Research

ComputerWorld can hold simulated buyers, readers and executives, give them information, fork the world before an intervention and compare what each branch does. These studies use that to ask how corporate identity, articles, pitches and comparative claims move simulated markets. Every person in them is a language model playing a persona and every company is fictional: the results are measurements of a simulator, labelled as such, with the human studies that would test them listed on each page.

Each study ends with the observations that would calibrate it. Across them the order is the same: run the same stimuli past real respondents (a randomised conjoint or vignette panel), compare the simulator's direction and ranking with theirs, then fit the factor model's intercepts to field outcomes (click-through, quote and order rates from a CRM) before trusting any number as a forecast.

Each factor is a JSON file naming its definition, the entities it applies to, its inputs, a typed output, its dependencies, how it is measured, where it came from and how far it has been validated. The measurement kind is never blurred: an observed factor is something a simulated agent did or was assigned; derived is a deterministic function of other values; llm-elicited is a model's report when asked; calibrated would be fitted to real outcomes and none is; a hypothesis is defined and not yet measured.

Machine-written candidates go through the same gate as hand-written ones: schema checks, deduplication against the registry by name and definition, and a status reset to unvalidated. A factor becomes benchmarked only when adding it to a study's baseline improves prediction on held-out scenarios, and null when it does not. The two ablations, baseline factor set against an expanded one, held-out log loss:

• Brand priors, leave-one-identity-out: 0.489 → 0.410 (log-loss reduction +0.079, 95% interval +0.036 to +0.117). The expanded set helps.
• Comparative positioning, leave-one-scenario-out: 0.292 → 0.316 (log-loss reduction −0.023, 95% interval −0.046 to −0.004). It does not: with six scenarios, scenario-level factors are not identified.

Neither shows that the factors predict real buyers; both models are fitted to one language model's answers. A larger registry is not more evidence: a factor counts when it carries held-out information, and most here do not yet.

python scripts/society/factors.py validate                      # load and check every factor
python scripts/society/factors.py propose candidates.json        # gate candidates (add --write to keep them)
python scripts/society/factors.py generate --context "..." --journal J   # ask a model, through the same gate

All 68 factors

Id	Name	Kind	Status	Applies to	Output
attention_displacement	Attention displacement	derived	unvalidated	buyer, artifact	number -50–50
claim_verifiability	Claim verifiability	derived	unvalidated	message	bool
commercial_intent_reader	Commercial-intent reader	derived	unvalidated	buyer, company	bool
counterfactual_intent_lift	Counterfactual intent lift	derived	unvalidated	artifact	number -1–1
endgame_engaged	Engagement end-game	derived	unvalidated	organization, decision	bool
evaluation_reply	Evaluation reply	derived	unvalidated	buyer, artifact	bool
expectation_gap	Expectation gap	derived	unvalidated	product, buyer	number -6–6
exposure_channel	Exposure channel	derived	unvalidated	buyer, artifact	category
linear_attribution_share	Linear attribution share	derived	unvalidated	artifact, buyer	number 0–1
name_descriptiveness	Name descriptiveness	derived	unvalidated	company	bool
name_pronounceability	Name pronounceability	derived	unvalidated	company	number 0–1
name_syllables	Name syllable count	derived	unvalidated	company	count 1–∞
price_ratio	Price ratio to reference	derived	unvalidated	product	number 0–10
protocol_compatibility	Protocol compatibility	derived	unvalidated	product, buyer	number 0–1
stakeholder_disagreement	Stakeholder disagreement	derived	unvalidated	organization, decision	number 0–3
switching	Switching	derived	unvalidated	buyer	category
brain_ltv	Brain-LTV	hypothesis	unvalidated	artifact	number -100000000–100000000
category_prototype_fit	Category prototype fit	hypothesis	unvalidated	company	scale 1–7
name_memorability	Name memorability	hypothesis	unvalidated	company	probability 0–1
value_per_intent_reader	Value per commercial-intent reader	hypothesis	unvalidated	company	number -100000–100000
anticipated_maintenance_burden	Anticipated maintenance burden	llm_elicited	benchmarked	product	scale 1–7
confidence_in_leadership	Confidence in leadership	llm_elicited	null	company	scale 1–7
expected_failure_rate	Expected three-year failure rate	llm_elicited	null	product	number 0–100
expected_product_maturity	Expected product maturity	llm_elicited	null	product	scale 1–7
expected_quality	Expected quality	llm_elicited	null	product	scale 1–7
expected_vendor_viability	Expected long-term vendor viability	llm_elicited	null	company	probability 0–1
feature_salience	Feature salience	llm_elicited	null	buyer	number 0–100
institutional_credibility	Institutional credibility	llm_elicited	null	company	scale 1–7
name_institutional_association	Institutional association of the name	llm_elicited	null	company	scale 1–7
name_premium_association	Premium association of the name	llm_elicited	null	company	scale 1–7
overall_impression	Overall impression	llm_elicited	unvalidated	company, product	scale 1–7
perceived_competence	Perceived competence	llm_elicited	null	company	scale 1–7
perceived_technical_depth	Perceived technical depth	llm_elicited	null	company	scale 1–7
perceived_vendor_risk	Perceived vendor risk	llm_elicited	benchmarked	company	scale 1–7
post_purchase_satisfaction	Post-purchase satisfaction	llm_elicited	unvalidated	product, buyer	scale 1–7
procurement_friction	Procurement friction	llm_elicited	benchmarked	company, buyer	scale 1–7
relative_preference	Relative preference	llm_elicited	unvalidated	product, buyer	scale -3–3
source_independence	Perceived source independence	llm_elicited	benchmarked	message	scale 1–7
stakeholder_cost_justification	Stakeholder-rated cost justification	llm_elicited	unvalidated	executive	scale 1–7
stakeholder_strategic_value	Stakeholder-rated strategic value	llm_elicited	unvalidated	executive	scale 1–7
stakeholder_support	Stakeholder support for engaging	llm_elicited	null	executive	scale 1–7
willingness_to_pay	Willingness to pay	llm_elicited	benchmarked	product, buyer	number 0–10000
artifact_engagement	Artifact engagement	observed	unvalidated	buyer, artifact	count 0–∞
artifact_exposure	Artifact exposure	observed	unvalidated	buyer, artifact	bool
artifact_impressions	Artifact impressions	observed	unvalidated	buyer, artifact	count 0–∞
binding_veto_incidence	Binding veto incidence	observed	unvalidated	organization, decision	count 0–10
budget_headroom	Budget headroom	observed	null	buyer	number 0–10000
buyer_risk_tolerance	Buyer risk tolerance	observed	null	buyer	category
buyer_segment	Buyer segment	observed	unvalidated	buyer	category
company_follow	Company follow	observed	unvalidated	buyer, company	bool
competing_build_proposal	Competing internal-build proposal	observed	unvalidated	organization, decision	bool
competing_team_exposure	Competing-team exposure	observed	null	organization, pitch	bool
decision_latency_weeks	Decision latency	observed	unvalidated	decision	number 1–10
diligence_breadth	Diligence breadth before proposal	observed	benchmarked	organization, decision	count 0–4
diligence_initiation	Diligence initiation	observed	unvalidated	buyer, product	bool
engagement_request	Engagement request	observed	unvalidated	company, buyer	bool
exec_forwarding_breadth	Executive forwarding breadth	observed	null	organization, pitch	count 1–5
focal_choice	Focal product chosen	observed	unvalidated	product, buyer	bool
identity_arm	Identity arm	observed	unvalidated	company	category
institutional_endgame	Institutional end-game	observed	unvalidated	organization, decision	category
listing_discovery	Listing discovery	observed	unvalidated	company, buyer	bool
need_fit_latency	Workload latency sensitivity	observed	null	buyer	bool
pitch_positioning	Pitch positioning	observed	unvalidated	pitch	category
price_treatment	Price treatment	observed	unvalidated	product	number 0–10000
proposal_contestation	Proposal contestation	observed	unvalidated	decision	count 0–10
repost_propagation	Repost propagation	observed	unvalidated	artifact	count 0–∞
repurchase_choice	Repurchase choice	observed	unvalidated	product, buyer	bool
true_feature_advantage	True feature advantage	observed	null	market	category