When to quit a project: $25,000 kill rule vs open-ended fix
When to quit a project: $25,000 kill rule vs open-ended fix
| Takeaway | Detail |
|---|---|
| Sunk costs distort risk perception significantly | Founders overestimate turnaround chances after sinking funds compared to outside base rates. |
| Rational abandonment requires strict math | If a project's value falls to $65 million after $30 million spent, abandoning is better than spending an additional $70 million. |
| Persistence often defies economic logic | In investment decision studies, 68% of participants never abandoned a failing project across multiple decisions. |
| Predefined exits prevent emotional escalation | Setting clear criteria like stop-loss orders helps avoid the trap of escalating commitment seen in historical market failures. |
Factory math offers a stark alternative to heroic fixing. Consider a project originally projected to yield $120 million; if value drops to $65 million after $30 million is spent, rational actors must abandon rather than inject another $70 million. Continuing only makes sense if the remaining value justifies the final outlay, such as when projections hold at $75 million. Disciplined kill rules replace emotional attachment with cold arithmetic, ensuring resources flow to viable opportunities rather than dead ends.
Historical collapses, from Bear Stearns to Black Wednesday, illustrate the catastrophic cost of ignoring exit signals. Studies show 68% of investors persist in failing endeavors despite negative indicators, driven by mental budgeting errors rather than marginal analysis. To counter this, leaders must establish predefined thresholds and conduct regular reviews. Quitting faster is not failure; it is the disciplined application of judgment that preserves capital for projects with genuine potential.
Barry Staw's R&D allocation experiment defined the first loop: escalation of commitment. When subjects were told they had personally chosen the failing division, they allocated substantially more follow-on funding than subjects who inherited the same failure from someone else. According to Frontiers, in a replication with 110 undergraduate students completing a series of investment decisions regarding a failing endeavor, responsibility for the initial choice predicted persistence. The mechanism is not stupidity; it is self-justification. Admitting the division is failing feels like admitting you are a bad chooser, so you buy more evidence that you were right.
The Escalation Circuit
Daniel Kahneman and Amos Tversky's prospect theory gives that feeling a value function. Losses hurt more than equivalent gains feel good, so managers in the loss domain become risk-seeking. A sure loss if you quit today feels unbearable compared with a gamble that might — just might — erase the loss if you fund one more fix. According to Academia, 68% of participants never abandoning a failing project across decisions shows how stable that gamble becomes. According to the PDF: A sequential learning analysis of decisions in organizations, the two potentially costly errors in sequential investment decisions are sticking too long to a failing course of action and abandoning a successful one, and under loss framing leaders systematically choose the first error.
Identity locks it in. According to SSRN: R&D Capitalization and Reputation-Driven Real Earnings, managers responsible for initiating an R&D project anticipate reputation damage as a result of abandoning the project when expenditures are capitalized, and abandoning requires asset impairment when R&D expenditures have been capitalized. In other words, quitting writes down both the balance sheet and the biography. Owners therefore spend extra days rationalizing continuation to protect competence reputation rather than updating on evidence. The British government on Black Wednesday in 1992 is the public-market version: according to Trustnet, it kept the pound in the European Exchange Rate Mechanism despite market pressures, and the UK spent billions in reserves to prop up the pound before withdrawing with substantial losses.
According to CB Insights' review of startup post-mortems, many ran out of cash after prolonged fix attempts rather than a single shock. That is the base rate leaders ignore: terminal depletion rarely looks like a sudden accident, it looks like months of rescue spending that slowly drains the account to zero while the narrative stays alive.
As a judgment researcher, I read that as a prior, not an anecdote. According to Harvard Business School professor Shikhar Ghosh's study, most VC-backed ventures fail to return invested capital. If three out of four funded bets with selection, diligence, and coaching do not pay back, the prior for a struggling internal project that already needs a fix is even lower. A turnaround bet is not a 50-50 coin flip. It starts underwater.
According to the Project Management Institute's 2023 Pulse report, 12.1% of project dollars are wasted on troubled-project escalation where continued funding produced no deliverable recovery. That waste is not exploration. It is funding after the signal has turned negative, with no recovery to show for it. In decision terms, the mechanism is commitment under uncertainty: once you have paid to open, you keep paying to avoid booking the loss.
| Circuit Stage | What It Feels Like Inside | Outside-View Check | Ledger Figure |
| Self-justification | I chose it, so I must prove it | Would a new owner fund this? | 110 students tested per Frontiers |
| Loss aversion gamble | Do not realize sure loss | Price the gamble vs quit and redeploy | 68% never abandon per Academia |
| Near-miss attention | Almost working means keep going | Require pre-registered traction win | $100 million cost vs $120 million value per Wikipedia Sunk cost |
| Identity protection | Quitting signals incompetence | Separate decision quality from outcome | Impairment on abandon per SSRN |
| Calibration audit | One more fix will turn it | Enforce stop-loss benchmark, invest $0 until pass | $0 vs abandon distinction per PDF Reciprocal Relationships |
Base Rates of Failure
According to Startup Genome's 2024 scaling report, premature-scaling startups that kept fixing burned 3.2x more cash and reached Series A at only a low rate versus a higher rate for teams that killed or pivoted early. The killers did not just save cash. They more than doubled their advancement rate by redeploying time and attention to a testable alternative instead of defending a failing path.
The textbook logic for when to walk away is brutal and clarifying. According to Wikipedia's account of sunk cost, after $30 million is spent, if value projection falls to $65 million the company should abandon rather than spending an additional $70 million to complete it. According to that same account, if value projection falls to $75 million after $30 million spent, a rational actor should continue the project. Sunk dollars drop out. Only forward cost versus forward value matters. Real estate shows how hard that is in practice: according to Decision Physics, developers are more likely to inject additional capital into stalled projects after significant write-downs, precisely when the forward math is worst.
That persistence gap is measurable in the lab. According to Aihavit's summary of a 2024 study in Judgment and Decision Making that tracked 1,847 participants across commitment scenarios, moderate investors showed 34% higher persistence rates at the six-month mark versus minimal investment. You do not need a character flaw to escalate. Mere prior investment raises persistence, which is why a pre-registered traction win in a 30-day test has to do the deciding before motivation does.
The status-quo story to kill here is that adding rescue money saves what you already spent. Escalation science shows the opposite: rescue spending to protect sunk spending typically extends the loss path and roughly doubles expected loss, because you pay the fix cost and still inherit the low base rate. The skill is to replace hope with a ledger: forward cost to fix, forward value if the traction test hits, and the outside base rate if it does not.
Quitting wins when the kill is cheap, reversible, and pre-committed. As a decision scientist I model this as an entry-exit-scrapping choice: According to An Investment Model with Switching Costs and the Option to Abandon, a project operating in a random environment should be evaluated on its forward payoff rate as a function of current state variables, not on what you already paid to enter. That is the entire logic behind the kill rule in this article — cap additional exposure, demand a pre-registered traction win in a 30-day test, otherwise time-box the fix and redeploy.
The Kill arm in this comparison is a hard cap on additional spend gated by a pre-registered traction milestone, with automatic termination and redeployment if the gate is missed. You write the kill criteria before the test starts: what counts as traction, how it will be verified, and what happens on a miss. No committee review, no extension by optimism. The mechanism matters more than the exact calendar length — the article uses a short 30-day test inside a bounded evaluation window — because pre-registration prevents you from moving the goalpost after you see disappointing data.
The Fix arm is the default in most companies: an engineering rescue effort plus a contractor patch with no milestone gate and open-ended follow-on authorization. According to Aihavit, heavy investors persisted even when the approach clearly was not working, better alternatives existed, or continuing caused harm. That persistence is the mechanism of escalation. Once you authorize a fix without a gate, each additional week creates its own justification to continue, and the team learns to optimize for survival of the project rather than discovery of truth.
| Decision case | Forward math | Verdict |
| Abandon case | $30 spent, projection $65, need $70 to finish per Wikipedia account | Kill - forward cost exceeds forward value |
| Continue case | $30 spent, projection $75 per Wikipedia account | Time-box - forward value supports conditional continue |
| Persistence risk | 34% higher persistence for moderate investors per 2024 Judgment and Decision Making study | Pre-register test - do not trust feel |
| Portfolio prior | Most fail to return capital per Ghosh; many cash-out after prolonged fixes per CB Insights | Quit early wins - redeploy beats rescue |
| Escalation waste | 12.1% of dollars with no recovery per PMI 2023; 3.2x burn with lower vs higher advance per Startup Genome 2024 | Kill early wins - cutters advance faster |
Kill Rule vs Open-Ended Fix
The expected-value hurdle makes the asymmetry explicit. Take the article's working example: a fix with a large upside if turnaround succeeds, risking the capped additional amount to try. Break-even turnaround probability equals cost-to-try divided by upside-if-successful. In that example the required probability to break even is above eight percent, while the relevant base rate for this class of turnaround is only around four percent. Expected value equals probability times upside minus cost, so when required probability exceeds base-rate probability, fix has negative expected value. Do not spend another increment to rescue sunk cost — that belief roughly doubles expected loss because you pay new money for a below-hurdle gamble.
Apply the Annie Duke quit-test combined with the Jeff Bezos Type-2 reversibility screen to execute. Duke asks: would you start this project today knowing what you now know? Bezos asks: is this a one-way door or a two-way door? If killing preserves option value because talent, capital, and attention can be redeployed within about two weeks to a higher-EV bet, then kill is the reversible Type-2 decision. The fix is the irreversible lock-in. Choose the reversible kill, keep the abandonment option alive, and let the scrapping model do its work.
Explicit winner for negative-EV projects is the kill rule described above. Your next action: before authorizing any rescue, write the traction gate, verification method, and redeployment plan in one paragraph and enforce automatic termination if the gate is missed.
Nassim Taleb’s antifragility variance explains why a small share of software pivots produce 10x outliers that disappear in mean failure rates. Average expected value is misleading for highly optional bets because it dilutes the impact of extreme positive skew. Leaders must recognize that mean metrics obscure the structural advantage of preserving exposure to rare events.
Survivorship measurement bias skews published base rates downward. According to internal corporate audit data, many kills are never logged in post-mortem datasets. This omission creates an illusion of higher fix success rates by removing negative outcomes from the statistical pool. Decision-makers relying on public case studies face a distorted view of reality.
| Dimension | Kill Rule | Open-Ended Fix | Winner |
| Cost ceiling | Hard cap on additional spend, auto-stop on missed gate | No gate, follow-on authorization expands | Kill 3 - Fix 1 |
| Reversibility | Two-way door, talent redeployed in days | One-way lock-in, team trapped in rescue | Kill 3 - Fix 1 |
| Learning velocity | Pre-registered test yields fast true/false signal | Ambiguous progress, slow learning | Kill 3 - Fix 1 |
| Team morale load | Clean stop, redeploy to winnable work | Prolonged rescue burns out owners | Kill 3 - Fix 1 |
The kill rule misfires under extended regulatory moats or platform network effects with extended traction lags. Delayed-feedback environments require longer gates because immediate signals do not reflect final viability. Leaders must adjust time-boxes to match the latency of the specific industry’s feedback loops.
What the Data Doesn't Tell You
By the time the test had returned a clear failure signal, Quibi held only a small number of paying subscribers with high churn after the 3-month free trial expired. That is not slow growth; that is disconfirmation of the core mechanism. People did not want appointment viewing in 10-minute mobile chapters divorced from living-room screens and social sharing. In decision science we distinguish noise from a failed consumption hypothesis, and this was the latter. No retention curve fixes a product whose assumed use case does not exist.
| Scenario | Payoff Distribution | Kill Rule Impact |
|---|---|---|
| Fat-Tail R&D | Convex (rare outliers) | Truncates upside |
| Linear Process | Normal | Optimizes efficiency |
From judgment science, the failure is not optimism, it is revisability. Once a fix is underway, the brain reframes quitting as a loss and continuation as rescue. According to Kellogg / Heath Petersen, results support mental budgeting over escalating commitment, which is why the winning tactic is to remove the decision from the moment of pain entirely. You decide the exit before you fund the next sprint, in writing, with someone else holding the veto.
Start with pre-registration. Before any rescue dollars move, write one traction milestone and one kill-budget cap in a shared memo and assign an outside referee with veto power. No gate, no spend. That referee is not an advisor; they can stop the transfer when the gate is missed. This is mental budgeting made operational: separate accounts for exploration versus rescue, so a struggling project cannot borrow credibility from healthy work.
During the test window defined above, require a fifteen percent week-over-week lift in retention or revenue or kill automatically with no extensions. Flatlines count as failures. Leaders in 2026 routinely negotiate with flat data by adding another cohort, another channel, another week. Pre-commitment blocks that negotiation. If the curve is not bending, you are not learning, you are renting hope.
| Data Source | Logged Kills | Actual Kills | Bias Direction |
|---|---|---|---|
| Public Post-Mortems | Partial logging | Full set | Overstates Success |
| Internal Audits | Full logging | Full set | Accurate |
Price discipline comes next. Solicit two independent fix quotes and kill immediately if the cheapest verified quote exceeds the kill cap defined above, blocking incremental five-thousand-dollar drip extensions. Drips are how escalation hides. Each small approval feels rational, but the sum violates the cap you already judged rational. The same logic applies to rescue spend above one thousand dollars: impose a two-day cooling period plus a five-hundred-word pre-mortem memo assuming the fix failed. Write what broke, who ignored what signal, and what the money could have bought elsewhere. If you cannot write a convincing failure story, you do not understand the risk well enough to fund it.
Quibi's Collapse
The myth to kill is that topping up what is already sunk somehow saves the investment, when escalation science shows it roughly doubles expected loss. Sunk dollars are not protected by new dollars; new dollars are exposed to the same base rate that created the sunk dollars. Redeployment is the alternative. Within seven days of a kill, reassign eighty percent of team hours to the top base-rate alternative, measured by shipped learning per dollar. A growth team I discuss with students did this after a personalization fix stalled: one memo, one referee, two quotes, failed lift, kill, and the same engineers shipped two onboarding experiments the next week instead of debugging a third month.
By the time the test had returned a clear failure signal, Quibi held only a small number of paying subscribers with high churn after the 3-month free trial expired. That is not slow growth; that is disconfirmation of the core mechanism. People did not want appointment viewing in 10-minute mobile chapters divorced from living-room screens and social sharing. In decision science we distinguish noise from a failed consumption hypothesis, and this was the latter. No retention curve fixes a product whose assumed use case does not exist.
Escalation logic then offered what it always offers: a rescue quote. Re-cutting the content library for longer episodes and television viewing was priced at an additional rescue cost, plus added cost for an app rebuild to add TV casting, sharing, and social features, for a large total rescue cost. This is where leaders confuse sunk costs with salvage value. The idea that adding more rescue spending somehow saves what is already burned roughly doubles expected loss, because you convert a bounded failure into an unbounded commitment without new evidence.
Put it in expected-value terms and Kill dominates Fix. Give the turnaround a generous chance times a large upside, which equals a modest expected gain. Subtract the large fix cost and you get a large negative expected value. Fix does not break even; it destroys value in expectation versus quitting and redeploying talent and capital. When verified fix cost exceeds the pre-registered kill threshold without a pre-registered traction win in the test window, the rational move is time-boxed exit, not another tranche.
The counterfactual makes the savings concrete. Had leadership killed early in burn once the mobile-only hypothesis failed, instead of deploying the full war chest, substantial funds could have been preserved for redeployment to higher base-rate bets. That is the skill to take away: pre-register the traction win, price the fix before you feel the loss, and let the ledger decide. If cost-to-fix exceeds the threshold with no win, kill and redeploy.
| Quibi Option | Verified Figure | What It Means For Decision |
| Initial thesis | Raised funds for 7.4M subscribers | Mobile-only bet that required pre-registered test |
| Reality at test end | Small payer base, high churn | Core hypothesis failed, not slow start |
| Content re-cut | Quote to undo format constraint | Pay to undo original format constraint |
| App rebuild | Quote to chase TV plus social use case | Pay to chase TV plus social use case |
| Total rescue vs EV | Cost exceeds expected gain | Net negative, so Kill wins |
| Early-kill counterfactual | Stop early in burn, save substantial funds | Redeploy to higher base-rate bets |
How to Choose Well
From judgment science, the failure is not optimism, it is revisability. Once a fix is underway, the brain reframes quitting as a loss and continuation as rescue. According to Kellogg / Heath Petersen, results support mental budgeting over escalating commitment, which is why the winning tactic is to remove the decision from the moment of pain entirely. You decide the exit before you fund the next sprint, in writing, with someone else holding the veto.
Start with pre-registration. Before any rescue dollars move, write one traction milestone and one kill-budget cap in a shared memo and assign an outside referee with veto power. No gate, no spend. That referee is not an advisor; they can stop the transfer when the gate is missed. This is mental budgeting made operational: separate accounts for exploration versus rescue, so a struggling project cannot borrow credibility from healthy work.
During the test window defined above, require a fifteen percent week-over-week lift in retention or revenue or kill automatically with no extensions. Flatlines count as failures. Leaders in 2026 routinely negotiate with flat data by adding another cohort, another channel, another week. Pre-commitment blocks that negotiation. If the curve is not bending, you are not learning, you are renting hope.
Price discipline comes next. Solicit two independent fix quotes and kill immediately if the cheapest verified quote exceeds the kill cap defined above, blocking incremental five-thousand-dollar drip extensions. Drips are how escalation hides. Each small approval feels rational, but the sum violates the cap you already judged rational. The same logic applies to rescue spend above one thousand dollars: impose a two-day cooling period plus a five-hundred-word pre-mortem memo assuming the fix failed. Write what broke, who ignored what signal, and what the money could have bought elsewhere. If you cannot write a convincing failure story, you do not understand the risk well enough to fund it.
The myth to kill is that topping up what is already sunk somehow saves the investment, when escalation science shows it roughly doubles expected loss. Sunk dollars are not protected by new dollars; new dollars are exposed to the same base rate that created the sunk dollars. Redeployment is the alternative. Within seven days of a kill, reassign eighty percent of team hours to the top base-rate alternative, measured by shipped learning per dollar. A growth team I discuss with students did this after a personalization fix stalled: one memo, one referee, two quotes, failed lift, kill, and the same engineers shipped two onboarding experiments the next week instead of debugging a third month.
| Rule | Gate to check | Action if gate fails |
| 1. Pre-register referee | One milestone plus cap in writing, outside veto assigned | Release zero fix dollars until gate is defined |
| 2. Lift or kill | Fifteen percent week-over-week retention or revenue lift | Kill automatically, flatline equals failure, no extension |
| 3. Two-quote cap | Cheapest of two independent quotes versus kill cap above | Kill immediately, block five-thousand-dollar drips |
| 4. Cool plus pre-mortem | Two-day pause plus five-hundred-word failure memo for spend above one thousand dollars | Deny rescue spend until memo is filed |
| 5. Seven-day redeploy | Eighty percent of hours to top base-rate bet within seven days | Track shipped learning per dollar, close old workstream |
Also worth reading: Overconfident product bets 2026: AI pre-mortem vs vote cuts 20% bad bets: Overconfident product bets 2026: AI · The Speed Paradox: Japan's 402 Tbps and What It Doesn't Fix About Modern Life: Speed Paradox: Japan's 402 Tbps · The $15K-$240K Hiring Cost Myth and Anthropology's Onboarding Fix: $15K-$240K Hiring Cost Myth and
What to do next
| Step | Action | Why it matters | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Calculate the verified cost-to-fix and compare it against the kill threshold. | Sunk costs distort risk perception; founders overestimate turnaround chances after sinking funds. | ||||||||
| 2 | Check for a pre-registered traction win within the last 30 days of testing. | Rational abandonment requires strict math; without this win, the project must be killed. | ||||||||
| 3 | If the fix exceeds the kill cap and no traction exists, execute an immediate kill decision. | Persistence often defies economic logic; 68% of participants never abandoned a failing project in studies. | ||||||||
Frequently Asked QuestionsAt what specific forward value threshold should a company abandon a project after $30 million has already been spent? If the project's value falls to $65 million after $30 million is spent, rational actors must abandon rather than injecting an additional $70 million. Under what financial condition does it make sense to continue funding a project that has already incurred $30 million in costs? Continuing only makes sense if the remaining value justifies the final outlay, such as when projections hold at $75 million. What percentage of participants in investment decision studies never abandoned a failing project across multiple decisions? In investment decision studies, 68% of participants never abandoned a failing project across multiple decisions. How much cash do premature-scaling startups burn compared to those that kill or pivot early, according to the 2024 scaling report? Premature-scaling startups that kept fixing burned 3.2x more cash and reached Series A at only a low rate versus a higher rate for teams that killed or pivoted early. What percentage of project dollars are wasted on troubled-project escalation where continued funding produced no deliverable recovery? According to the Project Management Institute's 2023 Pulse report, 12.1% of project dollars are wasted on troubled-project escalation where continued funding produced no deliverable recovery. How did moderate investors' persistence rates compare to minimal investors in a 2024 study tracking 1,847 participants? Moderate investors showed 34% higher persistence rates at the six-month mark versus minimal investment. Quick answers
Sources: Reddit, arXiv, arXiv, Reddit, arXiv Research Methodology & Editorial StandardsWe begin by defining the specific objectives the reader needs to accomplish. Primary product documentation and authoritative secondary sources are assembled into a verified research corpus; drafting occurs only after this foundation is in place. Every quantitative claim is subjected to dual-source verification. Any figure that cannot be independently corroborated is either qualified or omitted. Judgment Call Podcast Essays for people who make the callTechnology, philosophy, and society — long-form analysis for high-stakes judgment under uncertainty. Browse latest essays |