R07 Kickoff Prompts
Copy-paste prompts to launch the two independent deep-research passes for R07 (soil vs. coco vs. hydroponic root-zone systems). Drafted 2026-09-09 after the brief froze (v1.0) and the archive audit completed. Both researchers have already seen the R07 brief in draft (ChatGPT reviewed v0.96 line by line and approved the freeze; Gemini reviewed v0.9 from Drive), so these are launch prompts for the existing threads, not orientation prompts. Paste the frozen v1.0 brief text directly after each prompt — the Drive index lags on new documents, and the run must cite v1.0, not a draft. New-thread variants follow in §3 in case either thread has to be restarted.
1 · ChatGPT (existing thread — launch)
Thanks for the freeze call. R07 brief v1.0 is FROZEN as of September 9, 2026 on my confirmation; the text you approved is the text below, with only the version stamp changed. Please run your deep-research pass against it now and cite “R07 brief v1.0 (frozen 2026-09-09)” in your report.
Two things from Claude’s archive audit that you should know before you start, because they sharpen what the report has to do:
- My own teaching archive defined “hydroponics” as any soilless medium, coco included in every handout from 2016 through 2024 — definition (a) in the brief’s four-way split. So the pre-registered taxonomy is not academic: the report is un-pooling a category that my curriculum, and most of the trade, pooled for years. When a source says “hydroponic,” re-file it and say which of (a)–(d) it meant.
- The only quantitative comparison I ever put on a slide (2024) was “hydro: 4–5 cycles a year; organic: 1–2.” That is a per-year cycle-count claim that crosses R06’s nutrition axis with R07’s architecture axis. It is pre-registered here as M1’s strongest form; I want to know whether any cannabis source reports cycle duration by root-zone system at fixed genotype and environment, and if none does, say so as an absence.
Deliverables per the brief: the root-zone system taxonomy (A1) and the controlled-trial inventory (B3) as distinct tables; the system-property matrix (H24) with every cell graded direct / adjacent / general / unmeasured; the three-outcome-plus-time map (H25); the flow map (H26); the mechanism table (H27); calibrated verdicts on M1–M10; top-10–15 sources with reliability and commercial-position notes; explicit data gaps; 15–30 candidate claims each carrying its evidence-type tag, boundary line, root-zone system descriptor and nutrient-equivalence status; and the bounded search-stopping rationale. Standing rules unchanged: resolvable identifiers, citation-identity integrity, exact claim support, one paper is one source, “unsupported” ≠ “the opposite is true.”
One R06 note carried forward: your single identity miss last cycle was an author-year misattribution (“Cockson” for Llewellyn 2023) with the DOI correct — the tuple check is the whole check, not the DOI alone.
Frozen brief follows.
2 · Gemini (existing thread — launch)
Thank you for the review of the R07 draft — four of your technical items went into the frozen brief (reframed as research targets rather than stipulated mechanisms, per ChatGPT’s editorial disposition). Your numbering item did not: the brief numbers its questions continuously 1–28 across sections A–H, the R06 convention, so B3 and H24–H27 are correct as written.
R07 brief v1.0 is FROZEN as of September 9, 2026. Please run your independent deep-research pass against the text below and cite “R07 brief v1.0 (frozen 2026-09-09)” in your report. Do not use the v0.9 Drive draft you reviewed.
The standing note applies, with respect. Your substance — taxonomy, mechanism instincts, regional breadth — has contributed to every ledger, and your bibliographic metadata has failed verification in every cycle including the correction rounds. R06: 15 of 28 load-bearing sources passed identity (5 wrong-object, 1 nonexistent “Zheng, Frizzone & Bugbee 2021”, 2 fabricated DOIs on real papers); 7 claim-support failures, including a +26.5% figure that was your own arithmetic and wrong; and inside the correction round, one title was re-fabricated and one sentence re-attributed. One operating rule above all: resolve every tuple — authors, exact title, year, venue, volume/pages, DOI — against the live record before including it, and where you cannot, write “partial citation — not fully resolved” instead of completing it from memory. A flagged partial passes review; a confident wrong one costs the cycle a correction round. A number you attribute to a paper must be in the paper; do not derive it.
R07’s disciplines that matter most for your pass: (1) substrate and delivery architecture are orthogonal axes — never infer one from the other, and place every system on both; (2) “hydroponic” is pre-registered — when a source uses the word, record which of the brief’s four meanings it carries and re-file the system; (3) three outcomes plus the time axis — cannabinoid concentration, flower yield, total cannabinoid yield, and cycle time are four different results, and per-cycle and per-year are different functional units; (4) the root-zone system descriptor (the eleventh provenance field) on every cross-system comparison — a comparison without it is system-orphaned; (5) root-zone temperature is not air temperature, and the temperature × dissolved oxygen × oxygen demand × pathogen question is an interaction to investigate, not a threshold to assert.
Required sections: A1 taxonomy table; B3 controlled-trial inventory; H24 system-property matrix; H25–H27 maps and mechanism table; M1–M10 verdicts; top sources with commercial-position notes; explicit data gaps; 15–30 candidate claims with evidence-type tag, boundary line, descriptor and equivalence status; search-stopping rationale. All are required; a missing section is a missing deliverable.
Frozen brief follows.
3 · New-thread variants (only if a thread must be restarted)
3a · ChatGPT
New thread, ongoing project. You’re a senior researcher and methods advisor on my Cannabis Cultivation Knowledge Engine (CCKE) — a verified, evidence-graded research corpus on cannabis cultivation feeding a public learning library, courses, an app, and an updated edition of my book. The system: you and Gemini run independent deep-research passes from a frozen assignment brief; Claude (in Claude Code) is editor and ledger-keeper — archive audits, live two-check citation verification, cross-examination, permanent claim ledger (IDs like R06-C001) and source library (S-218 registered so far); I direct and sign off.
We operate under PROTOCOL v1.4: (1) every substantive claim cites a resolvable identifier; (2) citation-identity integrity — the full tuple resolves to one bibliographic object; (3) exact claim support — direction ≠ magnitude/population/period/boundary; (4) corrections inherit full verification; (5) end with candidate claims + a search-stopping rationale; (6) calibrated myth checks — “unsupported” never becomes “the opposite is true”; (7) one paper is one source. Every load-bearing claim carries an evidence-type tag (measured / hybrid / modeled / observed-administrative / regulatory / absence / methodological) and every quantitative claim an inline boundary line.
Status: seven reports complete — R05 (environments), R01 (organic agriculture; practices-not-labels), R02 (global landscape), R08 (living soil; no controlled multi-cycle indoor no-till comparison exists), R03 (legal & illicit; “which population, measured how?”), R04 (environmental footprint; “a map, not a number”), R06 (fertility; “nutrient systems, not labels” — outcomes are properties of chemistry, rate, timing, substrate, root-zone chemistry, delivery architecture, genotype and co-inputs, never of the labels organic and synthetic). Wave 1 is full; R07 closes Wave 2.
Next: R07 — Soil vs. Coco vs. Hydroponic Systems. Spine: root-zone systems named by measured properties and delivery architecture, never by the labels soil / coco / hydro; substrate and architecture as orthogonal axes; “hydroponic” pre-registered into four meanings; three outcomes plus a time axis (per cycle vs. per year); an eleventh provenance field, the root-zone system descriptor; and a system-property matrix as the synthesis target instead of a winner. R07 extends R06-C015 (recirculating vs. drain-to-waste raised total THC yield in one trial) and must not paper over R08-C030. The frozen v1.0 brief follows; run your pass against it and cite it.
3b · Gemini
New thread, ongoing project. You’re one of two independent deep researchers on my Cannabis Cultivation Knowledge Engine (CCKE) — a verified, evidence-graded cannabis-cultivation research corpus. You and ChatGPT each run an independent deep-research pass from a frozen brief; Claude is the editor — it live-verifies every citation (identity + claim support), cross-examines both reports, and keeps the permanent claim ledger; I direct and sign off. Seven reports are complete: R05, R01, R02, R08, R03, R04, R06.
We operate under PROTOCOL v1.4. The standing note, with respect: your substance has contributed to every ledger; your bibliographic metadata has failed verification in every cycle including correction rounds (R06: 15 of 28 identity; 7 claim-support failures; re-fabrication inside the correction round). One rule above all: resolve every tuple against the live record before including it, and where you cannot, write “partial citation — not fully resolved.” A number you attribute to a paper must be in the paper.
Next report: R07 — Soil vs. Coco vs. Hydroponic Systems. Its disciplines: substrate and delivery architecture are orthogonal axes; “hydroponic” is pre-registered into four meanings and every source is re-filed; three outcomes plus cycle time, per-cycle and per-year kept distinct; a root-zone system descriptor on every comparison; root-zone temperature is not air temperature. Required sections: A1 taxonomy, B3 trial inventory, H24 system-property matrix, H25–H27, M1–M10 verdicts, sources with commercial-position notes, data gaps, 15–30 tagged candidate claims, search-stopping rationale. The frozen v1.0 brief follows; run your pass against it and cite it.
4 · Post-critique prompts (2026-09-10)
4a · ChatGPT — critique disposition (existing thread)
Claude’s verification and critique of R07 is done (critique.md, verify-A, verify-B are in the R07 Drive folder; pasting the critique below). Your side: 31 of 31 tuples resolved on Crossref and OpenAlex — zero identity failures, and all 13 of your self-flagged partials resolved to the objects you described, which is exactly how the protocol is supposed to work. Claim support 24 YES, 1 PARTIAL: CAND-09 — Allred 2025 says plant height fell 15% in both cultivars; the cultivar-dependent effect was yield (−20% in ‘Cherry’ under both high-EC treatments; ‘Trump’ only under the hybrid 8/4). CAND-19’s rockwool leg isn’t in the R15/R17 abstracts and carries an abstract-level flag.
Gemini’s pass: two wrong-object DOIs (one on a paper that does not exist), three real papers with altered tuples, two corpus IDs attached to the wrong sources, and none of its 16 claims fully supported by its own citation — the new pattern is invented magnitudes in tables. Its description of Yep, Gale & Zheng 2020 (60 L DWC, ‘Zenith’/‘G1’, 21.5 °C) is fabricated; yours (11 L peat + 3 L DWC aquaponic, +42–116%) is what the paper says.
Two things I want from you: (1) your disposition on the critique — approve to proceed to ledger, or flag anything you’d adjudicate differently, especially §2b-3 (cycle time as absence), §2b-5 (coir chemistry as mechanism, not rule) and §2b-6 (M4/M7 unverified); (2) a quick opinion on whether Gemini’s A1 taxonomy and H24 matrix scaffolding should enter the ledger as
generalstructure with no numeric cells, or be held entirely until its correction round. Critique follows.
4b · Gemini — correction round (existing thread)
Thank you for the full R07 report — the taxonomy, substrate-physics section and matrix structure are the strongest you have produced, and Claude’s critique says so. The bibliographic layer did not pass. Two-check verification (Crossref + OpenAlex, live): 2 wrong-object DOIs, 3 altered tuples, 2 corpus IDs misattributed, 2 unresolved, and 0 of 16 candidate claims fully supported by the cited source. The new failure mode is invented magnitudes: numbers in your H25, H26, F2, F3, F5 and G1 tables and in claims C010, C011, C015 and C016 appear in no cited source.
This is a correction round under PROTOCOL v1.4 rule 9. For each of the 24 items below, supply either (a) a resolving tuple — authors, exact title, year, venue, volume/pages, bare DOI — and the passage that supports the claim as you stated it, or (b) a withdrawal. Do not complete anything from memory; “partial citation — not fully resolved” passes review, a confident wrong one does not. Every replacement is re-verified before use, and a corrected export that introduces new errors is kept as a record and not used as a source. Return the corrections as a numbered list matching the item numbers; do not re-issue the whole report.
Items follow (critique §4, 1–24).
4c · ChatGPT — ledger review (existing thread; send after 4a, or together)
Ledger v0.1 for R07 is drafted (pasting below; also in the R07 Drive folder as “R07 Claim Ledger v0.1 (readable)”). 36 claims, R07-C001–C036, sections A–H mirroring the brief; nine keystones — C001 (taxonomy rule: substrate × architecture orthogonal, “hydroponic” re-filed, descriptor-less comparisons are system-orphaned), C002 (Yep 2020 as the standing system-bundle comparison), C003 (coir is not one material), C013 (delivery architecture, R06-C015 confirmed in full text with the verified comparators), C018 (M1: no cycle-time datum by system), C022 (no class-b/c sensory study), C024 (pathogens in coir and stone wool, two labs), C029 (“90% less water” is borrowed), C034 (R08-C030 re-confirmed). Four rows are PROVISIONAL pending Gemini’s correction round (C005, C012, C019, C021). Twenty-two new sources, S-219–S-240, with rule-12 lab clusters flagged in the header (Hohenheim 4, Guelph 5, Punja 2, Bugbee 2, Prague 2).
Same review as R04/R06: check every row against its cited source for magnitude, population, boundary and outcome; flag rule-12 dependence I’ve under-flagged; challenge any evidence-status or confidence grade; and call out any place where a boundary line or the root-zone descriptor is incomplete for a quantitative row (four rows carry “abstract-level; confirm” on container volume — Caplan 2019/2017b, Burgel 2020, Massuela 2026 — and C013 assumes Velechovský’s substrate is inert-substrate fertigation; confirm or correct). Return APPROVE / APPROVE WITH CALIBRATIONS / REVISE with a numbered edit list. Ledger follows.
4d · ChatGPT — final sign-off pass on ledger v0.3 (existing thread)
Verify-D landed and v0.3 is cut (pasting below; Drive: “R07 Claim Ledger v0.3 (readable)”). Verify-D on Gemini’s round: 24 items ACCEPT 20 / PARTIAL 3 / REJECT 1 — the reject is item 10, which said the search-wrapped DOI links were stripped when six still were; identity layer was clean for the first time in a Gemini correction round (0 wrong-object, 0 fabricated); three small new errors inside the round (“β-pinene” for pinene in Yep 2020; an “11-week flowering cycle” the paper doesn’t describe; Schober’s substrate list changed between rounds) — none adopted. Holds lifted on C005, C012, C019, C021 and the S-224 citation; Schober from Gemini stays held, the row rests on your CAND-04. All twelve of your calibrations are applied (four were already in v0.2). Absence rows C024/C026/C030/C034 re-bounded to “within the documented search (cutoff 2026-09-09)”. Your final pass: APPROVE for Danny’s signature, or a last numbered list. Ledger v0.3 follows.
4e · ChatGPT — synthesis editorial pass (existing thread)
A practitioner-record passage is withheld from the public copy.
Editor’s note: brief v1.0 FROZEN 2026-09-09 (brief.md, Drive doc “R07 Brief v1.0 FROZEN 2026-09-09”); archive audit complete same day (archive-audit.md). Raw inputs land in raw-inputs/chatgpt-YYYY-MM-DD.md and raw-inputs/gemini-YYYY-MM-DD.md. Danny answers the ten practitioner seed questions (archive-audit §Practitioner-experience record) at the raw-inputs stage; answers are tiered item by item. The R03 post-signoff audit remains queued in knowledge-core/AUDIT_LOG.md; at most one backlog audit runs alongside this round.