---
name: gem-hunter
description: "Force the three-frame valuation format and a mandatory benchmark comp table into the gem-hunter output, so a screen can never again be delivered without frames or comps."
---

# Gem hunter

## OUTPUT FORMAT IS NOT NEGOTIABLE, READ THIS FIRST

**Every gem-hunter screen is delivered in the three-frame format from `three-frame-valuation`, with a benchmark comp table. There is no exception and no "this one is different".**

This section replaces the old instruction to follow `lowcap-screen`'s verdict format. That instruction caused a real failure on 31 Aug 2026: the VYNEX screen was delivered as a verdict-plus-receipts wall with zero benchmark comps, because this skill's own output section outranked Ben's standing format. It does not outrank it. It never did.

Order of operations for the output:

1. Plain-language conclusion, two or three sentences, answering the exact question asked.
2. **Frame 1 fundamental anchor** with a dollar range in the heading.
3. **Frame 2 live-product floor** with a dollar range in the heading.
4. **Frame 3 narrative ceiling** with a dollar range in the heading, resolved as BACKED or NARRATIVE with the receipt.
5. **The benchmark comp table.** Mandatory. Named same-sector projects, full contract addresses, cap, real TVL, MC/TVL, age. A screen without this table is incomplete and must not be sent.
6. Everything else much shorter than the frames: biggest risk, fatal flags or none, tier and trigger, exit depth at Ben's real size, the three falsifiers.

The seven checks below are the INPUT to the frames, not a substitute for them. Run them, then write frames.

No dashes or em dashes. Full contract addresses, never truncated. No hedging, no moralising, no "do your own research".

## How this differs from lowcap-screen

`lowcap-screen` answers **"what is wrong with this token."** It is a verification engine and it is excellent at it. Run alone, it produces objections instead of decisions, because nothing in it ever asks what the token could be worth.

This skill answers **"what would have to be true for this to be a 20x, and how much of it already is."**

**Run them in this order, always.** Thesis first, kill checks second. Reversing the order is the failure mode this skill exists to fix: opening the fatal list before a thesis exists reliably produces a no, including on the projects that go on to run.

## Step 0 — read memory BEFORE anything else

Call `memory_list`, then read `/preferences.md` and any `/areas/` file covering the chain or sector. This comes before invoking any other skill and before the first API call. The 31 Aug VYNEX miss happened because this step was skipped and the standing format plus the chain's comp table were never loaded.

Then load `ben-positions`. Actual sizing runs $2,400 to $34,976, weeks to months. The $100 to $500 lottery premise in the older parts of `lowcap-screen` is void for this book.

Consequences that follow and must be stated, not suppressed:

- Exit depth is a live constraint. State round trip cost at his actual size.
- Overlap is a live constraint. If the candidate correlates with something already at 15% of book, say so with the number.
- Do not lecture about diversification or risk management. State the figure and move on.

On a weeks to months horizon, **structure and dated catalysts lead; the tape is entry timing only.** A hostile hourly chart is a price question, never a thesis question.

## Step 1 — write the bull case before opening any data source

One paragraph, before the first API call, naming which of the four asymmetry sources this candidate is claiming. If none of them can be named even speculatively, stop and say so in one line. That is a saved hour, not a failed screen.

### The four sources of asymmetry

**1. Mechanism: holding it pays the holder, with no action required.**
The measured discriminator is not whether a product exists, it is whether the holder must operate anything. Measured 30 day medians on Robinhood Chain, Aug 2026: pays the holder automatically **+1,925%** (n=3), wears the chain's identity **+378%** (n=5), meme with a website **+341%** (n=9), infrastructure the holder must operate **+29%** (n=7). Payout tokens returned **66x** what infrastructure tokens returned over the same window, and both groups had real products, docs and teams.

When reading a landing page, score the copy: holder-directed payout phrasing is positive, a requirement that the holder operate something is a **negative**. A CDP whose docs say "vault", "distribution" and "backed by" is not a payout token.

**2. Venue: it could only exist here.**
Ask whether the project could ship unchanged on Base or Solana. If yes, downweight hard. The outperformers are pipes between the boring money already on a venue and the speculative money arriving on it. Receipt: Artificial Inu, a dog coin whose LP pair is tokenised NVDA, 42.7% of $14.7M daily volume quoted in NVDA, +1,925% in 30d, while NVDA itself did $66.7M/day and moved +13%.

**3. Narrative ownership: a named account is on the hook for the story.**
Not follower count. Posts that cite **this** contract, and an attention series with a rising slope.

**4. Structure: the float and the bid underneath it.**
`market_cap_usd == fdv_usd` means 100% circulating and zero unlock overhang, which is rare and worth stating plainly. Then find the real bid with `balanceOf(quoteToken, pool)`, never reported liquidity.

A candidate needs one source clearly or two partially.

**The fifth source is timing**, and it is a pipeline problem, not an analysis problem. It comes from protocol deploys before a pool exists, insider follow graph movement, and a monthly regime re-read. It never comes from staring at a chart.

## Step 2 — find what the token does, then price it

Run `lowcap-screen` check 1 in full and `onchain-forensics` for the endpoint quirks. Three rules that carry most of the value:

- **Read the holder list for function, not concentration.** Pull `token-balances` on every holder labelled TimelockController, Treasury, Vault, Escrow, Staking, Reserve, Locker, PoolManager or Safe. This one extra call per address is where theses live and it is the largest recorded miss: a TimelockController at 0.750% holding 321 tokenised NVDA (about $72,535) accrued from LP fees was skipped, and the verdict "pure memecoin, no product" was wrong.
- **Absence is also a finding.** If NO labelled protocol contract appears in the top 300, the token is probably not wired to the product it advertises. Say so explicitly. Receipt: VYNEX, 31 Aug 2026, whose site and docs never mention the token and whose factory, factory owner and vaults all hold zero of it.
- **Mechanism and magnitude are two separate required facts.** "There is a fee to treasury vault" and "it holds 0.55% of market cap" must both be reported. Either alone misleads. Receipt: VYNEX's VaultFactory is genuinely deployed with `vaultCount()` = 16 exactly as its docs claim, and the sixteen vaults hold $519.36 between them with `harvester()` set to the zero address, so no yield can ever be paid.

## Step 3 — the line. Sort every defect into one of two columns

This is the core judgement and the reason the skill exists. Amateurism is not extraction.

### The governing question

**Does the flaw sit somewhere you can watch, or somewhere you have to trust?**

Presentation is observable, so scruffiness is cheap. Permissions are what you would have to trust, so a permission flaw is total. A project can be ugly in every visible way and structurally honest. A project can be immaculate and one `owner()` call from taking everything.

### Survivable, price it by cutting size

- Anon team, no branding, an ugly one page site.
- Launchpad template contract rather than bespoke code. **4 of 8 chain winners used one**, so contract bespokeness is not a quality signal and was dropped as a token-level test.
- 60 to 80% off ATH **with a base and a flat or rising holder count**. Median drawdown among tokens up more than 100% over 30 days was **-41%**, and a third were worse than -45%. A deep drawdown is the normal resting state of an uptrend here. Never call a token late on the drawdown number alone.
- No product shipped, but contracts deployed and the deployer funded. This is the earliest signal available.
- A 5% buy and sell tax on a 10x thesis. Fatal on a 1.5x thesis, noise on a 10x one. Always state the round trip.
- A thin pool, **if** a real quoted pair exists to work manually. Receipt: ARROW's real depth sat on a USDT pair that routers never used; the trade was $5k entered manually against USDT, not abandoned.
- Docs ahead of the chain, roadmap slipping. Downgrade the tier, set the trigger, wait.
- An unverified contract **where the behaviour is observable in transfers**. Mechanism observable, logic unauditable, smaller size.
- An unverified contract where custody is the product, **when Ben is buying the token and not depositing**. That is a depositor's risk, not a holder's. Keep the two separate and say which one applies.

### Fatal, no position at any size

- Live mint authority, blacklist, pause, or modifiable anti-whale in one hand.
- LP unlocked and concentrated in the deployer against a real pool. Check `is_locked` per holder via GoPlus; do not assume either way.
- An unverified contract **where custody or the payout is the product AND Ben would be depositing**. Receipt: VozDex, zero source for a protocol whose whole proposition is custody of liquidity.
- Narrative migration: ticker chatter cites a different contract with more volume. You are screening the abandoned shell.
- Reported liquidity with near zero turnover. Receipt: a fake NET at $48.2M liquidity, near zero volume, 2 holders. **Never rank candidates by liquidity; rank by traded volume.**
- Contributor count peaking *after* price peaked. That is a distribution signature.
- Clone farm, uniform holder slices, or a one-way seller cluster **confirmed against a second token on the same chain**. Unconfirmed, it is chain-wide infrastructure, not evidence.
- The site itself says there is nothing there while you are pricing something. Receipt: PCC scored 70 on a fetch check while its homepage read "No utility. No promises." Read the copy, do not just fetch the URL.

**Every left column item costs size. Every right column item costs the position.** Confusing left for right misses the run. Confusing right for left funds someone's exit.

## Step 4 — size the doubt

Never size on conviction. Size on how much of the thesis has been *verified*. Less observable means smaller, and smaller is not zero.

| Tier | Condition | Share of full | Upgrade rule |
|---|---|---|---|
| Probe | One asymmetry source present but the mechanism is claimed, or observed and empty. Nothing fatal. | 10 to 20% | A written trigger, one checkable metric |
| Starter | Mechanism observed on chain and priced. Poor entry geometry, or one fatal-adjacent question open. | 33% | The question resolves, or price reaches the named shelf |
| Full | Mechanism observed and priced, narrative owner identified, dated catalyst inside the horizon, round trip under 5% at size. | 100% | None. Adding past full turns a thesis into an identity. |
| Zero | Any single right column item. | 0% | None. No price fixes a permission. |

Always state: round trip cost at his size (tax plus impact both ways), and net buying required for the target, `(L/2) x (sqrt(N) - 1)`, roughly 1.08x the pool for a 10x. Then the honest follow up: **is there any sign that money is arriving.** A cheap 10x nobody is bidding for is not a 10x.

The upgrade trigger is written **before** the probe goes in and names a metric, not a feeling. "`harvester()` returns something other than the zero address" is a trigger. "It starts looking stronger" is not.

## Step 5 — write the three falsifiers, then the ladder

Getting shaken out is a specification problem, not a discipline problem. If the thesis was never written, every candle is evidence against it.

At entry, produce all three in the same sitting:

1. **The mechanism sentence.** What the token structurally forces to happen, and the on chain place it is visible.
2. **Three falsifiers.** Specific, checkable, dated where possible. Only these can close the position early.
3. **The exit ladder.** Tier prices plus a trailing stop, from `meme-exits`, set while the position is boring.

### The fud filter, for hold questions

Not invalidation, unless already named as a falsifier:

- A 60 to 70% retrace with a base. The discriminator between base and abandonment is **the holder slope, not the volume**: volume drying up while holders hold is accumulation, volume drying up while holders fall is the end.
- The team going quiet. The team wallet *moving* is a different fact entirely.
- Bear posts from accounts with no position history.
- "It is just a fork", when venue asymmetry is the thesis.
- A competitor launching.

Also apply the LunarCrush reading: **never quote sentiment**, it held 77 to 86 through an entire 60% drawdown. Use `interactions` slope as the primary signal and `contributors_active` as the distribution tell.

### The mirror failure

Holding with no written exit is the opposite error and costs more. Reference case in `crash-thesis`: a source with a real early entry record who has **never issued an exit signal on any account in any timeframe**. A thesis you cannot be talked out of still needs a ladder that sells for you.

## Step 6 — wire the trigger to something that watches

The upgrade trigger and the falsifiers are only real if something checks them without Ben. Use `alert-bots`, and note that fleet has one fixed home: a duplicate running elsewhere silently doubles every alert.

## Verification, run before delivering any verdict

- **Is the output in the three-frame format, and is the benchmark comp table present with named comps, full addresses, caps and MC/TVL?** If either is missing, the screen is not finished. This is the check that failed on 31 Aug 2026.
- Was memory read before any skill was invoked?
- Did the thesis get written before the fatal list was opened? If not, the verdict is contaminated by objection-first framing. Redo step 1.
- Was `token-balances` called on every labelled contract holder, and was the ABSENCE of labelled holders reported if that is what was found?
- Is every claim carrying its figure, and is every mechanism carrying its magnitude?
- Was any defect placed in the fatal column that is actually survivable? Re-run the observability question on each one.
- Does any holder table come from a swept and RECONCILED ledger rather than an explorer endpoint? Blockscout `/holders` has served stale balances.
- Is exit depth stated at Ben's real size, and is overlap with the existing book checked?
- Are the three falsifiers checkable by a script or a bot, rather than by remembering?
- If a prior call in this thread is now wrong, is the correction explicit and does it name the number that misled?