Everyone says 'set exit criteria' — nobody gives you the number. We measured it.
When to quit a fading effort, measured. Quit when recent yield falls ~60% below its peak (a drawdown stop) — an interior optimum (too early and too late both lose) that beat mining to depletion by +239% on the same budget in our reference model. θ≈0.6 is illustrative, not a universal constant.
The claim. "Set exit criteria and ignore the sunk cost" is the most repeated career and business advice there is — and it's useless, because it never tells you the threshold. When exactly do you cut a fading project, ad campaign, research line, content series, or sales channel and move on? We built the smallest model of a depleting effort and measured the answer.
The rule. Track the recent yield of the effort and its running peak. Quit when the recent yield has fallen a fraction θ below that peak — a drawdown stop, the same idea a trader uses to cut a losing position. In our reference model the optimum was θ ≈ 0.6 (give back ~60% of your best, then cut) — but treat that as illustrative, not a universal constant: it shifts with the depletion shape, the switch cost, and the horizon, so calibrate it on your own data.
The counterintuitive part — it's an interior optimum. Quitting isn't "as early as possible" or "never." In the reference model (M depletable veins, fixed effort budget), mining each vein to depletion yields 757 findings; the θ=0.6 drawdown rule yields 2,569, and the curve has a peak: θ=0.4 → 2,010, θ=0.5 → 2,113, θ=0.6 → 2,569, θ=0.7 → 2,366, θ=0.8 → 2,053. Read the +239% honestly: the baseline (mine every vein bone-dry) is the worst policy — it spends a long tail of digs on a near-empty vein — so almost any early-exit beats it; the number shows how bad never-quitting is, not that θ=0.6 is special. Against the textbook-optimal rule (leave when your marginal rate hits the habitat average) the advantage is far smaller — and we measured how much. A full-information, exact-assessment (MVT-oracle) rule beats never-quitting by about the same amount (≈220% median over 40 seeds; the +239% above is one seed, range ≈175–316%), so the model-free drawdown stop captures roughly 98% of the full-information optimum — within ~2%. That is the decision-relevant result: a rule you run on your own output stream, knowing nothing about the "habitat average", lands within a couple of percent of the rule that knows everything. Why both tails lose here: too-late (large θ) grinds the dead tail; too-early (small θ) exhausts the finite set of veins before the budget is spent (there is no per-switch cost in this model — so the optimum is interior because veins are finite, not because switching is expensive).
Why it works. The marginal yield of any depleting effort falls as you exhaust it. The drawdown stop is a cheap, model-free detector of "this vein has passed its useful life" — it doesn't need you to know the vein's true richness in advance, only to watch your own recent output relative to your best. That's why a fixed θ generalizes across very different efforts. This is the foraging logic of the marginal value theorem (Charnov 1976) — leave a depleting patch when its marginal yield falls toward the habitat average — and the broader optimal-stopping tradition. In fact our drawdown trigger is a known giving-up heuristic — the foraging literature has a whole family of these threshold proxies (giving-up time, Krebs 1974; optimal giving-up times, McNair 1982; stopping rules, Green 1984; giving-up density, Brown 1988), which approximate but only slightly under-perform the exact assessment rule (here, by ~2%). The drawdown stop is also the trader's trailing stop (Leung & Zhang 2019; under steady positive drift Glynn & Iglehart 1995 show the optimal trailing stop is to never stop — a real boundary case), and the decision-theoretic optimum for switching between depleting "arms" is the Gittins index, distinct from the foraging MVT. So nothing here is new theory: "nobody gives you the number" is about the popular advice, not the science; what we add is a model-free, runnable proxy (no need to know the habitat average) and a measured illustration.
The honest caveat. This decides when a declining effort has declined enough to cut — a drawdown stop on yield. It is not a forecast of whether a brand-new bet will pay off, and the window length and θ are levers you tune to your domain (defaults: 25 periods, θ=0.6 from the reference model). Sharper limits, from red-teaming our own result: (1) the rule has no model of why yield dropped — on a noisy signal it can quit a good vein on a temporary dip; it assumes near-monotone depletion. (2) It quits permanently, discarding the option to return — wrong for a vein that can replenish (a field reopens, a method matures). (3) The reference numbers are one synthetic model, single seed, no confidence interval. So treat this as a runnable heuristic to calibrate, not a law: it turns "don't throw good money after bad" from a proverb into a rule you can run — and then test.
We packaged it as quitkit: one zero-dependency file (plus an MCP server, so an agent managing a portfolio of efforts can ask "have we hit the drawdown stop?"). Open-core and free — a sibling of our memory, RAG-freshness, statistics and self-reference tools. should_quit(recent_yields) gives you a verdict and the reason.
FAQ
What's wrong with “set exit criteria and ignore sunk cost”? It never gives you the threshold. The advice is empty without a number, so we computed one: when to abandon a depleting effort and move on, as an explicit optimal-stopping rule.
Is mining a vein to depletion optimal? No. In the reference model (depletable veins, fixed effort budget) mining each to depletion yields 757 findings, but a stopping rule at the optimum (θ ≈ 0.6) yields 2,569 — about 3.4× more on the same budget. But that baseline (mine every vein dry) is the worst policy, so the gain mostly shows how bad never-quitting is; against the textbook-optimal rule (marginal rate = habitat average) the gain is far smaller, and θ≈0.6 is specific to this one synthetic model.
What is the threshold θ? θ is the fraction-of-recent-yield trigger to quit a vein and switch. The yield curve peaks near θ = 0.6 with a 25-period window; both are levers you tune to your domain rather than universal constants.
What can't this tell me? It is not a forecast of whether a brand-new bet will pay off. It optimizes when to quit a known, depleting effort — not whether to start an untested one.