BlueMX · Internal · live-feedback round
Engine Room, team walkthrough
You're testing the live tool our carbon-development work will run through: six AI agents that draft, check, and stress-test VCS project documents, with a human (you) in the loop at every step. This page gets you productive in ten minutes.
1 · Getting in
Open the app link you were sent, enter the team access code (comes with the link, it's not written here on purpose), and you're in. Desktop browser recommended. Everyone shares the same workspace, so anything you do is visible to the whole team, that's by design.
2 · The two projects you'll see
BlueMX Aztlán (Nayarit, VCS #2842) Verification
Our registered project. The real filed PD is loaded, and the Verification tab holds all 70 real findings from the VVB's two review rounds. The job here: work the findings, draft responses with Arthur, edit them, and track each one open → responded → closed.
Veracruz (Alvarado Lagoon) Validation
The live project we're taking through validation. Core facts are pre-filled (600 ha, mangrove, Alvarado Fishing Cooperative). The job here: add what you know in Project Sections, upload documents, and let the agents drive the pipeline, they'll ask you for what they're missing.
3 · Reading the screen
- Sidebar, every project, with its current pipeline stage.
- The pipeline bar (top), one row that does both jobs: it shows progress (✓ done stages are tinted, the filled pill is where the project sits) and it's the navigation, click any stage to open that stage's output. Hover a stage to see which agent owns it. Verification (gold) is the post-registration findings tracker; Activity Log sits quietly on the right.
- Cost line, real API spend for that project, per call. Nothing is hidden.
- Question card, when an agent needs something only a human knows, it stops and asks there. Answer it and the pipeline moves on.
- Opening a project lands you where the work is: a project mid-pipeline opens on its current stage; a registered project (Nayarit) opens straight on Verification.
4 · Try this first, ten minutes
- Open Nayarit → Verification tab. This is our real audit record, statuses reconstructed from the two completed review cycles: findings the VVB closed are marked closed, ones we responded to are responded, and the ones the VVB kept open are open, that open list is the remaining verification work. Each finding carries our real filed response and the VVB's verdict.
- Filter to "open" and pick a finding, then hit "Draft with Arthur." ~30–60 seconds. Check: did it cite real sources? Did it build on our filed response and the VVB's evaluation? Did it say plainly what evidence is missing instead of bluffing? Edit, save, and update the status.
- Open Nayarit → PDD Draft → "Submission PDF." You'll get the full filed PD as a formatted 178-page document, cover page, contents, page numbers. This is the export we'd forward externally.
- Open Veracruz → Project Sections. Expand a section you know something about, stakeholders, site conditions, anything, type it in, save. This is how the project's record gets built.
- Answer the pending question (question card near the top) if one is waiting, and watch the pipeline advance on its own.
- Upload a document under Project documents, a survey, a permit, meeting notes. The agents read these and cite them by name.
5 · New since the first round, worth two minutes
- Validation readiness score (top of a validation project, e.g. Veracruz), how close the PDD is to what a VVB will check: applicability, completeness, evidence. Every gap links to the exact section that fixes it, with a draft with Arthur shortcut.
- ⚡ Auto-populate from evidence, dump a whole data room into Project documents (no filing needed), press the button, and Arthur drafts every empty section the evidence can support. Drafts arrive badged agent-drafted, review with their source files; saving a section marks it reviewed. No evidence → the field stays honestly empty.
- Monitoring tab, upload field-plot CSVs; get per-metric stats, monitored-vs-ex-ante deltas against the same VM0033 defaults the quantification used, and a Monitoring Report draft (.md/.docx).
- Program pages (sidebar), Dashboard (KPIs, findings burn-down, portal analytics), Finance inputs (replace the illustrative investment-model numbers; the flag is yours to clear), Partners → Access & codes (named partner codes, NDA status, per-document permissions, revocation, activity digest), and the partner Inbox.
- Partner portal (Preview portal in the sidebar), the external face: program overview, live-computed economics with the cash-position curve, data-room downloads, and a question form that lands in our Inbox. Everything a partner views or downloads is access-logged under their name.
- Guardian round-trip (Drafting toolbar), export the PDD as a Hedera Guardian credential, and re-import what comes back for a field-by-field reconciliation.
- Restoration plans (sidebar), the Veracruz delivery model as a working tool: the seven PAI sites with costed activity schedules, unit costs with work specifications, delivered quantities since 2022, an interactive map of every channel, firebreak and fence line from the GIS package, and the plan documents themselves.
- Agent instructions (sidebar), tell an agent what to do broadly, not just per response: a standing directive per agent (or for all) that rides on every call until you change it. Attributed and history-kept. It cannot relax the evidence rules.
- Restoration delivered (Monitoring tab), record hectares actually restored as of a date, with a source. This is the figure the partner portal shows, so it is entered by a named person and never inferred.
6 · The agents, in one line each
CodyOrchestrator, routes work, runs the guided interview, keeps project memory.
ArthurCompliance & drafting, cites the exact methodology clause for every claim; drafts PDD sections and finding responses.
QuincyCarbon math, deterministic code, not an AI guess; every formula logged for hand-checking.
LibbyEvidence, indexes every uploaded document and maps claims to sources.
SidPre-audit, adversarial reviewer trained on our real Nayarit findings; scores draft readiness 0–100.
CaseyInvestor comms, plain-language status reports from the technical record.
7 · Honest limits, please read
Everything an agent writes is a draft for your review. Nothing goes to a VVB, a registry, or an investor without a human signing off. The agents are built to cite sources or say "evidence missing", if you ever catch one inventing a fact, that's the single most valuable bug you can report.
- The first agent call after the site has been idle can take ~60–90 seconds (the model wakes up). After that it's fast.
- Nayarit finding statuses were reconstructed automatically from the audit record (responses + VVB verdicts). Spot-check them, if one looks wrong, correcting the status is useful feedback.
- When a new review cycle arrives, upload the VVB's findings document (.docx) with the "Import new cycle" control at the top of the Verification tab, matched findings get their statuses updated from the new verdicts, new findings are added, and nothing is ever deleted.
- Quincy deliberately refuses to show a net credit number until every term of the equation is real, a partial honest answer beats a complete invented one.
8 · What we want from you
Use it for real work, then tell us, rough notes are fine:
- Where did you get lost or annoyed? (First confusing screen, unclear label, dead end.)
- Where was an agent wrong, vague, or overconfident? Copy the exact text.
- What would make this save you real hours that it doesn't do yet?
- Would you trust the Submission PDF in front of a VVB? If not, what's missing?
Send feedback to Tucker directly, or drop it in the team channel.