Sure, 27 is a cliff. No, I'm not docking RBs twice to fall off it.
The cliffhanger I left
A month ago I went after the "sell every RB before 27" crowd, fully expecting to dunk on them. They were right. At the same Y0 workload, backs entering their age-27 season forfeit about three points a game off the 300-touch tier versus younger backs carrying the same load. Same fuel in the tank, less output. The cliff is real and it scales with usage.
But I was honest about the part I had not earned. The cohort told me the cliff exists. It did not tell me the engine was getting it wrong. To know THAT, I'd have to compare the cliff against what the engine already projects for these backs, and pull the age effect out cleanly first, an age-neutralized projection variant that does not exist in the codebase yet. So I left it on the table: real effect, ship decision pending.
The check that did not need a harness
Before building a whole Gate-2 rig, there's a cheaper question I should have asked first. What does the RB age curve already do at 27? That's not a research project, that's a config file. So I opened it.
| age step | curve factor | one-year drop |
|---|---|---|
| 24 to 25 | 0.995 to 0.976 | -1.9% |
| 25 to 26 | 0.976 to 0.953 | -2.4% |
| 26 to 27 | 0.953 to 0.897 | -5.9% |
| 27 to 28 | 0.897 to 0.837 | -6.7% |
Look at where the floor drops out. The curve cruises down about two percent a year through the mid-twenties, then the two biggest single-year cuts in a running back's entire arc land back to back, 26 to 27 and 27 to 28. Across those two years it sheds 12.2 percent. On a workhorse the engine projects around 15 a game, that's roughly 1.8 points evaporating, right at the age the cohort screamed about, before I add anything at all.
Why a new penalty would be a bug, not a fix
The engine's forward number for a returning back is, at its spine, last year's points run through that curve: ppg0 x (curve[age_next] / curve[age_now]). When a 26-year-old becomes a 27-year-old, the engine is already multiplying his production by that 5.9 percent cut. The decline is not missing from his projection. It IS his projection.
So picture bolting a standalone "age >= 27" penalty on top. You'd subtract the cliff the cohort measured from a number that already walked off the same cliff. That's not catching a market inefficiency, that's double billing. The board would mark down every aging back twice and start handing you back the durable bell-cows it should be valuing correctly.
This is the same trap I just spent a week clearing off the desk. The QB touchdown-rate "lift," the rushing-QB persistence signal, the pass-catching-RB bump, now the age-27 cliff. Four real, true, well-supported findings. Four signals already sitting inside ppg0 or the age curve. The skill is not finding the signal. Everybody finds the signal. The skill is knowing when the price already moved.
The part I'm still not pretending to know
I'm settling one question, not two. The curve clearly prices the DIRECTION of the cliff, steep decline, right on time. Whether it prices the MAGNITUDE, specifically whether the 300-touch workhorse should fall harder than the curve's blanket 12 percent because his legs are more mileage-worn than a committee back's, is a different question. And it's the exact same wall the original post hit: to measure that residual honestly I need the age-neutralized projection surface that still does not exist. (It's the same blocker parking the route-runner WR signal in the queue.) So the magnitude question goes back in the notebook with a named blocker, not a fake answer. What I will not do is ship a crude flat penalty as a stand-in for calibration I have not done.
What it means for your roster
Same as last month, and that's the whole point. Sell the high-mileage back before the age-27 season. The cliff is real, and your board already knows it, the engine drops his projection on schedule whether or not the group chat has caught up. What is NOT true is that the engine has been asleep on aging backs and a patch is coming to dock them further. The haircut already happened. Trade on the cliff the board shows you, and stop waiting for a second one that would only ever be the first one, counted twice.
Receipts
- The curve itself:
engine_config.json->age_curves.RB(the -5.9% / -6.7% steps are right there) - The cohort harness:
scripts/h30_rb_age_27_cliff.py - The verdict path: confirms-existing, the signal already lives in the age curve
- The original take: 27 turns RBs into pumpkins. Especially the workhorses.