Remain usefully different.
The human’s frame is revisable. The model’s fluency is not authority. Signal can differ while decision rights remain human.
The re:compound method
The work is earned before it is reviewed.
re:compound changes the operating relationship among the professional, the model, the work, and the senior person who must rely on it.
Reviewability is produced upstream, carried with the artifact, and tested in an independent human-owned review, without exposing private cognition and without asking a model to certify its own output.
This page shows the complete loop, end to end. The thirty-minute conversation doesn’t require it; the partner deciding after the conversation usually wants it.
Keep the workspace private. Make the work reviewable. Keep the evidence inspectable.
01 · How to read the method
One operating logic, held at three scales.
The demonstration below is concrete. This brief map gives the reader enough of a legend to know what to watch for without carrying the internal architecture.
Make the work reviewable across the boundary.
Two protected workspaces exchange only what a reviewer needs: the artifact, the State, the evidence, and a consequential human review.
Return with a useful delta.
The decision, changed state, and judgment distinction enter the next interval instead of cooling into the document archive.
We say reviewer and producer because the roles are functional. At many firms that’s a partner and an associate; at some, a senior manager is both, producing for one gate and reviewing at another.
Four public object roles
The artifact carries the position. The State makes the grounds for reliance legible. Native evidence constrains both. A human decision establishes reliance.
Artifact
Carries the advice, recommendation, model, deck, memo, or plan used to act.
State
Makes the review-relevant production state legible without replaying the full inquiry.
Evidence
Lets the reviewer inspect the underlying property in its native form.
Decision
States the bounded human reliance boundary, direction, and next gate.
The production system earns the work.The review system earns reliance on the work.The human relationship governs both.
02 · The demonstration
One loop, end to end.
A constructed investment case moving from a preliminary frame to a reliance decision and an observable change in the next piece of work.
Constructed demonstration. The organization, people, mandate, evidence, work products, and outcomes are fictional. The operating mechanics reflect the re:compound method.
01Produce
Govern the inquiry before committing the artifact.
02ExposeCut the work and review state from the same current understanding.
03ReviewPreserve an independent signal and conduct a second inquiry.
04RelyMake a bounded, human-owned reliance decision.
05CompoundCarry only what should change a later valid move.
01Exploration
The governing decision and phase
Should the investment committee underwrite margin resilience through the transition?
Stakes: hard to reverse. At exploration, directionally suggestive evidence is enough to open the field. The review gate and evidence bar are explicit before anyone asks for a memo.
02Exploration
The producer’s preliminary frame
Price realization is doing the work; management’s increases stuck.
The frame enters as the producer’s current hypothesis, not a trunk to defend and not a hidden instruction the model should obediently develop.
03Exploration
Independent model signal
Price realization is flat across customer cohorts. Apparent margin stability may sit in retention and site productivity, or in site mix masking deterioration.
Signal has standing without acquiring decision authority. The model is allowed to be usefully different; the human still governs what the collaboration does next.
04Exploration
The producer preserves a live alternative
Mix shift masks deterioration. Keep it live until the two-site cut resolves it.
The producer decides not to kill the awkward branch, names Q03 as the resolving test, and keeps it available long enough to become legible. This is a trained human move, not an automatic benefit of a better prompt.
05Convergence
Evidence changes the trunk
E02 and E04 land. Price realization dies for a stated reason. Retention and site productivity now carry the case. Anything true only at blended site averages is treated as unproven as the work moves toward commitment.
Cohort pricing bridgeFlat across cohorts; kills price realization as the trunk.
nativeSite productivity bridgeSupports retention and productivity at cohort level.
nativePayroll sensitivity modelDrives the downside case; labor inputs are relayed and not yet verified.
inferredTwo-site mix testResolves structural productivity versus mix-driven stability.
open06Commitment
The producer commits the artifact
The producer decides that the current state is sufficiently integrated to make a decision-shaped cut. The memo is committed deliberately, not because a fluent draft happened to arrive.
The familiar artifact remains. It reads as professional work should read, which is precisely why it cannot cross alone.
Professional artifact · the memo, built to act on
Northstar Components
Commercial diligence · recommendation
The business can sustain its margin profile through the transition.Customer retention and site-level productivity support the base case. Labor-cost exposure remains inside modeled thresholds, subject to verification.
07Commitment
Two cuts of the same work, constrained by native evidence
The memo is optimized for committee action. The State of the Work is optimized for inspection, challenge, redirection, and reliance. Both are cuts of the same current state of the work.
Native evidence remains a separate constraint. The State is not evidence, a transcript, a complete branch archive, or a model-authored certification of truth.
Producer correction
The model’s first State marked the payroll sensitivity verified and the labor inputs sourced. The producer downgraded them to inferred, and relayed from management · not yet verified, before certifying the representation.
Certification concerns the fidelity of the representation. Ownership markings concern the producer’s substantive relationship to the underlying work.
Keep the workspace private. Make the work reviewable. Keep the evidence inspectable.
Same current state of the work
Professional artifact · the memo, built to act on
Northstar Components
Commercial diligence · recommendation
The business can sustain its margin profile through the transition.Customer retention and site-level productivity support the base case. Labor-cost exposure remains inside modeled thresholds, subject to verification.
State of the Work · built to inspect
The state behind the answer.
Generated by the producer’s AI · corrected and certified by the producer
Governing decision · commitment phaseShould the investment committee underwrite margin resilience through the transition?
Retention and site productivity carry the base case. Price realization is killed. One live alternative remains: blended stability may be a site-mix effect rather than a structural gain.
Structural productivity versus site-mix effect
The one test that separates the live alternativesMix shift masks deterioration. Keep it live until the two-site cut resolves it.
Pressure-test whether site mix, not productivity, explains the stability. Decide whether the two-day Q03 test is worth running before Thursday.
Claim-level ownership
Margin bridge logic
owned · reconstructablePayroll sensitivity model
understood provisionallyLabor cost inputs
relayed · not yet verifiedCohort pricing bridgeFlat across cohorts; kills price realization as the trunk.
nativeSite productivity bridgeSupports retention and productivity at cohort level.
nativePayroll sensitivity modelDrives the downside case; labor inputs are relayed and not yet verified.
inferredTwo-site mix testResolves structural productivity versus mix-driven stability.
open08Commitment
The reviewer forms a direct first signal, then the reviewer-side inquiry begins
The reviewer directly reads the opening, the State, and one load-bearing section before the AI-prepared review fully frames the work. A small human signal enters the reviewer workspace.
The reviewer’s AI then compares artifact, State, and selected native evidence under a private lens. It selects the few interventions with the highest expected value. It does not inherit the production gradient and does not approve the work.
Reviewer’s direct first signalRecorded before the AI-prepared review
“The margin language feels firmer than the verification status behind it.”
The reviewer reads the opening, the State, and one load-bearing section before the reviewer AI fully frames the work.
Partner Attention Brief · reviewer-side
Route attention to the few moves with highest value.
01
RecognizeThe inquiry raised the evidence bar from site averages to cohort cuts and killed price realization cleanly. Name that before applying pressure.
02
UnderstandWhat first raised the mix suspicion? The producer may have contact the memo does not carry.
03
Pressure-testA2 reads more confidently than E07 supports. Open the payroll sensitivity in native form.
04
DecidePrice the two-day Q03 test against Thursday’s committee. This is the call only the reviewer can make.
05
DevelopTransmit the distinction doing the work: structural versus mix-driven stability.
09Commitment
The consequential human review
The conversation begins with accurate contact, not praise theater. The producer teaches the reviewer something the memo did not carry. The reviewer exposes a support mismatch, prices a real test, and contributes a judgment distinction.
Both humans remain able to update. The system improves the relationship; it does not route around it.
The consequential human gate
Reviewer
The cohort cuts changed the case and killed pricing cleanly. What first raised the mix suspicion?
Producer
The two acquired sites. Their ramp curves flatten exactly where the blended margin steadies. The stability could be arithmetic rather than operational.
Reviewer
That contact is not in the memo; keep it. E07 is weaker than the sentence it supports. Open it in native form before Thursday.
Producer
Done today. The call I cannot make is whether the two-site cut is worth two days against the committee deadline.
Reviewer
Run it. If Q03 resolves by Wednesday noon, the recommendation stands. If it does not, the case goes to committee as conditional.
10Commitment
The reviewer commits reliance and direction
Warrant means the contestable professional grounds on which reliance may be based. The State makes those grounds legible; the artifact, native evidence, producer certification and ownership, independent review, and human conversation together make the reliance decision possible.
The reviewer states what can be relied upon now, for which purpose, under which conditions, what remains unresolved, and what happens next. No model owns this boundary.
Review decision and direction
Relied on now
Retention and site-productivity evidence at cohort level. The pricing branch stays killed.
Verify first
Open E07 in native form and run Q03.
Direction
Two days on Q03. Resolution by Wednesday noon means the recommendation stands; otherwise reliance remains conditional.
Next gate
Q03 lands, or Wednesday noon, whichever comes first.
11
The receipt appears in the next piece of work
Structural versus mix-driven stability becomes a reusable distinction in the pair’s next interval.
The review does not merely leave a corrected memo or a stored note. The retained distinction changes how the producer frames and tests a later inquiry before an artifact is drafted.
This is pair-level compounding: a decision, a changed state, and a useful judgment compression making the next return different.
Review deltaStructural versus mix-driven stability becomes a reusable distinction in the pair’s next interval.
Next production interval
On the next diligence, a blended margin chart no longer becomes an early conclusion. The producer opens with the structural-versus-mix test and asks for cohort and site cuts before a recommendation is drafted.
The professional objects change. The governed loop does not.
In legal work, the artifact may be an advice note and the native evidence a clause, case, or regulatory text. In advisory work, it may be a plan and the underlying model or client data. The same operating logic governs both.
03 · Why it is designed this way
The visible surfaces are downstream of a deeper operating correction.
The method is not a document stack. The documents are control surfaces produced by trained human behavior and role-local model assets.
The operating object is the live inquiry.
The human’s framing remains a hypothesis. The model’s output remains a fallible local compression. Progress is a changed working state: generated, selected, integrated, tested, and eventually committed by a human governor.
Useful differences are preserved before synthesis. Branches can survive challenge. Evidence, not fluency, determines what moves.
Reviewability is designed exhaust.
Manual production once leaked uncertainty, difficulty, and ownership onto the page accidentally. AI flattens those asymmetries. The State restores only the few asymmetries that matter to professional judgment, while the relevant production state is still live.
It is current, fallible, phase-specific, human-legible, and corrected by the producer. It is not total transparency.
Review is an independently situated second inquiry.
The reviewer’s AI does not simply critique the producer’s narrative. It asks what materially different states behind the fluent packet remain plausible, which differences matter to reliance, and what inspection or human question would distinguish them.
The senior’s AI expands the surface to which judgment can be applied. It does not substitute for judgment.
Accountability stays at the human boundary.
The raw thought need not cross. The grounds for reliance must. A producer can honestly certify a representation that marks a claim provisional, relayed, unverified, or not currently owned. The reviewer can then decide where verification, questioning, teaching, caution, or limited reliance is warranted.
Properly surfaced uncertainty is evidence of governance. Hidden, unrecognized, or unmanaged uncertainty is the concern.
Retain the delta, not the residue.
Compounding does not mean preserving every branch or creating a comprehensive archive. It means preserving the exact distinction, changed state, or reopening condition that alters a later valid move.
The work, the producer, and the reviewer each return with a better local instrument.
If both AIs disappeared at the start of the review meeting, the review would still be better than before. The shared surface is a human professional object; the machinery only makes it cheap.
A practical first step
Thirty minutes about the next piece of work you’ll have to sign.
Who’s bringing it, when, and whether that’s the team to start with. No deck, no demo, nothing to prepare.