Product Management for Engineers · Page 3

What to Build and When: Appetite, Cost of Delay, Backlogs and Roadmaps

Admitting ideas, ordering them, sizing them and promising dates are four separate decisions that most methods bundle. This page takes the bundles apart: what RICE, cost of delay, appetite-and-bet, backlog attrition, cycles and outcome roadmaps each compute, what each optimizes, and where the arithmetic or the evidence stops. Afterwards you can choose a method for a given team, stress-test a RICE ranking, state what a date commitment costs, and say what happens to an idea that is not ranked.

Deep dive 60 min read 13 sources As of 2026-10-26
A queue of ideas with four decision points: admit, order, size, promise ideas requests bugs 1 Admit Sec. 6 backlog, or "maybe some day" waiting items 2 Order Sec. 2, 4 score, delay cost, or a bet Team capacity Sec. 5, 7 one team, one cycle shipped 3 Size each item estimate, or appetite · Sec. 3 4 Promise to people outside the team date, cycle, or objective · Sec. 8
Figure 1. The frame this page uses (the writer's, not a source's): ideas arrive faster than a team can ship them, so something admits or discards them, something orders what remains, something sizes each item, and something is promised to people outside the team. Each method in the sources answers a different subset of these four; section numbers show where each valve is taken apart. The capacity unit on the right is Shape Up's: one designer with one or two programmers for one six-week cycle1.

1. Prioritization is four decisions, not one

A list sorted by a score hides how much was decided before the sort. Something already decided which ideas were allowed onto the list (admission). The score or a meeting then decides order. Someone decided how much of the team each item would get (size), and afterwards someone tells other people what to expect (promise). The methods on this page differ less in how they order than in which of the four decisions they answer themselves and which they take as an input.

RICE answers order only. It takes size as an input, an effort estimate in person-months, and says nothing about admission or promise2. The cost-of-delay rule from Reinertsen's flow-based product development also answers order, with job duration as its input34. Shape Up answers all four: admission by a soft "maybe some day" and no backlog, order by a short meeting of the top people, size by a time budget fixed before design, and promise as a six-week uninterrupted bet1. Linear keeps a manageable backlog, runs n-week cycles and holds a roadmap of themes56. The outcome roadmap from Silicon Valley Product Group (SVPG) changes the unit that is ordered, from features to business objectives, and treats a date as a special, paid-for case789.

1.1 The supply side is small and rigid

The right side of Figure 1 is a few team-cycles. Shape Up's standard project team is one designer and two programmers, or one designer and one programmer, joined later in the cycle by a QA person for integration testing; the cycle is six weeks, followed by a two-week cool-down1 (Ch. 8, "Team and project sizes"). The authors name the alternative "Calendar Tetris": when people are free at different times because projects overlap, scheduling turns into a puzzle, and a fixed cycle gives "a standard project size both for shaping and scheduling"1. Every method below is a way to spend a small number of such units on a much larger set of candidates. None of the sources says how large that set is for a typical team. The only counts anywhere in them are Bing's: the headline change was one of "the hundreds suggested" and Bing "runs over 10,000 experiments a year"10 (Ch. 1, pp. 3, 5), a different unit of work from a six-week project.

1.2 What kind of evidence this page rests on

Two results below are mathematics: the sequencing rule for two jobs with delay costs (Section 4) and the mean waiting time of a single-server queue (Section 5). Everything else is a practitioner describing their own practice: Intercom on RICE, 37signals on Shape Up, Linear on its method, SVPG on roadmaps, and Reinertsen on survey answers from his own courses. The one measured outcome is a single Bing experiment (Section 6). No fetched source compares these methods on delivery speed, revenue or team morale, so "which method works" is answered below by mechanism and by stated failure mode, never by a result.

2. RICE: scoring impact per person-month

2.1 Intercom's definition and what its scales impose

Intercom's product team developed RICE as "our own scoring system for prioritization"; the post describing it is dated 5 January 20182. It scores each candidate on four factors. Reach is "number of people/events per time period", such as customers per quarter, taken "from product metrics instead of pulling numbers from a hat". Impact is the effect on one person, chosen from a "multiple-choice scale: 3 for 'massive impact', 2 for 'high', 1 for 'medium', 0.5 for 'low', and finally 0.25 for 'minimal'", and the number is a multiplier. Confidence is "100% is 'high confidence', 80% is 'medium', 50% is 'low'. Anything below that is 'total moonshot'", included "to curb enthusiasm for exciting but ill-defined ideas". Effort is person-months across product, design and engineering, in whole numbers or 0.5, and "divides the total impact"2. The formula is an image on the page; the prose gives it as Reach × Impact × Confidence ÷ Effort and describes the result as "total impact per time worked"2.

The post justifies the four factors by four temptations: working on pet ideas you would use yourself instead of broad-reach ones, on clever ideas instead of goal-linked ones, on exciting new ideas instead of ones you are confident about, and discounting the extra effort one project needs2. Each factor patches one temptation. Intercom's author is also explicit about the limit of the method: "RICE scores shouldn't be used as a hard and fast rule"2.

Table 1. RICE's scales as Intercom defines them2, with the step ratios between adjacent rungs (derived: ratios of the printed values).
FactorValues allowedUnitRatio between adjacent rungsSpan, top to bottom
ReachAny count for a chosen periodPeople or events per periodcontinuousopen
Impact3, 2, 1, 0.5, 0.25Multiplier per person1.5, 2, 2, 212×
Confidence100%, 80%, 50% (below: "moonshot")Percent1.25, 1.62×
EffortWhole numbers, or 0.5Person-months1→2: 2; 2→3: 1.5; 4→5: 1.25open

2.2 The worked example, with the scores the post does not print

The post scores three projects and shows the reasoning for each input but never prints the products. Project 1 is a signup-funnel change: 500 customers per month reach that step and 30% choose the option, so reach is 500 × 30% × 3 = 450 customers per quarter; impact 3; confidence 100% ("quantitative metrics for reach, user research for impact, and an engineering estimate for effort"); effort 2 person-months. Project 2 changes something every customer who uses the feature sees: reach 2,000 per quarter; impact 1; confidence 80% ("I'm unsure about the impact"); effort 4. Project 3 is "a one-time effect on 800 existing customers, with no ongoing effect": reach 800; impact 2; confidence 50% ("The reach and impact may be lower than estimated, and the effort may be higher"); effort 12. The scores below are derived.

Table 2. Intercom's three example projects2 with the RICE scores derived by the writer: \( \text{score} = R \times I \times C / E \).
ProjectReach (per quarter)ImpactConfidenceEffort (person-months)Score (derived)Rank
14503100%26752
22,000180%44003
3800250%18001

The project the author rates least certain ranks first. The mechanism is the divisor: Project 3's effort is the smallest, and its 50% confidence halves a raw 1,600 to 800 without removing the lead. The author's own justification for that 50% flags all three of Project 3's estimates as probably wrong in the direction that lowers its score. Derived: if reach were 20% lower, impact one rung lower (2 to 1) and effort 2 instead of 1, the score would be 800 × 0.8 × 0.5 × 0.5 = 160, below Project 2's 400. The confidence factor is a single scalar applied to a ratio of three estimates, and errors in a ratio multiply; the scale has no rung for "several things are probably off at once".

2.3 Stress tests (derived)

Margin against resolution. The gap between first and second place is 800 ÷ 675 = 1.185. The smallest step between adjacent rungs on the Impact and Confidence scales is 1.25 (100% to 80%, Table 1). A disagreement of one rung on Project 3's impact (2 to 1), or of one whole person-month on either leading project's effort, is therefore enough to reorder the top two. Table 3 gives, for each input of the leading projects, how far it has to move before the winner changes.

Table 3. How far one input must move before the winner of Table 2 changes (derived; all other inputs held at Intercom's values2).
Input movedPrinted valueWinner changes atRelative moveIs that one rung?
Project 3 confidence50%below 42.2% (ties Project 1 at 675)-15.6%No rung exists between 50% and "moonshot"
Project 3 impact2below 1.69-15.6%Yes: next rung is 1, score 400
Project 3 effort1above 1.19+18.5%Yes: next whole number is 2, score 400
Project 3 reach800below 675-15.6%Continuous
Project 1 effort2below 1.69-15.6%Yes: next whole number is 1, score 1,350
Project 1 reach450above 533+18.5%Continuous
Project 2 impact12 gives a tie at 800+100%Yes: next rung up
RICE scores of three example projects and Project 3's score as its confidence falls Scores from Intercom's inputs Project 3 score against its confidence 250 500 750 1,000 675 400 800 Project 1 Project 2 Project 3 RICE score = impact per person-month (derived) below 50%: "moonshot" 400 800 1,200 1,600 Project 1 = 675 Project 2 = 400 800 at 50% 42.2% 25% 100% 80% 50% 0 Project 3 confidence (Intercom's rungs marked)
Figure 2. Left: the scores of Table 2. Right: Project 3's score is \( 1600 \times C \) (reach 800, impact 2, effort 1). It falls to Project 1's 675 at \( C = 42.2\% \) and to Project 2's 400 at \( C = 25\% \). Intercom's scale has rungs at 100%, 80% and 50% and calls everything lower a moonshot2; both crossings lie in that unlabeled region. All scores are derived from Intercom's inputs.

The reach window. Reach is "per time period", and the three projects differ in what happens in the next period. Projects 1 and 2 recur: the funnel is measured monthly and Project 2 reaches every customer who uses the feature each quarter. Project 3 is a one-time effect on 800 existing customers2. Count \( n \) quarters and, assuming the recurring effects are stationary, the scores become \( 675n \), \( 400n \) and a fixed 800. Project 3 leads only while \( 800 > 675n \), that is \( n < 1.185 \) quarters. At \( n = 2 \) Project 1 scores 1,350 and Project 2 ties Project 3 at 800; at \( n = 4 \), a year, the one-time project falls from first to last: 2,700, 1,600, 800. The post chooses a quarter and does not discuss recurrence2, so the window is a hidden degree of freedom that can decide the ranking.

RICE with a counting window (derived; defaults are Project 3, the others are Table 2)

This project's score
Project 1 (recurring)
Project 2 (recurring)
Rank among the three

Formula: \( \text{score} = R_{\text{eff}} \cdot I \cdot C / E \) with \( R_{\text{eff}} = n R \) if the effect recurs, else \( R \). Projects 1 and 2 are held at their Table 2 inputs and assumed to recur. Set confidence to 42% against 43% to see the Figure 2 crossing.

2.4 What the score cannot see

Dependencies and table stakes. Intercom names both: "One project may be a dependency for another project, so it needs to happen first, or another feature might be 'table stakes' to sell to certain customers." Its remedy is bookkeeping, not modeling: "With a scoring system in place, you can clearly identify when you're making these trade-offs"2. The score is a record that makes an out-of-order decision visible.

Impact is a guess. "Choosing an impact number may seem unscientific. But remember the alternative: a tangled mess of gut feeling"2. The five rungs force a choice but add no information. The scale's 12× span (Table 1) is a convention: nothing in the post says "massive" is twelve times "minimal".

Confidence is not a probability. Intercom ties each rung to what evidence the author has ("data to support the reach and effort, but I'm unsure about the impact")2. That describes the quality of the inputs. Multiplying impact by it reads as an expected value only if the rung were the probability that the project works. This observation is the writer's, not Intercom's or Kohavi's: the measured base rates in the experimentation literature, where a third of ideas tested at Microsoft improved their target metric and 10-20% did on well-optimized Bing and Google measures10, are a different quantity, and RICE has no mapping from its rungs to such a rate. A "100% confidence" project is one with good data behind its estimates, not one that is certain to succeed.

Effort is an estimate made before design. That is the input appetite-based planning declines to take, and the next section is about the difference.

3. Appetite versus estimate: which way the number flows

3.1 Two directions

Shape Up states the difference in one line: "Estimates start with a design and end with a number. Appetites start with a number and end with a design"1 (Ch. 3, "Fixed time, variable scope"). An appetite is "a time budget for a standard team size", set in two sizes: a small batch, which one designer and one or two programmers can build in one or two weeks, and a big batch, the same team for a full six-week cycle. If the scope seems too large for six weeks, the shapers first try to narrow the problem and, failing that, "break off a meaningful part of the project that we can shape to a six-week appetite"1. The principle is "fixed time, variable scope": the authors' own illustration is a book deadline, where with one week left they can fix typos or add a section and the fixed date forces the choice1.

The calendar case in Chapter 2 shows the budget doing work. Customers asked for a calendar; "it can easily take six months or more to build a proper calendar", and earlier Basecamp versions had calendars that "only about 10% of customers used". That is why "we didn't have the appetite for spending six months". With one six-week cycle, "we could only build about a tenth of what people think of when they say 'calendar.' The question became: which tenth?"1. Page 2 follows the framing work that answered it.

Estimate-first starts from a design and ends in a number; appetite-first starts from a number and ends in a design Estimate-first Fix a design Estimate it A number: 6 months+ scope: fixed time: floats Calendar case: "it can easily take six months or more to build a proper calendar" Appetite-first Set the number Shape a design A design: 6 weeks time: fixed scope: floats Calendar case: "we could only build about a tenth of what people think of when they say 'calendar'" At the budget: appetite-first stops by default (Sec. 7); the RICE post does not address overruns
Figure 3. The two flows, with the quoted Basecamp calendar case1 (Ch. 2 case study, Ch. 3). Blue marks the dimension held fixed, orange the one that absorbs surprise. The RICE post2 takes effort as an input and says nothing about what happens when work exceeds it.

3.2 What sets the number

No formula. Chapter 3 says the appetite answers questions such as whether an idea deserves "a quick fix if we can manage", "an entire cycle", or only "a minor tweak"1. At the betting table, Chapter 9 lists five questions that people "might hear" while debating bets: does the problem matter, is the appetite right, is the solution attractive, is this the right time, and are the right people available1. None is scored. A stakeholder who rejects an appetite may be rejecting something else, and the chapter's probe is "How would you feel if we could do it in two weeks?", which can surface an objection such as "I don't want to introduce another dependency into that area of the app"1.

So the two methods place the same judgment, how valuable is this, in different places. RICE writes it into Reach, Impact and Confidence and sorts. Shape Up writes it into one number, the time the idea deserves, and argues it. Neither is more objective; one is recorded on a scale and the other is debated.

Shape Up also moves a filter upstream. It runs "two separate tracks: one for shaping, one for building"; shaping is kept private until a bet is made, which "gives the shapers the option to put work-in-progress on the shelf or drop it"1 (Ch. 2, "Two tracks"). Someone therefore chooses which raw ideas receive shaping time. Chapter 2 says only that shaping needs "a generalist or ... one or two other people" and Chapter 6 that pitches are posted for people on the betting table to comment on; neither says who picks what to shape or by what rule1. On the writer's reading, prioritization has not disappeared; it has moved into the shapers' attention, and the betting table chooses among pitches that already survived it. The cost of a pitch that loses is its shaping time, which the sources do not quantify.

3.3 Why not estimate first

Shape Up gives one mechanism: "the more specific the work is, the harder it can be to estimate", because "making the interface just so can require solving hidden complexities and implementation details that weren't visible in the mockup. When the scope isn't variable, the team can't reconsider a design decision that is turning out to cost more than it's worth"1 (Ch. 2). Reinertsen supplies a different, independent observation: in his courses, 95% of surveyed managers admit they begin design before all requirements are known, and "the average product developer begins design when 50 percent of requirements are known"3 (Ch. 1, p. 2). An estimate is made, in practice, with about half the requirements still open. He also states the structural alternative in one sentence about cadence: the orthodoxy "constrains the scope of work and drives variability into timing", and the new paradigm "constrains timing and drives variability into the scope of work"3 (p. 11). Shape Up does not cite Reinertsen; the match is the writer's observation. An appetite does not remove the uncertainty. It moves it from the date to the scope.

Table 4. Estimate-first and appetite-first planning, as the sources describe them12.
QuestionEstimate-first (RICE's effort)Appetite-first (Shape Up)
InputA design or scope descriptionA time budget for a standard team
OutputPerson-months, in whole numbers or 0.5A design that fits the budget
What absorbs a surpriseNot addressed in the postScope; by default the project stops at the budget
Where value is judgedReach, Impact, Confidence, scoredThe appetite, argued at the betting table
WeaknessShape Up's objection: a more specific design is harder to estimateThe writer's reading: needs a shaper who is "technically literate" and can design to a budget; no data on how often shaped work fits

4. Cost of delay: sequencing by what waiting costs

4.1 What Reinertsen's first chapter claims

Chapter 1 of The Principles of Product Development Flow is a polemic: "the dominant paradigm for managing product development is fundamentally wrong"3 (p. 1). The first of its twelve listed problems is "Failure to Correctly Quantify Economics" (Figure 1-1, p. 4). Developers, he says, optimize proxy variables such as cycle time and percent value-added time, but "when you ask them how much life-cycle profit will decrease due to a week of delay, they don't know"3 (p. 5). He wants every choice expressed in one unit, life-cycle profit impact, so that interacting proxies (higher utilization raises efficiency and lengthens cycle time) can be netted.

The numbers in the chapter come from the author's own course surveys, not from a representative sample. "Only 15 percent of product developers know the cost of delay associated with their projects" (p. 4). In a footnote he adds that at companies that do not calculate it, answers "have such high variance that they simply represent quantified ignorance. On average, ... people working on the same project will give answers that vary by 50 to 1"3 (footnote 1, p. 5).

On ordering, the chapter makes three claims. FIFO "is almost never economically optimal in product development" because projects "almost always have different delay costs, and they present different loads on our resources" (p. 2). Ranking by project profitability is "just an illusion": "we should not prioritize on the basis of project profitability, but rather on how this profitability is affected by delay" (p. 14). And minimum-slack-first, the rule behind Goldratt's critical chain, "can encourage you to apply resources to a low cost-of-delay job with a depleted buffer in preference to a high cost-of-delay job that is ahead of schedule. This is the opposite of what economics would suggest" (p. 14)3. He names computer operating systems as a source of better sequencing, since they "have developed much more advanced sequencing methods than the simple FIFO sequences used in manufacturing processes" (p. 23)3. The next subsection derives the rule that this line of argument leads to.

4.2 The two-job exchange argument (derived)

Setup: one team does one job at a time without preemption. Job \( i \) takes \( d_i \) weeks, and every week before it is delivered costs \( c_i \) dollars. Running A first delivers A at \( d_A \) and B at \( d_A + d_B \):

$$ \text{cost}_{AB} = c_A d_A + c_B (d_A + d_B), \qquad \text{cost}_{BA} = c_B d_B + c_A (d_A + d_B). $$

The difference is \( \text{cost}_{AB} - \text{cost}_{BA} = c_B d_A - c_A d_B \). Running A first is cheaper exactly when

$$ \frac{c_A}{d_A} > \frac{c_B}{d_B}. $$

Any order with an adjacent pair out of that sequence is improved by swapping the pair, so sorting by \( c_i / d_i \), delay cost per unit of duration, is optimal for any number of jobs. This is Smith's rule (weighted shortest processing time first), the classical result for minimizing total weighted completion time on a single machine, proved in scheduling textbooks by adjacent pairwise interchange11. Engineers know its unweighted form as shortest-job-first.

Delay cost as area: running job A first costs 170 thousand dollars, running B first costs 140 thousand Run A first Run B first $k per week of delay $k per week of delay A waits 3 wk × 30 = 90 B waits 4 wk × 20 = 80 0 1 2 3 4 weeks Total 170 A (3 weeks) B A waits 4 wk × 30 = 120 B waits 1 wk × 20 = 20 0 1 2 3 4 weeks Total 140 B A (3 weeks)
Figure 4. Derived illustration, hypothetical inputs: A takes 3 weeks and costs $30k per week of delay; B takes 1 week and costs $20k. Each project's delay cost is the area of its bar (rate times the time until delivery). A has the higher rate but a lower rate-per-week-of-work (10 against 20), so running B first saves $30k, which is \( c_B d_A - c_A d_B = 20 \times 3 - 30 \times 1 \). The ordering rule is Smith's11; the numbers are not data.
Table 5. Three simple ordering rules on two job pairs (derived; hypothetical inputs). A: 3 weeks, $30k per week; B: 1 week, $20k; C: 1 week, $1k. Totals are delay cost in $k; best in green, worst in red.
PairA firstOther firstHighest rate firstShortest firstHighest rate per week of work first
A and B (ratios 10, 20)170140A first: 170B first: 140B first: 140
A and C (ratios 10, 1)94121A first: 94C first: 121A first: 94

Neither "biggest delay cost first" nor "shortest first" is right on both pairs; their ratio is. The same quantity decides a dilemma that RICE-style value ranking handles awkwardly: a small perishable item against a large one. The calculator lets you move the four inputs.

Two-job sequencing (derived; inputs are knobs, not data)

A first
B first
Saved by A first (negative: B first is cheaper)
Rate per week of work: A
Rate per week of work: B

Formula: \( \text{saved by A first} = \text{cost}_{BA} - \text{cost}_{AB} = c_A d_B - c_B d_A \). A first is better when this is positive, which is the same as \( c_A / d_A > c_B / d_B \). Assumes one team, no preemption, linear delay cost.

4.3 WSJF as SAFe states it, and how it differs from RICE

The Scaled Agile Framework defines Weighted Shortest Job First as "a prioritization model used to sequence work for maximum economic benefit" and gives the estimate in prose: "In SAFe, WSJF is estimated as the relative cost of delay divided by the relative job duration". Its inputs are "relative user and business value, time criticality, risk reduction and/or opportunity enablement, and job size"4. That is the Section 4.2 ratio with both sides scored on a relative scale. The page opens with an epigraph attributed to Reinertsen's book, "If you only quantify one thing, quantify the Cost of Delay"4; the sentence itself was not found in the fetched book text. The book's table of contents lists "Sequencing Work" on p. 191 of Chapter 73; that it derives the same ratio rule is not confirmed here.

RICE and WSJF look alike, since both divide a benefit by a cost of capacity. Three differences matter, all derived from the formulas.

  • What is divided by. RICE divides by effort in person-months2, a measure of resource consumed. The exchange argument needs the calendar time the constrained team is occupied. The two coincide only if every project uses the same team size.
  • Urgency. RICE has no term for value that disappears after a date. WSJF's "time criticality" input is where such a deadline would go4.
  • Value delivered versus value lost to waiting. The sequencing rule needs the value lost per week of delay. For a benefit stream that starts when the project ships, a week's wait forgoes a week of it, so the two readings agree. For a one-time benefit of fixed size, delay merely postpones it. Intercom's Project 3 is a one-time effect on 800 customers2. Unless their problem expires or they leave, its cost of delay is small, and the sequencing rule would place it behind the recurring Projects 1 and 2 that RICE puts it ahead of. The post does not say whether the problem is perishable.

4.4 Where the rule is exact and where the inputs are not

The input is the problem. In Table 5 the two ratios of the first pair differ by a factor of 2. A factor-of-2 error in either project's delay cost produces a tie, and anything larger reverses the order. Reinertsen's footnote reports answers that "vary by 50 to 1" for the same project when cost of delay is not calculated3; at that spread the rule's output is set by who answered. The footnote gives no figure for how much calculating narrows the spread, so the claim that a shared money-per-week unit helps is his argument, not a measured result.

A fixed appetite fixes the denominator. Under Shape Up a bet's duration is its appetite, two or six weeks, by construction1. Among bets of equal appetite the rule reduces to ordering by delay cost, and no duration estimate is needed. This composition of appetite and delay cost is the writer's, not a source's.

Assumptions. The derivation assumes one resource, known durations, linear delay cost, no preemption and no dependencies. Real teams break all five. The contents list later chapters on queues (Chapter 3), variability (Chapter 4) and flow under uncertainty (Chapter 7)3, which were not available for this page, so what they say about these assumptions is not verified here. The rule is a clean baseline for asking what order costs, not a procedure that survives contact with a sprint board.

5. Queues: why a fully loaded team is a slow team

5.1 What Reinertsen says about queues

"Few developers realize that queues are the single most important cause of poor product development performance"3 (Ch. 1, p. 5). The inventory in such a queue is "design-in-process": financially invisible, because R&D is expensed and "if we ask the chief financial officer how much inventory we have in product development, the answer will be, 'Zero'", and physically invisible, because "our inventory is bits on a disk drive" (p. 6). "Today, only 2 percent of product developers measure queues", and the link to Section 4 is explicit: "Without a cost of delay, queues appear to be free and therefore, unworthy of attention" (p. 6)3. On utilization: "Executives coming to my product development classes report operating at 98.5 percent utilization in the precourse surveys", and once capacity cost and delay cost are put in the same unit, "we will always conclude that operating a product development process near full utilization is an economic disaster" (p. 7)3. The 98.5% is a self-reported survey answer from his own classes.

5.2 The delay curve of a single-server queue (derived)

Chapter 1 defers the mathematics to Chapter 3, which was not available. The standard result is short. Take one server, Poisson arrivals at rate \( \lambda \), exponential service with mean \( 1/\mu \), utilization \( \rho = \lambda/\mu < 1 \) (the M/M/1 queue). Balancing the flow between neighboring states gives \( \lambda p_{n-1} = \mu p_n \), so \( p_n = (1-\rho)\rho^n \). The mean number in the system is \( E(L) = \rho/(1-\rho) \). Little's law \( E(L) = \lambda E(S) \), which assumes the number in the system does not grow without bound, then gives the mean time in the system and the mean wait before service12 (§3.4, §4.2-4.3, eq. 4.9):

$$ E(S) = \frac{1/\mu}{1-\rho}, \qquad E(W) = E(S) - \frac{1}{\mu} = \frac{\rho/\mu}{1-\rho}. $$

In units of the mean service time, the wait is \( \rho/(1-\rho) \). That is Table 6, and it is the same hockey stick as mean latency against load on any single-queue server. Going from 95% to 98.5% utilization multiplies the wait by 65.7 ÷ 19, about 3.5.

Table 6. Derived from the M/M/1 formulas12 and the Kingman factor13 of Section 5.3. Waits are in multiples of the mean service time; the last column uses a mean item size of 2 weeks, a hypothetical.
Utilization \( \rho \)Wait, M/M/1Time in systemWait, fixed-size items (× 0.5)Wait, high size variance (× 1.5)Wait, M/M/1, 2-week items (weeks)
50%1.02.00.51.52
80%4.05.02.06.08
90%9.010.04.513.518
95%19.020.09.528.538
98.5%65.766.732.898.5131
99%99.0100.049.5148.5198
Mean wait in a single-server queue against utilization, rising steeply near 100 percent, for three levels of job-size variability 25 50 75 100 mean wait, multiples of the mean service time 98.5 65.7 32.8 high size variance (factor 1.5) exponential sizes, M/M/1 (factor 1) fixed-size items (factor 0.5) 0 50% 80% 90% 95% 98.5% utilization ρ
Figure 5. Derived, not measured: mean queueing delay of a single server in multiples of the mean service time, \( \frac{\rho}{1-\rho}\cdot\frac{c_a^2 + c_s^2}{2} \), with Poisson arrivals (\( c_a^2 = 1 \)) and job-size variance \( c_s^2 \) of 0, 1 and 21213. The orange point marks the 98.5% utilization that executives reported in Reinertsen's precourse surveys3 (p. 7); the dashed line is the same abscissa. The model says nothing about whether product development behaves like this queue.

5.3 Variability multiplies the wait

For a general single-server queue, Kingman's approximation scales the M/M/1 wait by the average of the squared coefficients of variation of interarrival and service times, \( E[W] \approx \frac{\rho}{1-\rho}\cdot\frac{c_A^2 + c_B^2}{2}\cdot E[B] \); the same lecture notes give Daley's tighter upper bound13 (§5.6.3). For M/M/1 both coefficients are 1. If every item took exactly the same time, \( c_B^2 = 0 \), and with Poisson arrivals the wait halves at every utilization; if sizes are widely spread (\( c_B^2 = 2 \)), it rises by half. Reinertsen's chapter names the same lever without the formula: "reducing batch size is usually the single most cost-effective way to reduce queues", because smaller batches reduce "unnecessary variability in flow"3 (pp. 17-18). Shape Up's standard sizes, a one-to-two-week small batch or a six-week big batch for a fixed team shape1, narrow the spread of item sizes. Shape Up itself makes no queueing argument; the connection is the writer's.

Queue wait and what it costs (derived; inputs are knobs, not data)

Mean wait before work starts
Delay cost per item, from the queue alone
Items waiting on average

Formula: \( W \approx \frac{\rho}{1-\rho}\cdot\frac{c_A^2+c_B^2}{2}\cdot \tau \), delay cost \( = c \cdot W \), items waiting \( = \lambda W \) with \( \lambda = \rho/\tau \) (Little's law). The queue wait excludes the item's own work time.

5.4 Slack by design

Both Shape Up and Linear leave capacity unallocated. Shape Up's two-week cool-down is "a period with no scheduled work"1: 2 of every 8 weeks, or 25% of calendar time (derived). Linear tells teams to "reserve some space on your product timelines for unplanned work that comes up unexpectedly" and "don't overload cycles with tasks"5. Their stated reasons are practical. The end of a cycle "is the worst time to meet and plan because everybody is too busy finishing projects", and teams need time to "breathe"1 (Ch. 8). Neither makes a queueing argument. On the writer's reading, unscheduled capacity is what lets a queue with random arrivals drain, and Shape Up goes further by taking arrivals away from the builders: "We do not allow the team to be interrupted or pulled away to do other things"1. The random stream meets the betting table, and the table's slate is cleared each cycle.

6. Backlogs, attrition and the Bing counter-case

6.1 Two policies that let ideas die

Shape Up rejects the backlog on cost grounds. "Dozens and eventually hundreds of tasks pile up that we all know we'll never have time for. The growing pile gives us a feeling like we're always behind even though we're not", and "the time spent constantly reviewing, grooming and organizing old ideas prevents everyone from moving forward on the timely projects that really matter right now"1 (Ch. 7). The default answer to a new idea is "Interesting. Maybe some day.", "a very soft 'no' that leaves all our options open", and explicitly "We don't put it in a backlog"1 (Ch. 3). The betting table sees only the last six weeks of pitches "or any pitches that somebody purposefully revived and lobbied for again"; a pitch not chosen is let go, with "nothing we need to track or hold on to". Support, product and programmers may each keep their own lists, but "none of these lists are direct inputs to the betting process"1.

What replaces the backlog is a filter on re-arrival: "Really important ideas will come back to you." A bug that matters "will come back to your attention when a customer complains again or a new customer hits it. If you hear it once and never again, maybe it wasn't really a problem"1. The premise is that how often an idea returns tracks how much it matters. The chapter offers no evidence for the premise and does not discuss users who are affected but never report.

Linear states nearly the same policy but keeps a list: "You don't need to save every feature request or piece of feedback indefinitely. Important ones will resurface, low priority ones will never get fixed. A more focused backlog makes it easier and faster to plan cycles, and ensures that work will actually get done"5. It also keeps feedback as "a research library" attached to issues5, so what Linear caps is the planning backlog, not the record of what users said. Both policies converge on the fate of a low-priority item, which is to stay undone. They differ in who carries the cost of the list.

6.2 A backlog is a queue, so its size has a waiting time (derived)

Little's law, \( E(L) = \lambda E(S) \), holds when the number in the system stays bounded12 (§3.4). Applied to a backlog, mean residence time equals items waiting divided by the rate at which items leave. Illustration with hypothetical numbers: a backlog of 120 items from which 3 items complete per two-week cycle implies 40 cycles, or 80 weeks, of mean residence for items that ever complete, provided the backlog is stable and nothing is deleted. If problems stay relevant for less than that time, the tail of the list is not waiting; it is expiring. This is the quantitative content of the Shape Up and Linear claims, and the check is simple: backlog size over completion rate, compared with how long an item stays worth doing. Reinertsen's frame would call the same list a queue with a cost3; the chapter available here does not say whether he would cap it or delete from it, and this page does not assert either.

6.3 The counter-case: an idea that sat for six months

Chapter 1 of the experimentation textbook opens with the strongest argument against trusting a ranking. In 2012 a Bing employee suggested lengthening ad titles by combining them with the first line of ad text. "The feature was prioritized low and languished in the backlog for more than six months until a software developer decided to try the change, given how easy it was to code." The experiment raised revenue by 12%, "which at the time translated to over $100M annually in the US alone, without significantly hurting key user-experience metrics"10 (p. 3). The authors draw four lessons: "It is hard to assess the value of an idea"; a $100M/year return "on a few days' work for one engineer is about as extreme as it gets"; "Bing runs over 10,000 experiments a year, but simple features resulting in such a big improvement happen only once every few years"; and "the overhead of running an experiment must be small", which at Bing meant ExP, "Microsoft's experimentation system"10 (pp. 4-5). Page 1 uses the same case for the economics of testing before building; this page asks what it says about queues.

What waiting cost (derived). Both figures are lower bounds ("over $100M", "more than six months"). If the lift would have started the day the idea was ranked and the annual rate held, six months of delay forwent at least half of the annual figure, 0.5 × $100M = $50M of US revenue, ignoring ramp-up and any change in revenue over the six months. As a rate that is about $100M ÷ 52 ≈ $1.9M per week against a duration of "a few days": a delay-cost-to-duration ratio (Section 4.2) far above any plausible neighbor in the queue, visible only afterwards.

What the case does not show. The textbook does not say how the idea was ranked, by whom, or whether any scoring model was used, so it is no evidence against RICE, WSJF or an attrition filter in particular. Nor did the idea die. It waited, then came back through an engineer's initiative. In Shape Up's terms that is "important ideas come back", and what the filter fails to do is shorten the wait. Where could such a try happen under Shape Up? Cool-down: programmers and designers there "are free to work on whatever they want", including to "explore new ideas, or try out new technical possibilities"1 (Ch. 8). The structure leaves a slot for a few days of one engineer without a bet. It has no measurement system in it, though, and the Bing try was cheap because evaluation was cheap.

The decision rule it supports (derived). Let \( C_r \) be what it costs to rank an item (scoring, grooming, argument across several people) and \( C_t \) what it costs to build the smallest version and measure it. If \( C_t < C_r \), ranking costs more than the test it would be deciding whether to run, and trying dominates, provided the try is measurable and reversible. The choice is per item: cheap, measurable changes skip the queue, and expensive or unmeasurable ones go through sizing and a bet. The tail still matters: wins of that size are rare among Bing's 10,000-plus experiments a year10, and most tested ideas do not improve their metric (page 1, Section 2), so the policy pays because the try is cheap relative to a heavy tail, not because most tries win.

Attrition filter: an idea gets a soft no, then survives only if it comes back; the Bing case annotated Bing 2012: waited more than 6 months then +12% revenue; delay cost at least $50M (derived) Raw idea Soft no "maybe some day" Comes back? yes Shape + pitch Betting table Bet 6 weeks Own lists per function no Dies quietly not chosen Let go nothing tracked Linear keeps a manageable backlog: "low priority ones will never get fixed" (the same fate, with a list retained). Re-arrival is the only filter in the flow; the sources give no evidence that it tracks importance.
Figure 6. Shape Up's replacement for a backlog: a soft no, decentralized lists that are not inputs to the betting table, and survival by being raised again1 (Ch. 3, Ch. 7, Ch. 8). The Bing annotation is the one measured counter-case10; the $50M floor is derived from "over $100M annually" and "more than six months". Linear's variant is from its method page5.
Table 7. What happens to an idea that is not ranked, under each policy, as the sources state it; cells marked "writer's reading" are inferred. No source measures how often an important idea is lost under any of them.
PolicyWhere the idea sitsWhat brings it backHow it is lostEvidence
Sorted scoring list (RICE)2Low on a sorted list; the post says to "sort your list and re-evaluate"Writer's reading: re-scoring after its inputs changeWriter's reading: a mis-scored item stays lowIntercom's practice; the post says scores are not a rule
No backlog (Shape Up)1Nowhere central; functions keep their own listsSomeone lobbies at a later betting table; customers complain againIt is not raised again, whatever it was worthPractitioner assertion; anecdote
Manageable backlog (Linear)5A focused backlogIt "resurfaces""Low priority ones will never get fixed"Stated practice; no data
Cheap trial (Bing)10On a list until someone tries itAn engineer decides to try itIt waits (more than six months here)One measured case

7. Cycles, bets and the circuit breaker

7.1 Six weeks plus two, against n weeks

Shape Up's cycle is "a six week period of time where teams work uninterruptedly on shaped projects", followed by a two-week cool-down "to do ad-hoc tasks, fix bugs, and hold a betting table"1 (glossary). The length is argued from experience. "Some companies use two-week cycles (aka 'sprints'). We learned that two weeks is too short to get anything meaningful done. Worse than that, two-week cycles are extremely costly due to the planning overhead." Six weeks is "long enough to finish something meaningful and still short enough to see the end from the beginning", and, in the chapter's words, "People need to feel the deadline looming in order to make trade-offs"1 (Ch. 8). Linear takes the opposite number: "2-week cycles are the most common in software building. They're short enough to not lose sight of other priorities, but long enough to build significant features"5. Neither source gives data. Linear's "most common" is an unsourced popularity claim, and Shape Up's "extremely costly" is an assertion without hours attached.

Sixteen weeks under Shape Up cycles and under two-week cycles, with planning points and the circuit breaker marked Shape Up Cycle: one bet Cool-down Cycle: one bet Cool-down betting table betting table circuit breaker: no extension by default circuit breaker n = 2 weeks planning point unfinished items roll over automatically (Linear) 0 4 8 12 16 weeks
Figure 7. Sixteen weeks of calendar under the two cadences. Shape Up: two six-week cycles, each followed by a two-week cool-down that holds the betting table, with unfinished bets stopped by default1. Linear: n-week cycles, shown at n = 2, the length it calls most common, with unfinished items moving to the next cycle5. Derived counts: 2 betting tables against 8 cycle-planning points in 16 weeks, or 6.5 against 26 per year; cool-down is 2 of 8 weeks, 25% of the calendar.

7.2 What a bet commits

The betting table is a meeting "held during cool-down where stakeholders decide what to do in the next cycle". At Basecamp it is "the CEO (who in our case is the last word on product), CTO, a senior programmer, and a product strategist", and "the call rarely goes longer than an hour or two"1 (Ch. 8). Derived: with four attendees and 6.5 cool-downs per year, that is 26 to 52 person-hours a year of top-level time, excluding the pre-reading and one-on-ones that the chapter says precede it. The chapter lists three properties of a bet. It has "a payout": a pitch defines a specific result that makes the bet worth making. It is a commitment: "If we bet six weeks, then we commit to giving the team the entire six weeks to work exclusively on that thing with no interruptions". And "a smart bet has a cap on the downside. If we bet six weeks on something, the most we can lose is six weeks"1. Bets are placed one cycle ahead only, to "keep the slate clean" and keep options open1.

The table's criteria are the five questions of Section 3.2, none scored. Chapters 6 to 8 give no ranking rule; Chapter 9 supplies the questions but no formula1.

7.3 The circuit breaker, against automatic roll-over

"Teams have to ship the work within the amount of time that we bet. If they don't finish, by default the project doesn't get an extension."1 The chapter gives three reasons. It "eliminates the risk of runaway projects": "If the project was only worth six weeks, it would be foolish to spend two, three or ten times that". A miss "means we did something wrong in the shaping", so the next shaping track reframes it, and the new pitch is reviewed "to see if it really changes our odds of success before dedicating another six weeks to it". And a hard deadline makes the team question scope as it builds1. Linear's default is the reverse: "Don't overload cycles with tasks and let unfinished items move to the next cycle automatically"5.

Table 8. Shape Up and Linear, on the dimensions where they disagree. Both columns are self-reports of practice, without outcome data156.
DimensionShape UpLinear
CycleSix weeks, then two-week cool-downn weeks; 2 "most common"
Unfinished workStopped by default; re-shaped and re-betMoves to the next cycle automatically
InterruptionsNone, except a real crisis ("true crises are very rare")Space reserved for unplanned work
BacklogNone central"Manageable backlog"
Who orders the workBetting table: CEO, CTO, senior programmer, strategistNot stated on the fetched pages; each project has a named owner who writes the brief and delivers5
Project sizeSmall batch 1-2 weeks, big batch 6 weeksSeveral people and more than two weeks makes a project6

The stop-loss reading (derived). Reinertsen gives a hypothetical that shows why a fixed plan is not sacred: a feature planned at 1 week of effort valued by 50% of customers turns out to need 10 weeks and be valued by 5%, "a factor of 100 change in its cost-to-benefit ratio. This emergent information completely changes the economics of our original choice. In such cases, blindly insisting on conformance to the original plan destroys economic value"3 (pp. 8-9). The circuit breaker is a coarse automatic response to the same fact. Reinertsen wants economics re-evaluated as information arrives; Shape Up fixes one re-evaluation point, the end of the appetite, and makes the default at that point to stop. The trade-off is in the trigger: a re-estimate needs a good estimate of remaining cost and value (the input problem of Section 4.4), while a time trigger needs none, much as a timeout bounds the worst case of a remote call without anyone estimating the remote work. What it costs is the partial work "as pitched", and it presumes shaping good enough that most bets finish. The glossary defines "de-risk" as improving "the odds of shipping within one cycle by shaping and removing rabbit holes"1, and no fetched source gives a completion rate.

8. Roadmaps, outcomes and dates

8.1 What the critique targets

Cagan's 2015 article defines the object of attack: "a prioritized list of features or projects that someone believes will actually solve some problem (even if that problem is often not explicitly stated or understood)". His example is a roadmap line, "integrate PayPal as an additional payment method", which could be there because a market cannot pay with the existing methods, because of international payment issues, to lower payment fees, or because someone sees a competitive disadvantage; the line does not say which7. He traces the artifact to "the old centralized command-and-control model", in which "a bunch of stakeholders meet, usually quarterly, in a room to come up with the priorities and the projects for the teams for the next quarter"7. A 2025 SVPG article on stakeholders repeats the claim in a milder form: "roadmaps of features so seldom generate the necessary business results", and teams should "reframe the many requests for features and projects (no matter where they originate)"8. Neither article has data. The 2015 piece suggests an audit instead: look back over the past year and grade the roadmap "on the success rate ... in terms of how many of the items actually met the business objectives"7. Page 1, Section 2, gives the measured base rates to compare such an audit against.

8.2 A roadmap does two jobs

The same article grants the roadmap two legitimate purposes. Management "wants to make sure that the teams are working on the highest business value items first", and "there are cases where they actually need to make date-based commitments, and the roadmap is where they see and track those commitments". Hence a design criterion: "any alternative approach to roadmaps must address these needs at least as well as they are addressed today"7. Cagan's answers are objectives for the first job and high-integrity commitments for the second. The criterion also works as a test of the other methods in this page. Shape Up answers the first job with the betting table and the second only inside a cycle: the bet is a six-week promise with floating scope1. For a two-year product, HEY, the chapter says "Every bet on HEY was placed one at a time", the team "made no specific commitments about what would go into those cycles", and the July 2020 launch followed two cleanup cycles that "significantly cut back the feature set"1 (Ch. 9). The betting table "didn't know they would be working on HEY for two years", so on the chapter's account the launch date was an outcome, not an up-front promise1.

8.3 The outcome roadmap

Cagan replaces the feature list with two components: a product vision, and "specific, prioritized business objectives for each product team". The method is to "tell the team what you need them to accomplish, and how the results will be measured, and let the team figure out the best way to solve the problems", with OKRs as his "current favorite" (mechanics are on page 4)7. His illustration is explicitly hypothetical: "suppose your product currently requires 30 days for a new customer to on-board" and needs 3 hours or less to scale, giving the objective "dramatically reduce the time it takes for a new customer to go live" and the key result "average on-boarding time less than 3 hours"7. For a runtime team the analogue would be an objective on, say, time from model download to first token on a device class, with the numbers left to the team; none is supplied here.

What moves is the unit being ordered: management now prioritizes "the importance of business results, rather than perceived value of features"7. A team "is not off the hook just by delivering a requested feature"; the feature must move the key result "otherwise the team needs to try a different approach"7. What does not move is the question of which solution a team tries first within an objective. The article is silent on that, and Sections 2 to 4 still apply there.

8.4 What a date costs: high-integrity commitments

Cagan's 2012 article locates the trouble in timing: commitments "are made too early. They are made before we know if we can actually deliver on this obligation, and even more important, if what we deliver will actually solve the problem for the customer"9. The trade he proposes is "a little bit of give and take": stakeholders give the team time in discovery to validate the solution for value and usability with customers, for feasibility with engineers, and for viability with stakeholders. Afterwards the team can "make an informed and high-confidence statement about when we could deliver this, and what type of business results we can expect. We can also decide if it's even worth doing"9. The article adds a point about queues that Section 5 explains: delivery managers "are essential to actually coming up with any commitment dates", because "just because your engineers believe something might take only two sprints to deliver, what if that team is already occupied on other work, and they can't start on this work for another month?"9. A date is queue wait plus duration. The 2025 article keeps the rule: "a date you can count on, but it requires some product discovery work ... These come at a cost, so they should be used sparingly"8.

The sources name one cost, discovery time; two more follow from the writer's reading and from the arithmetic of Section 4.

  • Discovery before the date exists. The articles give no duration for it. The date is paid for in calendar time during which no date can be quoted9.
  • Less room to change course (the writer's reading). Shape Up's promise fixes the time and lets the scope move1. A date that also fixes the deliverable has nothing left to give when the work turns out larger. SVPG's pieces say only that dates "come at a cost"8.
  • Displaced work (derived). A committed date reserves a slot. If it forces an order that differs from the delay-cost order, the penalty is the Section 4.2 difference. In Figure 4, forcing A before B costs \( c_B d_A - c_A d_B = 20 \times 3 - 30 \times 1 = 30 \), that is $30k of delay, on the displaced work. The sources do not mention this cost, but it is what a date commitment charges every other item in the queue.

The evidence is thin. The mechanism rests on three articles from one organization over thirteen years (2012, 2015, 2025), and each article ends with a promotion for SVPG's workshops. No fetched source reports how accurate high-integrity dates turn out to be.

8.5 A themes roadmap, and what each horizon means

Linear keeps a roadmap. In planning, "we define a high-level roadmap with themes that we want to address in the next six months or so. And we break those themes further into concrete features that we can build. Those features often become the projects themselves". The descriptions start "quite open-ended": a planned analytics feature "turned into four or five different projects in the end". Projects are things with multiple people and more than two weeks of work, and "the majority of projects should have end dates"6. The method page adds that initiative meetings review "projects and their target dates" and cycle planning "pull[s] from projects"5. Neither page says who ranks projects, how end dates are set, or how often they hold.

How far ahead each approach plans, in weeks from today Linear cycle 2 weeks Shape Up bet 6 weeks, then 2 to decide the next Quarterly priorities one quarter (feature list) Linear themes about six months, themes to projects High-integrity date discovery date, set by the work 0 6 13 26 weeks ahead
Figure 8. Planning horizons in weeks from today. Linear: cycles of 2 weeks and a themes roadmap of "about six months"56. Shape Up: a six-week bet and a two-week cool-down1. Quarterly priority-setting is the practice Cagan criticizes7. A quarter is drawn as 13 weeks and six months as 26 (the writer's conversions). A high-integrity date has no fixed horizon and the sources give no discovery duration, so its bar is schematic.
Table 9. Four stances on the roadmap, by what is planned and what a date means. All evidence is the proponents' own account.
StanceUnit plannedHorizonWhat a date meansWho orders the work
Feature roadmap (Cagan's target)7Prioritized features or projects"Usually quarterly" planningA delivery promise per itemStakeholders in a room
Outcome roadmap (SVPG)79Vision plus prioritized objectives per teamNot statedOnly for a high-integrity commitment, after discoveryManagement orders objectives; the team picks solutions
No roadmap, bets (Shape Up)1A shaped projectOne cycle aheadEnd of the cycle; scope floatsThe betting table
Themes roadmap (Linear)6Themes, then features, then projects"Six months or so""Majority of projects should have end dates"Not stated

9. Choosing a method for a context

None of the sources says when its method should be used instead of another; each describes its own practice. The table lists what each method needs and where it fails, taken from the sections above, and the conditions after it follow from those mechanisms.

Table 10. The methods side by side. "Evidence" is the strongest kind the fetched sources offer for that method, not a measure of how well it works.
MethodNeeds as inputOptimizesTypical failureEvidence
RICE2Reach from metrics, impact and confidence rungs, effort in person-monthsImpact per person-month across independent itemsClose calls decided by rung choice and counting window (Sec. 2.3); confidence read as a probabilityIntercom's practice; derived sensitivities
Cost of delay, WSJF3411Delay cost per week in one unit; durationsTotal delay cost of a sequence for one teamInput spread (50 to 1 reported); relative scores hide magnitudeSmith's rule is a theorem under stated assumptions; survey figures are the author's own
Appetite and bet1A shaper who can design to a budget; senior time at the tableShipping bounded work; downside capped at one appetiteShaping capacity; ideas die unless raised again; upstream filter unspecifiedBasecamp's experience; anecdotes
Manageable backlog, n-week cycles56Named project owners; a cadenceA list short enough to plan fromAutomatic roll-over has no cap; who ranks is unstatedLinear's account
Outcome objectives7Ranked business results; teams able to run discoveryTying work to results, not outputsDoes not rank solutions within an objectiveEssays, no data
Cheap trial10A change buildable in days; cheap measurementFinding rare large wins without ranking overheadNeeds an experimentation system; most tries failOne measured case plus base rates
High-integrity date89Discovery time; someone who knows capacityA date others can plan aroundDiscovery cost; displaced work (Sec. 8.4)SVPG essays
  • Many small, independent items with measurable reach, one team. A ratio ranking is cheap and adequate. Fix the counting window first, then run the break-even checks of Section 2.3 to see whether the order survives one rung of disagreement.
  • Value that perishes, or a hard external deadline. Use delay cost. RICE has no term for urgency (Section 4.3).
  • Large, uncertain items whose cost is the unknown. Fix a budget and shape to it rather than estimating a design that does not yet exist (Section 3).
  • A change cheaper to try than to discuss, with a measurement already in place. Try it (Section 6.3).
  • Someone outside the team needs a date. Pay for discovery first and price the displaced work (Section 8.4).
  • A list that never shrinks. Divide its size by the completion rate and compare with how long items stay relevant (Section 6.2).

Table 11 applies the conditions to a hypothetical on-device runtime team; no fetched source covers such a team, and it carries no figures.

Table 11. Hypothetical mapping of five kinds of work in an on-device inference runtime team to the mechanism that fits, by the logic of Sections 2 to 8. An illustration of the reasoning, not evidence.
Kind of workProperty that mattersMechanism that fits
Crash on one device class that blocks an integrator's launchPerishable; smallDelay cost is the criterion; Shape Up interrupts a bet only for a real crisis (data lost, the app grinding to a halt, a huge swath of customers affected)1
Support for a newly released model architectureValue decays as attention moves onDelay cost with a deadline; a small appetite
Kernel optimization with an existing benchmarkCheap and measurableTry it; the benchmark plays the role Bing's experiment system played10
Rewrite of the graph executorGrab-bag risk; unclear "done"Narrow to one problem and an appetite; "2.0" is the warning label1
A public API an integrator wants by a dateExternal promiseHigh-integrity commitment, with the displaced work priced

Where experts disagree

Two disagreements run through the sources. The first is about the artifacts: whether to reshape, abolish or keep the roadmap and the backlog. The second is about the ordering itself: whether to estimate and rank, sequence by economics, fix a budget and bet, or distrust ranking and make trying cheap. No fetched source compares any of these on outcomes, so each position is stated by its mechanism and its own evidence.

Roadmaps and backlogs: reshape, abolish or keep?

Reshape: objectives, with dates only as high-integrity commitments

Cagan; Jones and Cagan (SVPG, 2012, 2015, 2025)789

A roadmap does two jobs, ordering by value and holding dates. Replace the feature list with vision plus ranked business objectives, let teams choose solutions, and quote a date only after discovery has made it trustworthy.

  • Serves both jobs the roadmap serves
  • Gives the team a problem and a measure rather than a spec
  • States what must happen before a date exists
  • No outcome data in any of the three pieces; each carries workshop promotion
  • Silent on ordering solutions within an objective
  • Needs stakeholders who accept discovery time

Abolish: no roadmap, no central list, bet one cycle ahead

Singer, Shape Up (37signals)1

Backlogs are a weight; ideas are cheap; important ones come back. Keep decentralized lists, shape a few pitches, and bet six weeks at a time with a circuit breaker.

  • Removes grooming cost and the always-behind feeling
  • Downside of a bet capped at one appetite
  • Keeps options open; the slate is cleared each cycle
  • Practitioner assertion and anecdote; no data
  • Assumes a top-level betting table with authority and senior shapers
  • Re-arrival as importance is untested; the upstream choice of what to shape is unspecified

Keep a light one: themes, end dates, a short backlog

Linear (method page; engineering and product account)56

A roadmap of themes for about six months, projects with end dates, a manageable backlog, two-week cycles, and space for unplanned work.

  • Gives outsiders a horizon without committing to features
  • Keeps the list short enough to plan from
  • Self-report from a small, senior team with one PM
  • Says nothing about who ranks projects or how end dates are kept
  • Unfinished work rolls over, so overruns are uncapped
Assessment

All three move away from a dated feature list set by stakeholders, and differ on what replaces it. Which fits depends on three conditions that no source tests: how many outside parties need dates and how far ahead, whether the company can hold a betting table with real authority, and whether teams can do discovery. The audit Cagan proposes, grading a year of roadmap items against the objectives they were meant to meet, run under each artifact on comparable teams, would settle it; none is published in these sources. Meanwhile, the evidence on the light options is two self-reports: Linear describes a small, senior team with one PM6, and Basecamp describes its own betting table1. Whether either carries over to a larger company with many outside parties who need dates is untested here. For a date promised outside the team, SVPG's rule is the only one of the three that says what must happen first. For the backlog under any option, measure it as a queue (Section 6.2).

Prioritization: score, sequence by economics, bet, or try?

Score and rank

McBride, Intercom (RICE, 2018)2

Reach times Impact times Confidence over Effort turns a debate into a sortable column and exposes out-of-order decisions.

  • Cheap; forces assumptions into the open
  • Makes it visible when you break the order
  • Ordinal rungs; close calls flip on one rung (Table 3)
  • Result depends on the counting window
  • Confidence is evidence quality, not probability

Sequence by what waiting costs

Reinertsen; SAFe's WSJF3411

Quantify delay cost in one unit and order by delay cost per unit of duration; FIFO and ROI ordering are not economic.

  • Exact for one team under stated assumptions
  • Handles urgency and explains why queues are costly
  • Inputs vary 50 to 1 in the author's surveys
  • Single resource, no dependencies; later chapters unfetched
  • Evidence is the author's course surveys

Fix a budget, bet, and let the rest die

Shape Up; Linear for the attrition rule15

Appetites start with a number and end with a design; judge a few shaped pitches by five questions; unraised ideas lapse.

  • Moves uncertainty from date to scope
  • Low planning overhead; capped downside
  • No formula; judgment of a few senior people
  • The attrition premise has no evidence behind it

Distrust the ranking; make trying cheap

Kohavi, Tang and Xu10

"It is hard to assess the value of an idea." A low-ranked Bing change worth over $100M a year waited more than six months; keep the overhead of testing small.

  • The only position with a measured outcome
  • Makes the cost of a wrong rank small
  • One case; needs an experimentation system
  • Most tried ideas fail, so volume is required
Assessment

The positions answer different questions. RICE and the delay-cost rule are close cousins, both a benefit rate over a cost of capacity; they differ in the unit and in whether urgency is priced, and the Project 3 example shows they can invert (Section 4.3). Appetite changes what is estimated: cost becomes an input and the design the output. The Bing case attacks ranking itself. Taken together with the measured base rates (a third of ideas improved their target metric at Microsoft, 10-20% on well-optimized Bing and Google measures10), it suggests the accuracy of the ranking matters less than the price of a wrong pick: a bet capped at one appetite, a roll-over with no cap, or a trial that costs days. That is the writer's inference, not a published result. Use a ratio rule for many small items with measurable reach, delay cost where value perishes, an appetite where the cost is the unknown, and a trial where it is cheaper than discussion. A comparison of methods on the same backlog with outcomes measured would settle the rest; none appears in the sources.

What is changing

Several of these sources are living pages. The Shape Up web edition is stamped ©1999-2026 with no edition year1, Linear's method page is undated5, and the SVPG stakeholder article is dated 1 December 20258; the Intercom post is from 2018 and the Reinertsen sample from 200923. The main open question for this page is cost: if the first working version of an idea gets cheaper to build, the Bing logic extends to larger items and the number of things worth ranking shrinks. Page 1 (Section 8) covers the claims about AI and the cost of a prototype, and none of them is measured in the sources collected here.

References

  1. officialSinger, R. (37signals / Basecamp). Shape Up: Stop Running in Circles and Ship Work that Matters. Web edition (the page states only "Copyright ©1999-2026 37signals LLC"). https://basecamp.com/shapeup. Accessed 2026-10-26. Used for: Ch. 2 ("Two tracks", calendar case, over-specifying and estimation, shaping as strategic work); Ch. 3 (appetite, fixed time and variable scope, "good" is relative, raw ideas, permissions example, grab-bags); Ch. 6 (appetite in the pitch); Ch. 7 (no backlogs, decentralized lists, important ideas come back); Ch. 8 (cycles, cool-down, team sizes, betting table, the meaning of a bet, uninterrupted time, circuit breaker); Ch. 9 (questions to ask, HEY); glossary. Practitioner account and anecdote; no measured data.
  2. blogMcBride, S. (Intercom). RICE: Simple prioritization for product managers. Intercom blog, 5 January 2018. https://www.intercom.com/blog/rice-simple-prioritization-for-product-managers/. Accessed 2026-10-26. Used for: the four factors and their scales, the three-project worked example (inputs only; the formula is an image on the page and the scores in Table 2 are the writer's arithmetic), the limits the author states.
  3. textbookReinertsen, D. G. The Principles of Product Development Flow: Second Generation Lean Product Development. Celeritas Publishing, 2009. Publisher's sample: front matter, contents, Chapter 1. http://lpd2.com/wp-content/uploads/2013/06/ReinertsenFLOWChap1.pdf. Accessed 2026-10-26. Used for: Ch. 1 pp. 1-2 (requirements known at design start, FIFO), p. 4 and footnote 1 on p. 5 (15 percent, 50 to 1), pp. 5-7 (queues, 2 percent, 98.5 percent utilization), pp. 8-9 (conformance), p. 11 (cadence, scope and timing), p. 14 (ROI, critical chain), pp. 17-18 (batch size), p. 23 (operating systems); table of contents for Chapters 2-9. Statistics are the author's course surveys. Chapters 2-9 were not available.
  4. officialScaled Agile, Inc. Weighted Shortest Job First (WSJF). Scaled Agile Framework. https://framework.scaledagile.com/wsjf. Accessed 2026-10-26. Used for: SAFe's definition of WSJF as relative cost of delay divided by relative job duration, its listed inputs, and the epigraph attributed to Reinertsen. Only the introduction was visible; the "Details" section is behind a login.
  5. officialLinear. Linear Method: Principles & Practices. Undated. https://linear.app/method/introduction. Accessed 2026-10-26. Used for: "Work in n-week cycles", "Keep a manageable backlog", "Set strategic product initiatives" (space for unplanned work), "Connect daily work to larger goals with projects", project owners, feedback as a research library. Stated practice; no data.
  6. blogYu, N., Roman, S. and Largman, R. (Linear). How we run projects at Linear. Linear, 5 October 2023. https://linear.app/now/how-we-run-projects-at-linear. Accessed 2026-10-26. Used for: what counts as a project, the six-month themes roadmap, end dates, one PM and engineer-led projects. Self-reported by a small senior team.
  7. blogCagan, M. (SVPG). The Alternative to Roadmaps. Silicon Valley Product Group, 7 September 2015. https://www.svpg.com/the-alternative-to-roadmaps/. Accessed 2026-10-26. Used for: the critique of the feature roadmap, its two purposes, vision plus prioritized objectives, the hypothetical onboarding objective, the roadmap self-audit. No data in the article.
  8. blogJones, C. and Cagan, M. (SVPG). Stakeholders and the Product Model. Silicon Valley Product Group, 1 December 2025. https://www.svpg.com/stakeholders-and-the-product-model/. Accessed 2026-10-26. Used for: reframing requests, the claim about feature roadmaps, the definition of a high-integrity commitment and its cost.
  9. blogCagan, M. (SVPG). High-Integrity Commitments. Silicon Valley Product Group, 15 December 2012. https://svpg.com/managing-commitments-in-an-agile-team/. Accessed 2026-10-26. Used for: why commitments fail (made too early), the give-and-take with discovery, and the role of delivery managers in dating. Linked from the 2015 article above.
  10. textbookKohavi, R., Tang, D. and Xu, Y. Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing. Cambridge University Press, 2020. Chapter 1 and front matter, from the authors' companion-site PDF. https://experimentguide.com/wp-content/uploads/TrustworthyOnlineControlledExperiments_PracticalGuideToABTesting_Chapter1.pdf. Accessed 2026-10-26. Used for: the Bing ad-headline case (Ch. 1, pp. 3-5), the volume of experiments (p. 5), and the idea success rates (Tenet 3, p. 13).
  11. textbookBasic Scheduling Algorithms for Single Machine Problems (lecture handout, "Single machine problems"). Hosted under the NYU Stern faculty web space of M. Pinedo, path "scheduling/shakhlevich"; the handout itself shows no author or date. https://web-static.stern.nyu.edu/om/faculty/pinedo/scheduling/shakhlevich/handout02.pdf. Accessed 2026-10-26. Used for: Section 2, Theorem 1 (the weighted shortest processing time rule is optimal for total weighted completion time on one machine; proof by adjacent pairwise interchange) and its attribution to W. E. Smith, 1956. Smith's original paper was not fetched.
  12. textbookAdan, I. and Resing, J. Queueing Systems. Lecture notes, Eindhoven University of Technology, 26 March 2015. https://www.win.tue.nl/~iadan/queueing.pdf. Accessed 2026-10-26. Used for: §3.4 (Little's law \( E(L) = \lambda E(S) \) and its stability assumption); §4.2-4.3 (M/M/1 equilibrium distribution, mean number in system, eq. 4.9 for mean time in system, mean waiting time).
  13. textbookTran-Gia, P. and Hossfeld, T. Performance Modeling and Analysis of Communication Networks: A Lecture Note. Würzburg University Press, 2021, DOI 10.25972/WUP-978-3-95826-153-2; online version, §5.6.3 "Kingman's Approximation of Mean Waiting Times". https://hossfeld.github.io/performance-modeling/chapter5_nonmarkovianSystems/ch5-6-KingmanApproximation.html. Accessed 2026-10-26. Used for: Kingman's approximation of the mean waiting time of a GI/GI/1 queue and Daley's upper bound.