Citation-grounded triage
for Solana DAO delegates.
Red flags, treasury impact, stakeholders, and the questions worth asking. For any Solana DAO governance proposal, in 30 seconds. The skill is hard-ruled to never recommend a vote.
Marinade MIP-21. Treasury authority change. 4 red flags, every claim cited.
MIP-21 proposed transferring 3,350 mSOL ($586,932) from the DAO treasury to a Council wallet, redirecting protocol fees to that wallet, and committing to monthly MNDE repayments at "equivalent value."1
The triage flagged: vague_kpi (no quantified target for "supporting protocol growth"), missing_milestone (no timeline for when fees revert to DAO), backdoor_permission (Council gains ongoing revenue authority), and technical_risk (no specified valuation methodology for "equivalent MNDE value").2
Six proposal-specific delegate questions. Four stakeholders identified, each cited. Treasury direction internal_transfer, confidence medium. Output: a markdown summary a delegate can read in 30 seconds before opening Realms.
Proposal in. Structured triage out.
Two phases. The first reads end-to-end and scans systematically for risk categories. The second emits a JSON object with six required sections, every claim grounded in a verbatim citation. No on-chain step in the user-facing tier.
Triage
Reads the full proposal including every appendix. Identifies stakeholders, locates monetary amounts, scans systematically for KPI measurability, milestone bounding, permission changes, supply effects, and off-chain dependencies. Buried permission changes get equal scrutiny to the main body.
Output
JSON object with six required sections: TL;DR, treasury impact, red flags (categorized + cited + severity-rated), stakeholders, open questions, honest gaps. Vagueness in the source is itself surfaced as a finding. Output renders as markdown a delegate can read in 30 seconds.
How we prevent hallucination.
Hallucination in governance costs delegates real money. The triage carries three load-bearing constraints, encoded as hard rules in the skill prompt. A failure on any one invalidates the output. Citation rate across N=10 live proposals tested on Marinade, Jupiter, and Jito: 100%.
- Grounding. Every claim cites the proposal text verbatim, or the claim is omitted. If the skill cannot point to a quote, the finding does not appear.
- No interpolation. Don't fill from general DAO knowledge or external context. The triage analyzes what's in the proposal, not what should be.
- Vagueness is a finding. Don't paper over soft KPIs, open-ended scope, or missing milestones. Surface them as
vague_kpi,unbounded_spend,missing_milestone.
The eight categories the skill scans for, each with a verbatim-citation requirement:
- vague_kpi: success metrics are non-measurable, no baseline, no thresholdflagged + cited
- unbounded_spend: budget open-ended, no dollar cap, no review triggerflagged + cited
- missing_milestone: disbursement not tied to deliverables or checkpointsflagged + cited
- backdoor_permission: authority change buried, omitted from summary, framed innocuouslyflagged + cited
- centralization: explicit concentration of stake, voting power, fee flow, executive authorityflagged + cited
- token_dilution: any modification to supply trajectory, regardless of directionflagged + cited
- technical_risk: oracle dependency, off-chain execution risk, mechanism uncertaintyflagged + cited
- team_track_record: proposer history materially relevant, undisclosed, or contestedflagged + cited
Honest disclosure is preferred over confident guessing. When the skill cannot resolve a question, it degrades the confidence field and surfaces the gap as open_question or honest_gap rather than fabricating detail. The hard rules are load-bearing in this direction.
Five active delegates ran a real triage.
Five active Solana DAO delegates ran a PsycheForge triage of Marinade MIP-21 (the recent treasury authority proposal that passed 99.6 / 0.4) and answered four questions about whether the structured format actually helps decision-making. All five reported strong-yes value. Quotes verbatim, names withheld at delegates' request.
Three tiers. Same skill, same calibration. Different breadth.
Pricing is calibrated to delegate workflow intensity, not feature gates. Every tier returns the same triage quality on every proposal it covers. The difference is how many DAOs and how much surrounding context.
- 1–2 DAOs of your choice
- Full triage on every proposal in those DAOs
- Citation-grounded, no hallucination
- Markdown + JSON output
- Multi-DAO across active Solana governance
- Everything in Lite
- Priority on contested votes during active cycles
- 14-day pilot for early users
- Comprehensive coverage across Solana DAOs
- Everything in Standard
- Alpha-grade context: on-chain stakeholder lookups, historical pattern surfacing
- Direct line to the team
What works. What doesn't. What's paused.
PsycheForge is a PoC. Customer discovery completed 2026-05-05 with 5 anonymized delegate interviews. Real value confirmed. No paying users yet. The public Lite tier is beta-only, not generally available. We expect to learn things in the beta that update the copy on this page.
Works
- Triage skill produces grounded, cited analysis on real proposals
- Citation grounding: 100% across N=10 live proposals tested
- 5 / 5 strong-yes from anonymized delegate validation cohort
- Calibration discipline: v0.3 regression caught and rolled back same week
- v0.4 incorporates three corrections from delegate feedback
- Open-source: skill prompts, test fixtures, validation reports public
Doesn't (yet)
- Zero paying users
- Public Lite tier not GA, beta-only
- No automatic URL fetching for DAO forums
- Multilingual stress-test still narrow
- Some DAO coverage partial (Sanctum)
Paused (honest disclosure)
- Vote-outcome predictions: at N=7 weighted accuracy was 51% (coin-flip)
- Coin-flip accuracy on contested votes does more damage than missing predictions
- Returns as a public-facing claim only when N≥50 calibrated accuracy is demonstrable
- On-chain reputation infrastructure: built, runs on devnet, paused as a marketing claim
- Until then: triage only
Two minutes to a working triage on a proposal you actually care about.
Send a DM with the URL of a Solana DAO proposal you're currently weighing. We reply within 4 hours with a Lite-tier triage of that proposal, free, no commitment. If it changes how you read the proposal, the 14-day pilot is on us in exchange for one short interview at the end. If it doesn't change anything, no follow-up. We're not interested in selling tools that don't earn delegate trust.
DM us a proposal URL
Send a Marinade, Jupiter, Jito, or Sanctum proposal link. Free Lite-tier triage back within 4 hours. No card, no commitment.
Run it yourself, open-source
For builders. The whole pipeline is MIT-licensed: skill prompts, test fixtures, validation reports, the Lite-tier wrapper. Self-host with any OpenAI-compatible LLM backend.