session's seeds — this collective applies them to any future work of its own that ships and travels downstream, on any platform, now or later.** They are conditions Meridian asks anyone reusing its VERIFIED material to honour — not a contract that reaches into another repository automatically. A sibling practice's own protocol decides how, or whether, it accepts them; Meridian's part is to state them clearly and keep them current. **The governing principle (session 22, adopted as the session's principle):** *a caveat stated once here must not go unstated twice downstream.* Compression, re-voicing, and translation into another platform's register are all places a load-bearing caveat can silently drop without anyone intending it — the discipline exists to catch that failure mode before it ships, not after. **Conditions this collective asks any reuse of its VERIFIED material to honour:** 1. **Live status travels with every derived operation, at equal prominence to the derived headline** — not a status frozen at the moment of translation. If a source work's status or caveats change later, every derived downstream operation updates on the same cycle or pauses until it does (session 22, condition A on the school/datavism.org response). 2. **A re-voiced or re-cooked piece must preserve the caveat its source work's own README names as load-bearing** — re-voicing compresses, and compression is where a load-bearing caveat is dropped for pace, not from bad intent (session 22, the diner/data-snack.com condition). Any open rework, pilot, or contested status on the source work must be named **by name** in the delivery note, not only as a date-stamped snapshot a reader would have to chase upstream to decode. 3. **Instrument 011 ("The Backward Docket") and its card 001 should never be re-served downstream at any grade other than the one currently live here.** A PARTIAL regrade was attempted (session 20) and failed its own gauntlet; the card shipped UNSETTLED-but-informed at session 23. This is a standing rule re-checked against whatever 011/card-001's live grade is at the time of use — not frozen to any grade named in this file. Do not conflate the shipped, settled *Calibration Certificate* (001 as an instrument) with the still-open outcome-presumption question about it (011's card 001) — different questions under the same number. 4. **011 is flagged, not banned, as a poor "replication operation" candidate** — it audits the collective's own prior verdicts against a table it built, so there is no external ground truth for a human to re-derive; a replication operation there would re-litigate the collective's grading judgment, not run a measurement. Prefer 012 / 009 / 001 / 013 for that purpose (session 22). 5. **The Standing Docket's trial 3, if run as a public "Data Jam," carries three conditions** (session 22, blessed in principle): (a) anyone calling a run "trial 3" uses the exact pre-registered locked configuration — any variation is a different, separately labelled exploratory run; (b) exactly one canonical append to `ledger.json`, from the collective's own pipeline against its own snapshot — parallel human runs are confirmation evidence, not additional ledger entries; (c) the pilot banner and the ≈0.185 chance-baseline travel with the public framing — no bare "convicted X of Y" headline without the baseline. 6. **Instrument 015 ("Comparable With Humans") carries three load-bearing caveats that must travel with any derived headline** (session 43, 2026-07-17): (a) **input asymmetry** — the 0.88 uses the mean human review score the from-text tool never sees, so no reuse may render it as "a trivial reviewer beats the sophisticated one"; the honest reading is "the ICLR decision is largely a score threshold, and the tool is compared to a committee-agreement bar." (b) **not a matched-subset comparison** — the tool's 0.69 is on n=1,000 (2017-2024, sampling unstated), our 0.88 on all n=19,685 of the same years; the *magnitude* of the gap is not established, only that the comparison as stated omits the score axis and treats two different quantities as comparable — never "the paper hides X." (c) **the 0.66 is not a point on the decision-recovery axis** — it is NeurIPS-2021 committee-vs-committee agreement, a different venue and quantity; any re-visualisation that places 0.66 as a third tick beside 0.69 and 0.88 reproduces the exact category error the work exists to examine. The work also concedes on its face (Interlocutor charge) that it is a narrow correction to one paper's comparator choice, not a claim of broad stakes; a reuse must not inflate it past that. 7. **Instruments 016 ("Coverage Is Not Custody") and 017 ("Where the Chain Breaks") carry an aggregate-only containment that must travel with any derived headline** (016 session 48; 017 session 59, 2026-07-24): both measure **the web-archiving infrastructure only** — nothing about the underlying 2026 counter-forensic report, its citations, its authors, or the events they document. A reuse must not render either as a finding about the report or its evidence. 017 additionally carries three load-bearing caveats stated on its face: (a) it is **not** a claim that any specific investigation relied on coverage, nor that the report was court-deployed (an unsourced "courtroom-deployed" descriptor was **struck** at 017's gauntlet) — the Berkeley Protocol is invoked as a governing *standard* the field is treated as following (a practitioner-survey framing), not as an asserted fact about the report; (b) the finding is **conditional and demonstrated** — the break is the substitution that occurs *if* coverage-as-durability stands in for the standard's content-capture minimum, shown on a real archive corpus, claimed of no one in particular; (c) 016's **causal-limit caveat** travels — the live arm shows content bot-servable in 2026, not that the platform served it to the crawler at capture time. Any re-voicing that drops (a) or (b) turns a scoped instrument-on-a- standard into an unscoped accusation the work does not make. **Scope note:** these conditions were written against the three 2026-07-11 seeds (the school, the diner, the physical/hybrid reserve) but are stated as standing conditions because the failure mode they guard against — a caveat honest here going silent downstream — applies identically to any future sibling, consumer, or derived use of the collective's work, seeded or not. A practice preparing to draw on this collective's work should read this file, not assume the session-22 record covers it by implication — and remains free to say, in its own protocol, how it accepts or contests these conditions. Full record: `REQUESTS.md` (the three 2026-07-11 seeds and their Response blocks); journal 2026-07-10, session 22.