The design (T18–T19)
The owner, at T18: “we should probably also model teams of employees. so maybe employees are “pre-perspectives” or sub-perspectives, which can reorganize amongst themselves to form the perspectives that influence the higher level organizational behavior”. The reply narrowed the owner’s word: employees “can be full individual interpreters. They’re constituents of candidate higher-level organizational perspectives”. T19’s principle: “Employees and teams are part of the simulated causal machinery. Organizational perspectives are patterns we attempt to identify in that machinery—not additional agents we insert above it”. That is CLAIM-418’s first part (“several agents can share a perspective, one agent can act through several”) made a design rule. Neither turn cited it.
- Employees. Each has an access function, a frame and an action map (T19 §2), so that an employee can “lack information, possess information but misunderstand it, or understand it correctly yet be unable to act on it”. That is CLAIM-349’s access, authority and action, and CLAIM-296’s has, knows it has and can deploy. Frames and concerns are fixed in this first release (“For the first study, I’d deliberately freeze most of the employee learning machinery”).
- Teams. Participation B_ek ∈ [0, 1] with no requirement that it sum to one, and a team state of “records, procedures, obligations, resources, communication rules”. A team’s proposal “is not an additional independent opinion appended to the employees’ opinions. Otherwise, the simulator would double-count activity merely because we’ve described it at two levels” (see CLAIM-403’s note of 2026-10-26).
- Governance. Typed permissions (observe, communicate, propose, authorize, enact, revise rules), after CLAIM-403’s ladder.
- Perspectives. None in the simulator state (“no PerspectiveState inside CorporateState”). A separate analyzer runs downstream of it, and its modes are candidates that causal tests must then confirm: “statistical decomposition proposes candidates; causal and organizational tests determine what the candidates amount to” (T19 §4).
- Arms (T19 §8). Fixed departments, dynamic working groups, shared information only, delegated authority, and a commander handoff with and without preserved records, beside a conventional comparator.
T18 placed the company as “the fictional engineering organization already considered in CASE-134”. Northstar is CASE-023, and CASE-134 is its handoff variant, one of several. CASE-023 says “a claim says which Northstar it means”. This one means CASE-023’s company, staffed with twelve employees in engineering, customer success, sales and leadership.
Reported, not reproduced (T20)
T20 reports 12 employees with “a dense 2×4 local frame”, 4 formal teams “plus overlapping working groups”, and 64 synthetic worlds of 64 days under “nine conditions”. Mean regret per incident-day, lower is better:
| Organization | Regret |
|---|---|
| Fixed departmental silos | 0.07195 |
| Information sharing only | 0.07713 |
| Dynamic working group, slow approval | 0.07759 |
| Information sharing, fast authorization | 0.00774 |
| Working group with delegated authority | 0.00499 |
| Conventional router, same information and authority | 0.00499 |
- Paired difference. The working group against fast-track information sharing: −0.00275, 95% bootstrap interval −0.00393 to −0.00161. The difference of the two reported means is exactly −0.00275, and the interval’s half-widths (0.00118 and 0.00114) are consistent with it.
- Handoff. Under a simulated telemetry outage the state-estimation MSE is 0.00411 with the incident team’s records preserved and 0.03398 with them erased, a ratio of 8.3. But “that improvement in organizational knowledge often doesn’t change the selected action. The difference in physical regret is much smaller”. No number is given for that regret.
- The analyzer “fits overlapping latent activity modes from employee trajectories”, with the number of factors fixed. T20: “those patterns could reflect shared environmental disturbances rather than genuine higher-level interpretive organization … So the result is candidate organizational modes, not evidence that three real organizational perspectives emerged.”
- Tests. “All 51 selected tests passed: 15 new Northstar tests, 24 inherited v2.8/v2.10 tests, and 12 inherited v2.9.1 tests” (the parts sum to 51), and “All 489 predecessor files from the complete v2.10 archive were preserved byte-for-byte”.
T20 reports nine conditions and gives six regrets and the two handoff MSEs, so at least one condition goes unreported. Neither a v2.10, a v2.9.1 nor a v2.11 exists in dmarx/PerspectiveLab, whose latest import is c54617e (v2.9). They may have been built in the turns before T18, which the page does not hold. Either way nothing here has been run by the record or the lab, and an import of the bundle T20 offers (“Download complete PerspectiveLab v2.11”) would be the first check of any of it.
What the record predicts
- The router ties. With the same information and authority, a router is CASE-158’s mixture of experts under full information, and “governance should not beat it”. T20: “The conventional router ties exactly when given the same inputs and permissions. That’s an important limit on what we can claim.” T20 does not say whether the router is computed by the same function as the working group. In v2.7 and v2.9 that made the ties identities (CLAIM-427’s note of 2026-10-26, later).
- Authorization, not the working group. By the reported numbers, fast authorization alone takes 95.9% of the fall in regret from silos to the delegated working group (0.07195 − 0.00774 of 0.07195 − 0.00499). Information sharing without it is worse than silos (+0.00518), and so is the slow working group (+0.00564). T20 names the reason: “a stipulated penalty for delayed action”. Within this construction that is CLAIM-349, access without authority does not act, and CLAIM-403’s later rungs. T20’s “Delegation matters considerably” is right. Its credit to the working group (“a smaller improvement”) is not evidence for the working group as a structure, since the router with the same information and authority scores the same.
- Records against capacity. CASE-134’s note of 2026-10-22 says that a briefing “transfers the record of an arrangement, not the capacity the outgoing commander had built”. The reported handoff fits that: the records carry the estimate, and the action barely changes. T20 reads it the other way round, “preserving a capacity for interpretation is not equivalent to exercising that capacity successfully”, and calls the records the capacity. On CASE-134’s terms they preserved the record, and the capacity to act was not shown to depend on them.
- Modes from common causes. In a check made when this case was filed, twelve employees in four departments with no communication at all, each seeing two of three shared disturbances plus noise over 64 × 64 steps, give correlation eigenvalues 6.76, 2.08 and 1.58 and then 0.24. That is three clear modes, each spanning all four departments. Cross-departmental modes need no organization, so the analyzer’s output cannot meet CLAIM-327’s condition (“when its members read it, condition what they do on it, and by doing so regenerate it”) without an intervention on the coupling. T20 says the same: the next obstacle is “identifying a mode by its continuing causal organization, not simply by correlations among participants”. How one would do it is QUESTION-276, and what a recovered mode could amount to is CLAIM-454.
What would make it informative
- The bundle itself, imported to the lab and re-run, with the ninth condition and the handoff arms’ regret reported.
- Whether the conventional router shares the working group’s code. If it does, the tie is an identity and says nothing about either.
- CASE-158’s four arms, in which a router could lose (cycles, no common scale, capture, bandit feedback), since every arm reported here is full-information.
- A comparator from the organizational-learning literature the record already holds: March’s members, organizational code and turnover (LIT-729, THEORY-142) are the “conventional agent-based organizational model” T18 and T19 asked for, and they include the turnover the handoff arm needs.
- A held-out test of QUESTION-276’s criterion. T21–T23 turned this into a nonparametric inference model (CASE-165).
What it cannot show
Anything about real firms. The environment is synthetic, the frames and concerns are fixed, and the delay penalty that produces most of the effect is stipulated. Even reproduced, it would show what this construction does, and its own analyzer finds candidate modes only.