Intervention and accountability

A PM owns a decision and it is going badly. The leader can see it.

So the leader helps. They join the next customer call, "just to listen." They take the draft and tidy the analysis section, because it was faster than explaining. They chair the working session, because the last one drifted. Each act is generous, each takes ten minutes, and none of them is announced.

Four weeks later the PM no longer holds the decision. Nobody told them. There was no conversation, no failure named, no date, and no way back, because there was never a moment at which something formally changed. The team saw all of it and drew the obvious conclusion. The PM is now managing upward instead of managing the problem, which makes their work worse, which confirms the leader's read.

That is the first failure: the silent takeover. It feels like support and functions as a demotion. The ambiguity is what makes it cruel — a person told plainly "I am taking this decision for six weeks, here is why, here is how you get it back" is in a far better position than a person quietly relieved of it one helpful gesture at a time.

The second failure runs the other way. A leader who has learned not to undermine ownership waits. They do not want to be the manager who takes things back. So they say nothing while the pattern repeats, until the consequence lands. Then the conversation is a post-mortem about a person rather than a correction of a decision, and every cheap option — a question, a raised standard, one added consultation — has expired.

Both failures come from the same missing thing. Intervention was never treated as a decision with its own scope, duration, and exit. It was treated as a mood.

The two failures look like opposites and share a root:

flowchart TD
  P["A decision someone owns<br/>is going badly"] --> H["Help quietly<br/>join the call, tidy the draft,<br/>chair the session"]
  P --> W["Wait, so as not to<br/>undermine ownership"]
  P --> C["Name it<br/>rung, scope, duration, exit"]
  H --> HD["Ownership gone, with no moment<br/>at which anything changed"]
  W --> WD["The consequence lands.<br/>Every cheap rung has expired."]
  C --> CD["A pattern is corrected,<br/>and ownership has a way back"]

Neither of the first two paths ever produces a conversation the other person can answer. The third one is the only path with a return edge on it.

Ask yourself: the last time I stepped into someone's work, could that person have stated, on the day, exactly what had changed, how long it was for, and what would return it? If not, you did not intervene. You occupied.

An intervention is a temporary, explicit change in decision rights. It has four required parts, and it is not an intervention until all four exist.

Part	What it must contain	What happens without it
Named failure mode	The specific pattern going wrong, stated as an observable, with instances	The person defends their character instead of fixing a pattern
Proportionate scope	Which decisions move, and which explicitly stay	A single-class failure becomes a whole-person demotion
Duration and exit	How long, the observable that returns ownership, and who decides	"Temporary" is a word; everyone reads it as permanent
Support attached	What you are adding — information, an interface, a decision to practise on	Intervention becomes surveillance

An intervention with no exit is a demotion that has not been announced. If a demotion is what you mean, say that instead. It is a legitimate decision and it deserves its own honest conversation.

The ladder

Intervention is not one thing. Pick the lowest rung that addresses the failure mode you named, and say out loud which rung you are on.

Rung	Action	Decision rights
1	Ask a calibrated question (LD-03)	Unchanged
2	Raise the evidence standard, or add a checkpoint before commitment	Unchanged; standard raised
3	Require a named consultation before the call	Narrowed to level 3
4	Move this class to recommend-and-I-decide, temporarily	Level 4, for this class only
5	Take the decision, temporarily, with a stated return date	Level 5, for this class only
6	Change the role or the scope permanently	Not an intervention

Two rules govern the ladder.

Lowest sufficient rung. If a raised evidence standard would address the failure mode, rung 4 is an overreaction that costs you the person's ownership for no additional safety.

Early and low beats late and high. Waiting until you are certain removes the cheap rungs from the board. Rung 2 at week two is a smaller event than rung 5 at week six, for you, for them, and for the work.

Name the pattern, not the person and not the instance

A failure mode is an observable pattern with instances attached. "Is not strategic enough" is a trait, and a trait cannot be corrected — it can only be defended. "In the last four weeks, the problem statement has been rewritten three times and no decision has been proposed" is a pattern, and it can be addressed by Friday.

One bad call is data, not a pattern. Three of the same shape is a pattern. Say which one you have, and if it is one instance, use rung 1 or 2 and watch.

Accountability is not blame

Accountability means the person carries both the consequence and the correction. Blame means they carry the story.

PJ-05 gave you the discipline of judging a decision from what was knowable at the time. Apply it here before you intervene. A bad outcome from good reasoning is not a failure mode, and intervening on outcome alone teaches the team a precise and damaging lesson: avoid uncertainty rather than manage it. The people who learn that lesson fastest are the ones who bring you the safest work.

So state which you are responding to — the reasoning or the outcome. If it is the outcome, and the reasoning was sound, the honest intervention may be on the system rather than the person.

The exit is the contract

Write the return before you write the takeover. Four elements:

1. The observable that returns ownership. Not "when I have confidence again." Something the person can watch themselves.
2. The date it is checked, whether or not anyone raises it.
3. Who decides the return.
4. Where it is announced — to the same audience that saw the intervention. A public intervention with a private return is not a return. The team's model of who owns this decision does not update, and neither does the person's.

Laid out on a clock, an intervention has four moments, and three of them are written before the first one happens:

timeline
  title An intervention with a named exit
  Before you act : Failure mode named, with dated instances : Scope fixed - which decisions move and which stay : Exit observable, return date and decider written down
  The takeover : Rung stated out loud, to the person and the team : Support attached - an information gap you close, a decision they keep
  The return date : Exit checked whether or not anyone raises it : Ownership returns, announced where the intervention was
  If unmet : No second extension : The honest role conversation instead

The last column is the one that gets skipped, and skipping it is what converts an intervention into a demotion nobody announced. An intervention with no written return has already decided the role question; it just has not told anyone.

If the exit condition is not met by the date, you have one honest choice and one dishonest one. The honest one is to say that the class is not returning and to have the role conversation. The dishonest one is a second extension. An intervention that keeps getting extended is a decision you are avoiding, and everyone involved can already tell.

All of that arrives through one opening sentence. Here it is written both ways, for the PM holding Noted's large-workspace response-time decision:

The difference is not kindness. The second version is harder to say and easier to receive, because it is the only one that can end.

Boundary

Intervention is the wrong instrument whenever the cause is structural.

Run this test first: would a capable person in this seat, with this direction, these decision rights, this operating system, this evidence, and this interface, have hit the same wall? If the answer is yes, intervening on the person fixes nothing, and it teaches the team that system failures are settled by finding someone to correct.

The structural causes have addresses in this phase. No ordering to resolve the trade-off is LD-01. Unclear authority or a missing evidence standard is LD-02. A signal that never reached a room is LD-04. A capability nobody holds, or an interface that was never real, is LD-05. Check those four before you write an intervention contract, and be prepared to abandon the contract.

Two further limits.

Emergencies are exempt from proportion, not from honesty. A security, legal, or customer-harm situation may require you to take the decision immediately, at rung 5, with no ladder. Do it. Then separate the two things afterwards: the emergency action, and any assessment of the person's judgment. Folding the second into the first produces a verdict nobody can appeal, reached under time pressure, on evidence collected during a fire.

Some interventions are on you. If you have intervened on three different people in the same decision class, the pattern is in the class, not in the people. That is a direction or rights problem wearing a personnel costume.

Read the Noted case if you have not already.

Assume the first of the two funded PM hires has now been in seat for six weeks. Under your LD-02 rights map, they hold Signal 3: P95 response time rose 34% for large workspaces at level 2 — decide and inform.

Here is what you have observed, and it is all you have observed. In six weeks the problem statement has been rewritten three times. Each week produces a new analysis and no proposed decision. Segmentation has not been done. No option has been put to engineering. Meanwhile the end-of-support clock on Signal 4 has kept running the whole time, and the paid seats concentrated in those 4% of workspaces are still exposed.

You have not yet asked them why.

Your task. Decide whether to intervene, and if so, write the contract.

1. State the failure mode as an observable pattern, with its instances. Then check yourself: is this a pattern or a single instance repeated in your memory? Write the instances down with dates before you decide.
2. Run the structural test. Would a capable person in this seat have hit the same wall given your LD-01 direction, your LD-02 rights map, your LD-04 operating system, and the capability picture from LD-05? Answer honestly and name which of the four is weakest. If the wall is structural, say so and stop — then write what you fix instead.
3. Separate reasoning from outcome. Say which one you are responding to. If the reasoning is sound and only the outcome is bad, say what that changes.
4. Choose the rung. Justify why the rung below it is insufficient for the failure mode you named. If you chose rung 5, say honestly whether you chose it for the failure mode or for the end-of-support clock — those are different reasons and only one of them is about this person.
5. Write the scope: which decisions move, and which explicitly stay with them. Name the ones that stay. Unstated scope expands.
6. Write the exit: the observable, the date, who decides, and where the return is announced.
7. Write the support you are attaching. Not encouragement — a specific information gap you will close, an interface you will open, or a decision they keep in order to practise.
8. Write the opening sentence, verbatim, that you will say to them. Then check it for the disguised version: any sentence that begins "let me just help with" and does not name a change in rights is a silent takeover.
9. Write what happens if the exit condition is not met by the date, including the sentence you would say if the honest answer is a role change.

Expect to be pushed on: whether your failure mode is a pattern with dated instances or a trait in disguise; whether step 2 was run honestly or was a formality on the way to the conclusion you had already reached; and whether your exit observable in step 6 is something the PM can watch themselves, or is a restatement of "when I trust them again."

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

• A failure mode written as a pattern with dated instances, not a trait, and an honest check on whether it is three instances or one remembered three times.
• The structural test run against LD-01, LD-02, LD-04, and LD-05, with the weakest of the four named — and a visible willingness to stop and fix the structure instead if that is where the test lands.
• A statement of whether you are responding to the reasoning or the outcome.
• The lowest sufficient rung, with the rung below shown to be insufficient. If rung 5, an honest line on whether the clock or the person drove it.
• Scope that names what stays with the PM; an exit with an observable they can watch, a date, a decider, and a place of announcement; an opening sentence that names a change in rights in plain words.
• The most common weak move is an exit written as your state of mind — "when I'm confident again." It is weak because the person cannot watch it, so it functions as no exit at all.

Produce artifacts/LD-06-intervention-contract.md using the template below.

Write it to be shared with the person it concerns. That constraint does most of the work. A contract you would not hand over is one whose failure mode is a trait, whose scope is vague, or whose exit does not exist — and each of those is easier to see when you imagine the other person reading it in front of you.

In the Product Decision Case, this is the counterpart to LD-02. That map granted authority under stated conditions. This one records what happens when those conditions are not met, on terms that were set before anyone was under pressure.

# Intervention contract · <person or role> · <decision class>

- **Date:** <YYYY-MM-DD>
- **Decision class affected:** <one class, named>
- **Current authority level (LD-02):** <1-5>
- **Level during this intervention:** <1-5>
- **Rung on the ladder:** <1-6> — <name it out loud in the conversation too>

## Structural test — run this before anything else

> Would a capable person in this seat, with this direction, these decision
> rights, this operating system, this evidence, and these interfaces, have hit
> the same wall?

- **Answer:** <yes / no>
- **Weakest of the four:** <direction (LD-01) / rights (LD-02) / operating system
  (LD-04) / capability and interfaces (LD-05)>
- **If yes:** <what I am fixing instead of intervening — and this contract stops
  here>

## Failure mode

<An observable pattern, with instances. Not a trait, not a standard, not other
people's feelings.>

| # | What I observed | Date | Where I observed it |
|---|---|---|---|
| 1 | <…> | <…> | <the artifact, meeting, or decision> |
| 2 | <…> | <…> | <…> |
| 3 | <…> | <…> | <…> |

- **Pattern or single instance?** <three of the same shape is a pattern; one is
  data. Say which you have.>
- **Responding to reasoning or to outcome?** <reasoning / outcome> — <if outcome
  with sound reasoning, say what that changes, and consider whether the
  intervention belongs on the system instead>

## Scope

| Decisions that move to me | Decisions that explicitly stay with them |
|---|---|
| <…> | <name them — unstated scope expands> |

- **Why this rung and not the one below:** <the rung below is insufficient
  because …>
- **If I chose rung 5, the honest reason:** <the failure mode / the clock — these
  are different reasons and only one is about this person>

## Exit

- **Observable that returns ownership:** <something they can watch themselves>
- **Checked on:** <YYYY-MM-DD — arrives whether or not anyone raises it>
- **Who decides the return:** <person>
- **Where the return is announced:** <the same audience that saw the
  intervention. A public intervention with a private return is not a return.>

## Support attached

<Not encouragement. A specific information gap I will close, an interface I will
open, or a decision they keep in order to practise.>

- <…>

## What I will say

- **Opening sentence, verbatim:** "<…>"
- **Check:** <does this sentence name a change in decision rights? Any version
  that starts "let me just help with" and does not name the change is a silent
  takeover.>

## If the exit condition is not met

- **The honest options:** <return anyway with a revised standard / the class does
  not return and we have the role conversation>
- **What I will not do:** <extend a second time. An intervention that keeps
  getting extended is a decision I am avoiding.>
- **Sentence I would say if the answer is a role change:** "<…>"

## Gaps

<Everything marked `<unknown>` above, including any instance you believe
happened but cannot date or locate.>