---
name: base-rates
description: Base rates and calibration for Ben's token screens, computed from the closed rows in [[screen-log]], [[screen-log-2]], [[screen-log-outcomes]] and [[screen-log-outcomes-2]]. Not yet computable. Read before quoting any hit rate or any claim about how often a screen is right.
sources: [cowork]
aliases: [calibration, hit-rates, screen-base-rates]
---

# Base rates

- [stated] Ben agreed on 31 Aug 2026 that base rates are the right third priority in the stack, after the screen log and the operator graph.

## Status, 6 Sep 2026
**0 closed rows of the 15 required. 16 rows open** (the 15 below plus COMMOTITTES, screened and called SELL on 5 Sep, still open in the log because a same-day SELL is not a 90 day close). No computation run. No row closed and no row hit a dead threshold on 6 Sep; **ARC is the closest ever at liq $7,306 = 1.46x the $5,000 floor and vol24 $822 = 1.64x the $500 floor.** Day 6 standings are at the bottom of this file. First 7 day marks were written at day 6 inside the +/-1 day window on 6 Sep and refresh at exact day 7 on 7 Sep.

## Status, 5 Sep 2026
**0 closed rows of the 15 required. 15 rows open (8 screened 31 Aug at day 5, RIPE screened 1 Sep at day 4, 5 tokens from the two 2 Sep multi-token sweeps at day 3, and STONK screened 4 Sep at day 1). No computation run.**

**Note the trap in that sentence: 15 OPEN rows is not the 15 CLOSED rows this file needs.** A row closes at 90 days or on a dead threshold, and none has done either. Two rows sat within 2x of the liquidity floor for a third day without closing (ARC liq $8,705, SCL liq $11,623). The 5 Sep run recorded a standing note about exactly this: **stop forecasting which row closes first, report the distance to the threshold.**

Nothing here is a base rate. Until this file has numbers, no statement of the form "screens like this usually" is supported by anything in the stack. First scheduled close is 29 Nov 2026 at the 90 day mark. **First 7 day marks land 7 Sep 2026** for SCL, BOW, DICE, BACKED, both RAM hub rows and ARC; RIPE on 8 Sep; the five 2 Sep tokens on 9 Sep.

Distance to the dead thresholds on 5 Sep, stated as a distance and not a forecast: **ARC liquidity $8,705 (1.74x the $5,000 floor), vol24 $2,365 (4.7x the $500 floor). SCL liquidity $11,623 (2.32x), vol24 $5,941 (11.9x).** Both were called "closes first" on three consecutive runs and neither has closed.

## What gets computed at 15 closed rows
- By category (payout, chain-identity, meme-site, infra-operated, other): n, median multiple from mc_screen to max_mc_since, share reaching 2x, share reaching 10x.
- By verdict (PROBE, STARTER, FULL, NO) separately.
- **The NO rows matter most. They are the only measure of missed opportunity and the only thing that stops the screen drifting conservative.**

## THE RULES THIS FILE HAS EARNED

### 1. A multiple inside 48 hours of a screen is noise, set 3 Sep 2026
| Row | Day 1 | Day 2 | Day 3 | Day 4 | What the early print said |
|---|---|---|---|---|---|
| ARC | 1.59x | 0.27x | 0.074x | 0.031x | 1 Sep called it "the widest gap against a verdict of any open row" |
| RIPE | 2.39x | 0.57x | 0.52x | | 2 Sep called it "the widest gap in the book" and "the single most useful entry in this file for calibration" |

Both inverted inside 24 hours, both were vertical-candle prints on tokens whose caps were mismeasured or unmeasurable by aggregators. **The first scoreable horizon in this stack is 7 days.** Delete the habit of naming a "widest gap" from a daily print.

### 2. A fired falsifier is not a price signal, set 4 Sep 2026
**BOW fired its structural falsifier on 3 Sep at 95.01% of borrowing sitting in disabled markets, and on 4 Sep printed +73.15% and a new high of $8,901,188, above its own mc_screen of $7,747,316.** The exposure got worse on the same day (95.03%, with wsNET-NN alone at 80.11% of the book against a 2,000 USDG cap, a 143x breach) and the price went up anyway. A falsifier is a statement about the mechanism; it carries no timing. **Report the falsifier and the multiple as two separate facts and never let one explain the other.** Twice now this file has written a fired falsifier next to a falling multiple; that was coincidence, not causation.

### 3. Score the routing mechanism, never the revenue curve, set 4 Sep 2026
Both directions appeared in the same run:
- **SLVR** revenue made a sixth consecutive daily ATH at $78,089/day while the price fell 24.85% to a 0.59x multiple.
- **O** made a new price high at $93,246,344 while o1 Launchpad revenue fell 73% in two days, $411,118 to $111,054.

Revenue predicts nothing about a token that has no claim on it. Routing does.

## Day 4 standings, recorded only so nothing can be retrofitted later
n is 14 and the horizon is two to four days. **This is not a base rate.**

| Verdict | Row | Multiple, day 4 | Reading |
|---|---|---|---|
| NO | ARC | 0.031x | right by 31.9x, closing as dead first |
| NO | SCL | 0.11x | right |
| NO | SELL RIPE | 0.52x (day 3) | right, after being 2.39x behind at day 1 |
| NO | SLVR (PROBE candidate) | 0.59x (day 2) | gate met and strengthening while the price falls |
| NO | RVH | 0.64x (day 2) | right |
| NO | DICE | 0.85x | right |
| YES lottery | BACKED | 0.95x | roughly flat, behind its entry |
| PROBE small | PAIR | 0.97x (day 2) | flat |
| NO | O | 1.07x (day 2) | behind, new high on collapsing fees |
| STARTER | BOW | **1.15x** | ahead of entry, and its falsifier is fired |
| NO | RAM 4th pass | 1.18x (FDV) | behind, narrowing three days running: 1.64x, 1.20x, 1.18x |
| cheapest in sector | PONS | **1.71x** (day 2) | ahead, and the best-performing verdict in the book |

**Every NO is right or flat except RAM and O. The two verdicts that are ahead of entry are the two positive calls, BOW and PONS. At day 4 the screen is not drifting conservative.** That is a two-to-four day observation and nothing more.

## Day 5 standings, 5 Sep 2026. n is 15, horizon one to five days. **Still not a base rate.**

| Verdict | Row | Multiple, day 5 | Change vs day 4 |
|---|---|---|---|
| NO | ARC | **0.026x** | 0.031x, still falling |
| NO | SCL | **0.049x** | 0.11x, halved |
| NO / SELL | RIPE | **0.37x** (day 4) | 0.52x, widest the SELL has been right |
| NO | RVH | 0.47x | 0.64x |
| NO (PROBE candidate) | SLVR | 0.62x | 0.59x |
| NO | DICE | 0.67x | 0.85x |
| YES lottery | BACKED | 0.77x | 0.95x |
| NO | RAM 4th pass (FDV) | **0.88x** | 1.18x. **The NO went ahead for the first time**, series 1.64x / 1.20x / 1.18x / 0.88x |
| NO | O | 1.00x | 1.07x, exactly flat at day 3 |
| STARTER | BOW | 1.03x | 1.15x, off its high, falsifier still fired |
| BUY | STONK | 1.21x (day 1) | first score, day-1 print, not evidence |
| PROBE small | PAIR | **1.43x NEW HIGH** | 0.97x, +50.6% in a day |
| cheapest in sector | PONS | **1.64x** | 1.71x, off its high, still the best verdict in the book |

**Reading, and it is the same one as day 4 with more separation: every NO is right or flat, and the four verdicts ahead of entry are the four positive calls (PONS, PAIR, STONK, BOW). The screen is not drifting conservative at day 5 either.** RAM, the only NO that was behind, went ahead today. That is a one-to-five day observation and nothing more.

## THE FIRST REAL BASE RATE IN THIS STACK, computed 5 Sep 2026 on the COMMOTITTES screen
Not from the screen log (still 0 closed rows). Computed directly from chain data, so it is available immediately and is reusable on every new launch.

**Method, reusable. Dune public query 8613013, `performance:"medium"`, ~90 seconds.** From `erc20_robinhood.evt_Transfer`: first-ever block time per contract = t0, then `count(distinct "to") filter (where evt_block_time <= t0 + interval '90' minute)` = **h90, the holder count at 90 minutes**. Window 2 Sep 06:00 to 5 Sep 06:10 UTC.

**THE DENOMINATOR IS THE FINDING: 119,588 tokens were launched on Robinhood Chain in 3 days.** ~40,000 a day, ~450 an hour. GeckoTerminal `new_pools` confirms it independently: 60 pools covered only the most recent 8 minutes. Any statement about a new launch being "early" has to be read against that.

### Distribution of h90 (holders in the first 90 minutes), n = 119,588
| h90 at least | tokens | share of all launches |
|---|---|---|
| 541 | 1,199 | **1.00%** |
| 400 | 1,594 | 1.33% |
| 300 | 2,066 | 1.73% |
| 200 | 2,975 | 2.49% |
| 100 | 5,948 | 4.97% |
| 50 | 10,225 | 8.55% |
| 25 | 19,087 | 15.96% |

### Survival by first-90-minute bucket: does the token have ANY transfer after hour 24?
| h90 bucket | n older than 24h | still alive at day 1 | rate |
|---|---|---|---|
| 541+ | 1,195 | 551 | **46%** |
| 300 to 541 | 863 | 424 | 49% |
| 200 to 300 | 898 | 348 | 39% |
| 100 to 200 | 2,958 | 581 | **20%** |
| 50 to 100 | 4,247 | 506 | 12% |
| 20 to 50 | 9,711 | 472 | **5%** |

**READ, and it cuts both ways. A strong first 90 minutes is a real filter: 46-49% survival at the top versus 5% at the bottom, roughly a 9x separation. But the ceiling on that signal is low: even a top-1% opening leaves better-than-even odds the token is dead inside a day, and "alive" here means one transfer, not a bid.** So h90 is a good screen-OUT tool and a weak screen-IN tool. Never quote a strong opening as evidence a launch will run.

**Caveat to carry: airdrop and spam contracts sit in the top of the h90 table** (one contract shows h90 49,319 with h_all 281,139, and several show tx90 ≈ h90, the airdrop signature). They inflate the count above any real token, so a percentile computed this way is CONSERVATIVE for a genuine launch. Filter with `tx90 / h90 < 1.5 and h90 > 1000` if a tighter number is ever needed.

**Companion live check, same day: only 20 of the chain's top 100 pools by 24h volume were under 24 hours old, and the best same-age comp (CATSTRO, 2.7h, $6,937,039 of volume on a $1,143,901 FDV) was doing 11x the volume of the token being screened at the same age.** Rank a new launch against the day's OTHER new launches, not against zero.

## Day 6 standings, 6 Sep 2026. n is 16, horizon one to six days. **Still not a base rate.**

| Verdict | Row | Multiple, day 6 | Change vs day 5 |
|---|---|---|---|
| NO | ARC | **0.018x** | 0.026x, still falling, closest to a dead close |
| NO | SCL | **0.057x** | 0.049x, up on 7 buys against 19 sells |
| NO | SELL RIPE | **0.31x** (day 5) | 0.37x, widest the SELL has been right |
| SELL, called intraday | COMMOTITTES | **0.27x** (day 1) | first score after the 5 Sep SELL, right by a further 2.08x |
| NO | DICE | 0.38x | 0.67x, round-tripped the announcement pump |
| NO (PROBE candidate) | SLVR | 0.58x | 0.62x, **falsifier 1 now 0.15% from firing** |
| NO | RVH | 0.66x | 0.47x |
| YES lottery | BACKED | 0.85x | 0.77x |
| NO | O | 0.96x | 1.00x, **the NO went ahead for the first time** |
| STARTER | BOW | 1.01x | 1.03x, falsifier still fired, periodFinish 48h out |
| NO | RAM 4th pass (FDV) | 1.05x | 0.88x, **the NO went behind again after one day ahead** |
| cheapest in sector | PONS | **2.41x NEW HIGH** | 1.64x, revenue ATH the same day |
| BUY | STONK | **3.09x NEW HIGH** (day 2) | 1.21x |
| PROBE small | PAIR | **3.38x NEW HIGH** (day 4) | 1.43x |

**Reading: every NO is still right or flat except RAM, and the four verdicts ahead of entry are the four positive calls (PAIR, STONK, PONS, BOW). Same shape as day 4 and day 5, with far more separation.** One to six days, and nothing more.

### 4. A green day is as much noise as a red one, set 6 Sep 2026
PONS, PAIR and STONK all printed new highs on the SAME chain-wide risk-on day (RH chain fees24h $21,335,628, Pons V2 revenue at an ATH). Two of the three were at day 2 and day 4. **Rule 1 already says a multiple inside 48 hours is noise; the failure mode this file has not yet had is reading a flattering print as vindication.** The discipline is symmetric: do not promote a verdict on a green day, and do not name a "best verdict in the book" from a daily print any more than a "widest gap".

### 5. A verdict is priced at the cap it was written at, set 6 Sep 2026
**STONK's BUY was written at a $19.67M cap on a 1.60x revenue multiple and an ANSEM comp gap of 13x. At $60.78M the multiple is 4.29x and the gap is about 4x.** The row is 3.09x and the thesis that produced it is largely spent. **Mark a row RE-SCREEN when the cap moves far enough that the frames would come out differently. Carrying a verdict forward unpriced is how a screen log turns into a scoreboard for a position.**

### 6. Record a near-miss falsifier as a distance, not as "not fired", set 6 Sep 2026
On 6 Sep two legs would have read identically as "not fired" in a binary column and both are the most likely things to change the book this week: **SLVR revenue $40,060/day against a threshold of under $40,000/day, 0.15% away**, and **BOW's StakingRewards periodFinish 1788858349 = 8 Sep 09:05 UTC, which fires on the calendar in 48 hours if the period is not refunded.** Same rule as the dead-threshold distances: report the distance.

### 7. A missing run is not a reading, set 6 Sep 2026
**Dune hit an account-level datapoint limit on 6 Sep and step 2 did not run at all.** A quota failure and a "nothing new" result read the same in a summary line. **Never let a missing run count toward a streak, a base rate, a cluster's "unchanged" status, or the v4.fun zero-day counter.** Written into [[operator-graph-2]] as the run header for that day.

## Status, 7 Sep 2026. **1 CLOSED ROW OF THE 15 REQUIRED. No computation run.**
**The first row in the book closed today: ARC `0xce845443428867d271929bd739148B85524B536a`, OUTCOME dead, at vol24 $56.59 against the $500 floor.** 15 rows still open. Full detail in [[screen-log-outcomes-5]].

- **ARC closed RIGHT.** Verdict NO at mc_screen $342,605 on 31 Aug; dead at $6,081 on 7 Sep = **0.018x**, a 56x avoidance. mc_screen to max_mc_since = $546,291 / $342,605 = **1.59x**, so the NO forwent a theoretical 1.59x and avoided a 98.2% loss
- **Closure basis: DEAD THRESHOLD at day 7, not the 90 day calendar.** Record the basis on every closed row. A table that mixes death closures and calendar closures without the flag will understate time-to-resolution and overstate how fast the screen resolves
- **Volume killed it, not liquidity.** liq $7,195 was still 1.44x above its floor when vol24 hit $57 with 0 buys in 24 hours. Second observation of the same shape (SCL and COMMOTITTES are both tracking it). **When a row is near dead, quote the volume ratio first**
- First 7 day marks were written at exact day 7 today for SCL, BOW, DICE, BACKED, both RAM hub rows and ARC. RIPE 8 Sep, the five 2 Sep tokens 9 Sep, STONK 11 Sep, COMMOTITTES 12 Sep. First scheduled 90 day close is still 29 Nov 2026

## Day 7 standings, 7 Sep 2026. n is 16, horizon one to seven days. **Still not a base rate.**

| Verdict | Row | Multiple, day 7 | Change vs day 6 |
|---|---|---|---|
| NO | **ARC** | **0.018x, CLOSED DEAD** | 0.018x, first closed row |
| NO | SCL | **0.042x** | 0.057x, -27.65% on 2 buys against 16 sells |
| SELL, called intraday | COMMOTITTES | **0.089x** (day 2) | 0.27x, -64.88%, right by a further 6.4x |
| NO | SELL RIPE | **0.26x** (day 6) | 0.31x, widest the SELL has been right |
| NO | DICE | 0.35x | 0.38x |
| NO | RVH | 0.51x | 0.66x, gave back yesterday's +39.75% |
| NO (PROBE candidate) | SLVR | 0.72x | 0.58x, **falsifier 1 is 1.72x clear on complete days, not 0.15% away** |
| NO | RAM 4th pass (FDV) | **0.85x** | 1.05x, the NO went ahead again. Series 1.64 / 1.20 / 1.18 / 0.88 / 1.05 / 0.85 |
| YES lottery | BACKED | 0.94x | 0.85x, first redemption in the series |
| NO | O | 0.97x | 0.96x |
| STARTER | BOW | **1.19x NEW HIGH** | 1.01x, falsifier still fired, periodFinish 32h out |
| cheapest in sector | PONS | 2.07x | 2.41x, off its high |
| PROBE small | PAIR | **4.83x NEW HIGH** | 3.38x |
| BUY | STONK | **9.01x NEW HIGH** (day 3) | 3.09x, +198% in a day |

**Reading, unchanged in shape for the fourth day: every NO is right or flat, and the four verdicts ahead of entry are the four positive calls (STONK, PAIR, PONS, BOW). The separation keeps widening in both directions.** One to seven days, and nothing more.

### 8. A verdict can be 9x right and more wrong than when it was written, set 7 Sep 2026
STONK is **9.01x** at $177,255,027 and rule 5's re-screen mark applies harder than when it was set: **cap/revenue is 12.5x on complete-day revenue against 1.60x at screen, and the ANSEM comp gap that carried the 4 Sep upgrade (STONK earning more revenue at 13x less cap) is now zero.** ANSEM.io $259M on $210,456 of 7d revenue; STONK $177M on $271,887. **The multiple column and the thesis column are different columns. A base-rates table built on the multiple alone will teach that the STONK call was a triumph of the method, when the method's own inputs stopped supporting it at roughly 3x.** Record, for every closed row, both the multiple AND whether the thesis was still standing at the close.

### 9. Rule 7 compounds, set 7 Sep 2026
**Dune was out of quota for a SECOND consecutive run.** The v4.fun zero-day counter now has gaps at both 6 and 7 Sep, and four falsifier legs (SCL holders, DICE fee-EOA transfers, COMMOTITTES holders, the chain-wide holder gap) have gone unread for seven runs. **A single missing run is a gap; a run of them is a hole in the base rate itself.** When this file finally computes, the rows whose falsifiers were never evaluable must be flagged, not silently scored as "not fired".