Ongoing research publication · current through Phase B
Can history become part of capability?
What changed, what stayed the same, and what became newly interesting across two completed phases.
The experiment in one minute
Formation built eight histories. Generalization changed the information world.
Formation
What happens when the same collaborators repeatedly work together?
Eight fixed pairs, identities, roles, and four-turn collaboration.
Accumulated history, task exposure, and later research-path freedom.
Different searches, sources, challenge styles, language, and candidate habits.
Many candidate habits weakened later; query wording and source reuse often did not survive topic changes.
Personality, relationship, preference, or a causal history effect.
Generalization
Same pairs. Same histories. New information world. What travels?
Partners, identities, histories, roles, and collaboration scaffold.
Broader scholarly research plus a disclosed provider/runtime transition.
Broad collaboration persisted; differences lived more in evidence choices and corrective routes than final answer shape.
Exact phrases, some Formation orientations, and Pair 04’s striking structural split did not reliably persist. Pair 08’s failure-mode wording was one inquiry.
That Formation history caused the behavior or created partner-specific capability.
The next question
Where do the differences live?
In the individual agent, assigned role, accumulated experience, specific partner, accessible history, or the interaction itself? Phase C is intended to begin separating those explanations. Its exact protocol is not yet ratified. Phase C has not started and has no execution authority.
01
Current scientific boundary
Formation is complete, valid, closed, and frozen at dosage [12,12,12,12,12,12,12,12]. Generalization is complete. Qualifying observations are PB-GEN-04, PB-GEN-05, and PB-GEN-06 only; dosage [3,3,3,3,3,3,3,3]. Descriptive generalization is supported. A causal Formation-history effect is not supported. Shared-History Change is next, not started, and has no execution authority. Round 17 is a Formation chronology marker, not the current scientific frontier.
02
The central question is not whether agents follow their roles.
Wren investigates and Ada challenges because the protocol tells them to. The more interesting question is whether eight pairs given similar jobs begin doing those jobs differently—and whether those differences survive a changed environment.
03
Formation built eight inspectable histories.
Every original pair completed twelve qualifying shared experiences. Independent reading found organized, heterogeneous histories that were insufficient for a supported developmental or relationship-formation finding. That Formation interpretation remains frozen.
04
Generalization asked what traveled into a new information world.
The same pairs, identities, partners, and histories entered a broader scholarly environment. Across three qualifying inquiries, source-bounded, caveat-preserving, partner-responsive argumentative organization persisted. Final form converged more strongly than pair-specific routes.
What persisted
Broad form and partner-responsive uptake.
Across three new inquiries, qualification changed the scope or structure of the next synthesis. Final answers preserved caveats while compressing the exchange.
What differentiated
The route more than the destination.
Pairs differed most visibly in source choices, examples, emphasis, and corrective mechanism. Many ended in similarly cautious argumentative forms. Many of those local routes were prompt-compatible and not unique.
05
Three representative cases, not a second eight-pair catalog.
The full longitudinal set lives on the pairs page. The cases below are the same three used on the homepage: a prompt-entangled recurring candidate, a strong local pattern that did not persist, and a suggestive candidate with cohort-common alternatives.
Category decomposition appeared, but it is prompt-compatible and not unique.
No distinctive wording persisted, and other pairs also decomposed broad causal objects.
The distinctive structural move peaked in PB-GEN-04 and attenuated.
Final reasoning form still converged with the cohort. This pair is the strongest available counterexample to assuming that a striking Generalization move will persist.
Overclaim resistance is a plausible continuity and is also cohort-wide. Failure-mode specificity is a PB-GEN-04 episode, not a three-inquiry signature.
Pair-08 received disproportionate earlier close reading; the behavior is also expected from role design and appears in other pairs.
06
What disappeared is part of the result.
Pair 04’s three-level structural split peaked in PB-GEN-04 and attenuated. Exact query wording and selected-source continuity often failed to survive topic changes. Several candidate Formation orientations did not reappear consistently. Hiding disappearing hypotheses would overstate pair stability.
07
Conceptual forms survived more clearly than exact phrases.
- Within each Generalization exchange, partner-introduced distinctions were often explicitly acknowledged and reused in the next synthesis. (direct observation)
- Across inquiries, conceptual forms survived more clearly than exact phrases: causal unit, comparison group, exposure definition, outcome separation, and implementation boundary. (recurring descriptive pattern)
- Common phrases such as 'I agree', 'best-supported', 'conditional', and 'under these conditions' are cohort-wide register and protocol effects, not pair signatures. (counterevidence)
- No pair-specific phrase inheritance across the Formation-to-Generalization boundary is established. (not demonstrated)
Why this may matter for teams
Human team concepts give us questions—not conclusions.
These are theoretical scaffolds. The study does not claim that the agents have human emotional or team states.
Experienced teammates learn how the other person thinks. Technical term: shared mental models
Does Wren begin anticipating the distinctions her particular Ada will make?
The Lineage question is whether any of that anticipation is visible in the sealed text after a changed information world.Teams learn who is good at what. Technical term: transactive memory
Does each partner increasingly rely on the other for a cognitive function, creating efficiency and blind spots?
The Lineage question is whether role overlap is pair-history or just the assigned jobs.Teammates notice when the other is drifting. Technical term: mutual monitoring
Does Ada catch overreach without explicit prompting, and can Wren detect a weak challenge?
The Lineage question is whether monitoring is partner-specific or prompt-and-role generic.Sometimes one steps into the other's job. Technical term: backup behavior
Do collaborators begin performing pieces of one another’s cognitive role when needed?
The Lineage question is whether backup is accumulated history or the four-turn scaffold.Experienced teams often need fewer words. Technical term: communication compression / shorthand
Does less need to be said without losing important caveats?
Compression was observed descriptively. A team interpretation remains untested.08
The failures are part of the instrument’s history, not the qualifying result.
PB-GEN-01/02/03 exposed distinct layers of identity, stored history, trajectory, and provider-native continuity. They remain preserved as commissioning evidence; PB-GEN-04/05/06 are the qualifying scientific observations.
09
What the record does not establish.
- Formation history caused any observed Phase B behavior.
- Any pair developed a personality, preference, relationship, bond, trust, attachment, consciousness, or team identity.
- Pair-specific trajectories are stable or causal.
- Ada independently verified Wren's evidence.
- Exact phrase inheritance persisted across phases.
- Phase C has begun or its protocol is ratified.
Contract V2: Formation → Generalization also included a disclosed provider/runtime handoff. Cross-boundary change therefore cannot isolate information environment or shared history as the cause.
Evidence ladder
How far the current record actually climbs.
- Observed difference → YES
- Recurs across tasks → YES — some recurring candidates
- Distinguishes this pair → NOT ESTABLISHED
- Survives partner/history change → NOT ESTABLISHED
- Causally linked to shared history → NOT SUPPORTED
Different is not necessarily durable. Durable is not necessarily unique. Unique is not necessarily pair-specific. Pair-specific would still not prove that shared history caused it.
This is a bound, not a score. Descriptive regularities were learned. Zero supported findings means zero supported causal pair-history findings.
Questions we are now asking
The evidence made the next experiment more specific.
- Do different pairs develop different ways of correcting the same kind of error?
- Does Wren begin anticipating her particular Ada rather than the generic role?
- Does a conceptual distinction survive when its exact wording disappears?
- Do partners begin performing pieces of one another’s cognitive jobs?
- Can two pairs reach the same answer through reliably different corrective routes?
- Which apparent differences disappear as soon as the topic changes?
- Does knowing a collaborator eventually become visible in what no longer needs to be said?
Next · not started
Shared-History Change asks whether accessible accumulated history is doing causal work.
The intended contrast follows directly from the uncertainty above. Detailed interventions, assignments, measures, and authority remain subject to prospective governance. Phase C has no execution authority.
Meaning first. Evidence immediately behind it.
Inspect the method and sealed record.
Methods · Evidence · Technical archive · Canonical current-study projection · Cross-phase analysis record · Formation report · Generalization report