Decision rights and progressive autonomy

A leader tells a PM: "You own the migration decision. I trust you."

The PM spends a week with engineering, weighs a three-to-five week estimate against a ten-week deadline, and decides to spread the work across the quarter. They announce it. The leader hears about it in a standup, goes quiet, and then says: "I think we should just do it now and get it behind us."

Both people leave that exchange with a different story. The PM believes the ownership was fake. The leader believes the PM made a call above their weight class. Neither is lying. The word "own" was never defined, so each person filled it in with their own meaning, and the meanings only collided at the moment it was most expensive.

What actually went wrong is not trust. It is that four separate things were bundled into one word and none of them was written down. Who decides. Who carries the consequence if it goes badly. What evidence would have made the decision acceptable regardless of the answer. And what condition should have brought it back before it was announced.

Trace the handoff and the collision has a location. It is not the announcement. It is the first line:

sequenceDiagram
  participant L as Leader
  participant P as PM
  L->>P: You own the migration decision. I trust you.
  Note over L,P: authority, consequence, evidence standard<br/>and escalation boundary all unstated
  P->>P: Weighs the three-to-five week estimate<br/>against the ten-week clock
  P->>L: Announces the split plan
  L->>P: I think we should just do it now
  Note over L,P: the two private meanings of "own"<br/>meet at the most expensive moment

Everything after the first message was going to end this way. The note in the middle is the whole lesson: four things were bundled into one word, and neither person could see the other's version until it cost something.

The failure is hard to see from inside because delegation feels generous in the moment. "You own it" is a pleasant sentence to say. It costs the leader nothing at the time and costs them nothing later either — they can always reclaim the decision, because they never specified the conditions under which they would not. Vague grants of autonomy are always revocable. That is what makes them comfortable to give.

There is a quieter cost. Because the leader never wrote an evidence standard, the PM had no way to know whether their reasoning was good enough before they committed. They found out by being overruled. That teaches compliance, not judgment.

Ask yourself: the last time I gave someone a decision, could they have written down, in advance, the condition under which I would have taken it back? If not, you did not delegate a decision. You loaned one.

A decision right is four statements, not one.

Component	The question it answers	Failure when it is left implicit
Authority	Who makes the call, at what level	"Own it" collapses into "propose it" at the worst moment
Consequence	Who carries the cost if it is wrong	The decider optimizes for approval, not outcome
Evidence standard	What must be shown before deciding	Quality is judged after the fact, by taste
Escalation boundary	What condition sends it back up	The leader intervenes on mood, and it reads as arbitrary

Write all four or you have written none.

Levels of authority

"Autonomy" is not binary. Name the level explicitly.

Level	What it means	Who announces the outcome
1 · Decide alone	You decide. No notification required before acting.	The decider
2 · Decide and inform	You decide, then tell named people within a stated window.	The decider
3 · Decide after named consultation	You must hear from specific people first. You are not required to agree with them.	The decider
4 · Recommend, leader decides	You do the work and take a position. The call is not yours.	The leader
5 · Leader decides	The decision is not delegated.	The leader

Level 3 is the one that gets corrupted. Consultation is not consensus. If the named people can block, you have written level 4 and called it level 3, and the person will discover the difference in public.

The other common corruption is level 2 with a quiet veto: "decide and inform, but check with me first." That is level 4. Say level 4. A person operating at level 4 who knows it is level 4 will do better work than a person told they are at level 2 and corrected each time.

Two questions place most decisions on that scale — whether the decider can see the consequence, and whether the choice can still be undone this cycle:

flowchart TD
  D["A decision in this class"] --> C{"Can the decider see<br/>the consequence of being wrong?"}
  C -->|"no — it lands<br/>on someone else"| L3["Level 3<br/>named consultation with<br/>whoever carries it"]
  C -->|"yes"| R{"Still reversible<br/>this cycle?"}
  R -->|"yes"| L12["Level 1 or 2<br/>decide, inform inside<br/>a stated window"]
  R -->|"no — one-way<br/>once it starts"| L4["Level 4<br/>recommend, leader decides"]
  L12 -.->|"escalation boundary:<br/>the stated threshold is crossed"| L4

The dotted edge is the escalation boundary, and it is the only line on this diagram that has to be written down in advance. Undrawn, it still exists — it just gets drawn after the fact, by the leader, on the day they dislike an answer.

Consequence is separate from authority

These come apart more often than people expect, and the split is legitimate.

A PM may hold authority over what gets built while the engineering lead carries the consequence of a system that becomes unmaintainable. A support lead may hold authority over what a customer is told while the PM carries the consequence of a commitment that constrains the roadmap.

Name both. When authority and consequence sit with different people, the person carrying the consequence needs a named channel — usually a documented objection, not a veto. Unnamed, they will use the only channel available to them, which is escalating to whoever will listen.

Evidence standard: the part almost everyone omits

An evidence standard says what has to be true about the reasoning for the decision to be acceptable, independent of which option is chosen.

Written on Noted's migration decision, the two versions look like this:

Note what this does. It lets the person know, before they commit, whether they have done enough. It also means that if they meet the standard and you still disagree with the answer, the disagreement is about judgment, not process — and you must either let it stand or admit you are overruling. Both are honest. What is not honest is deciding after the fact that the process was insufficient because you disliked the outcome.

Progressive autonomy

Autonomy is granted per decision class, not per person. The same PM can be at level 1 on activation instrumentation and level 4 on anything that changes an external customer commitment. That is not an insult. It is precision.

Movement should have a stated rule. The usable rule is calibration, not tenure and not likeability:

You move a level in this class when your last three decisions in it carried a written prediction made in advance, and the outcomes fell inside the confidence you stated.

This rewards the thing you want — stated uncertainty that turns out to be honest — rather than the thing that is easy to see, which is confident delivery. A person who is right 90% of the time and claims certainty every time is worse calibrated than a person who is right 70% of the time and said so.

Autonomy can also move down. Say that in advance too, and say what would trigger it, or the first reduction will land as a punishment.

Boundary

Decision rights cannot be granted for a decision whose consequence the person cannot see or carry.

If a PM cannot observe the cost of a maintenance burden they created, granting them authority over it does not build judgment — it builds a habit that goes uncorrected because the feedback lands on someone else. In that case the right move is not to grant the decision. It is to change what they can see first, or to place them at level 3 with the person who does carry it.

A second limit: a rights map is written against a situation, and situations change. A decision class that was cheap to reverse can become one-way — the moment a migration starts, the moment a commitment reaches a customer, the moment a public announcement goes out. Rights maps that do not name their expiry get applied after they stopped being true. Write the condition that voids each row.

A third limit, and the one that catches product leaders: you cannot delegate specialist authority you do not hold. You can decide who decides. You cannot convert a PM into the person qualified to judge a technical risk, and writing "PM decides" on that row does not make the risk go away — it just removes the one person who could see it from the decision.

Read the Noted case if you have not already.

The standing condition from LD-01 still applies: two PM hires are funded, the first starting in about three weeks, and their strengths are <unknown>.

Take Signal 4: a core dependency loses support in ten weeks.

What the case gives you: engineering estimates the migration at three to five weeks with wide uncertainty; two engineers want to migrate now and two want to spread the work across the quarter; after end-of-support, security patches stop and nothing breaks immediately. The case does not tell you who has authority today. It does tell you the engineering lead owns technical standards and repository workflow, and that some planning assumptions have not been checked against the code.

Your task. Write the decision rights for this decision, then write the version that survives your absence.

1. Split the one question into at least three distinct decisions. "Migrate now or spread it" is one. There are others hiding inside it — for example, what evidence is required before committing to the estimate, and what happens if the migration overruns week four. Name them separately.
2. For each decision, assign the authority level from the table above. Use the numbers. If you write "shared" or "collaborative," you have not assigned it.
3. For each, name who carries the consequence if it goes wrong, and say whether that is the same person as the decider. Where it is not, name the channel the consequence-carrier uses to object.
4. Write the evidence standard for the main decision: what must be shown before anyone chooses, stated so that both possible answers could satisfy it. Then check your own standard for bias. If it can only be met by the option you prefer, it is not a standard.
5. Write the escalation boundary: the specific observation that sends this decision back up, with a threshold. "If it looks risky" is not one.
6. Write the expiry condition for the whole row set. What event makes this map wrong?
7. Now rewrite the map as it will read on the incoming PM's fourth day, when they have no history with the engineering lead and you are unreachable for two days. Change whatever stops working.
8. Commit: name the one decision here you are keeping at level 5, and state what would have to be true for you to move it to level 4.

Expect to be pushed on: whether any of your level 2 rows are level 4 wearing a friendlier label; whether your evidence standard could be satisfied by the answer you do not want; and whether you have quietly assigned a technical judgment to someone who cannot see the consequence of getting it wrong.

Committed to a position? Check it against what a strong answer holds

• The one migration question split into at least three distinct decisions, each with a numbered authority level. No row says "shared" or "collaborative".
• A named consequence-carrier for each decision, and a written objection channel wherever the carrier is not the decider.
• An evidence standard that migrating now and spreading the work could both satisfy, with the bias check done in the open.
• An escalation boundary with an observable and a threshold, and an expiry condition for the whole map.
• A version that works on the incoming PM's fourth day with you unreachable — and a refusal to place a technical risk judgment with someone who cannot see the consequence of getting it wrong.
• The most common weak move is a level 2 with a quiet veto — "decide and inform, but run it past me first." It is weak because the person discovers it was level 4 in public, and what they learn is compliance.

Produce artifacts/LD-02-decision-rights-map.md using the template below.

Write it for two readers at once: the person receiving the authority, who needs to know exactly where their call stops, and the person carrying the consequence, who needs to know how to object without escalating. If either reader has to ask you what a row means, the row is not written yet.

This attaches to the direction from LD-01. Direction resolves the recurring class. This map assigns everything direction deliberately left to a person, and records the conditions under which that assignment changes.

# Decision-rights map · <team or decision area>

- **Date:** <YYYY-MM-DD>
- **Author:** <name or role>
- **Covers the period:** <start> to <end>
- **Voided when:** <the event that makes this whole map wrong>

## Decisions in scope

<Split bundled questions into separate decisions. "Own the migration" is not a
decision; it is three or four.>

| # | Decision, stated as a choice | Authority level (1-5) | Decider | Reversibility |
|---|---|---|---|---|
| 1 | <choice> | <1 decide alone / 2 decide and inform / 3 decide after named consultation / 4 recommend, leader decides / 5 leader decides> | <person or role> | <reversible / costly / one-way> |
| 2 | <choice> | <level> | <person> | <…> |
| 3 | <choice> | <level> | <person> | <…> |

<Check: if any row says "shared", "joint", or "collaborative", it is not
assigned. Pick a number.>

## Consequence

| # | Who carries the cost if this is wrong | Same as decider? | Channel to object |
|---|---|---|---|
| 1 | <person or role> | <yes / no> | <documented objection to whom, by when — not a veto unless you mean a veto> |

## Evidence standard

<For the main decision. What must be shown before anyone chooses, stated so that
every option on the table could satisfy it.>

Before deciding #<n>, the decider must have:

1. <required input, with its range or basis>
2. <the consequence of failure stated separately from its probability>
3. <one rejected option and the reason>
4. <the observation that would trigger reversal>

**Bias check:** <Can this standard be met by the option you do not prefer? If
not, rewrite it.>

## Escalation boundaries

| # | Observation that sends this back up | Threshold | Sends it to |
|---|---|---|---|
| 1 | <observable, not a feeling> | <number or event> | <person> |

## Progressive autonomy

| Decision class | Current level | Moves up when | Moves down when |
|---|---|---|---|
| <class> | <1-5> | <calibration rule: predictions written in advance, outcomes inside stated confidence, over how many decisions> | <the stated condition — write it now, not after the first failure> |

## Specialist authority not delegated

<Judgment this map does not move, because the person receiving authority is not
qualified to hold it. Name who holds it instead.>

- <area> → <person or role>

## Portability check

<Rewrite the map as it will read to someone who joins in three weeks, has no
history with anyone named above, and cannot reach you for two days. Note what
you had to change.>

- **What broke without me in the room:** <…>
- **What I changed as a result:** <…>

## Gaps

<Everything marked `<unknown>` above, listed as open questions with who could
close them and by when.>