docs: README and About copy for the tuned additive model

The idea table now shows 83/69/18 (the soft-program default) with a 'who you
teach matters more than how many' framing; the model section describes the
forwarding composite (baseline, raised by novelty, lowered by harm awareness)
and the soft education lever instead of the old fixed-rate / never-forward
wording. Background acknowledges the team's literature sourcing of the
parameters and frames this demo as the accessible layer of the funded
instrument (calibration, full parameters, optimisation). Preset About copy
updated to match.

README.md

@@ -24,26 +24,34 @@ the *same* outbreak in the *same* simulated social network under different
 education strategies, so the only thing that changes is who gets educated.
 With the default world (120 students, educate 30% of them):
 
-| Strategy                   | Reached by the fake |
+| Strategy                       | Reached by the fake |
-| -------------------------- | ------------------- |
+| ------------------------------ | ------------------- |
-| No program                 | 82%                 |
+| No program                     | 83%                 |
-| Educate 30% at random      | 58%                 |
+| Educate 30% at random          | 69%                 |
-| Educate the 30% best-connected | 6%              |
+| Educate the 30% best-connected | 18%                 |
 
-Same budget, different targeting, an order-of-magnitude difference. Making
+Same budget, very different outcomes: educating at random barely dents the
-that lever visible (and later: searching for the best intervention under a
+spread, while educating the best-connected students contains it to under a
-budget) is the point of the tool.
+fifth of the school. Who you teach matters more than how many. Making that
+lever visible (and later: searching for the best intervention under a budget)
+is the point of the tool.
 
 ## How the model works
 
 - **Network**: Holme-Kim preferential attachment with triangle closure
   (Holme & Kim, 2002), producing the hubs and clustered friend groups of
   real social networks; a port of networkx's `powerlaw_cluster_graph`.
-- **Spread**: an independent cascade (Kempe, Kleinberg & Tardos, 2003);
+- **Spread**: an independent cascade (Kempe, Kleinberg & Tardos, 2003). Each
-  each directed edge gets one random forwarding draw, shared across all
+  student's chance of forwarding is a literature-grounded composite: a baseline
-  scenarios, so strategies are compared in the same world.
+  rate, raised by how novel or shocking the fake is, lowered by the year
-- **Education lever**: an educated student still receives the fake but
+  group's ambient harm awareness. Per-edge random draws are shared across all
-  never forwards it. Strategies: no program, uniform random, most-connected.
+  scenarios, so strategies are compared in the same world. Because
+  deepfake-specific data are scarce, the parameters are sensitivity ranges
+  rather than fitted values, drawn from the project team's literature review.
+- **Education lever**: the program cuts an educated student's forwarding
+  probability by its effect size (a strong but imperfect reduction, not a
+  perfect wall). Strategies pick who is educated: no program, uniform random,
+  or the most-connected students.
 - **Determinism**: every source of randomness flows from seeds in the
   config; identical configs produce identical results, pinned by tests.
 
@@ -138,6 +146,16 @@ Python prototype used in that project's pitch. The subject is handled from
 the prevention side only: the tool models how harmful content spreads and
 what education changes, nothing about creating such content.
 
+The model's diffusion and education parameters were sourced by the project
+team from open-access literature on misinformation spread, complex contagion,
+peer aggression, and comprehensive sexuality education.
+
+This public demo is the accessible, intuition-building layer of the proposed
+research. The funded project would build the full instrument behind it:
+calibrating the model to real-school data, incorporating the complete
+literature parameter set, and searching for the best intervention under a
+budget.
+
 References: P. Holme & B. J. Kim, *Growing scale-free networks with tunable
 clustering* (Phys. Rev. E 65, 2002). D. Kempe, J. Kleinberg & E. Tardos,
 *Maximizing the spread of influence through a social network* (KDD 2003).

web/src/presets/deepfake-school.ts

@@ -13,12 +13,13 @@ export const deepfakeSchoolPreset: StudyPreset = {
   disclaimerShort: 'Illustrative model, not validated',
   disclaimerLong:
     'spreadlab runs a seeded agent-based toy simulation: a synthetic ' +
-    'friendship network of one school year group, a fixed chance to forward ' +
+    'friendship network of one school year group, a chance to forward the ' +
-    'the fake along each friendship per round, and an education program that ' +
+    'fake along each friendship that rises with how novel it is and falls ' +
-    'teaches some students to refuse. Its parameters are chosen for ' +
+    'with the ambient harm awareness, and an education program that strongly ' +
-    'illustration, not fitted to data, so it is not a validated prediction ' +
+    'but imperfectly reduces forwarding for educated students. Its parameters ' +
-    'of any real school. Use it to build intuition about who to educate, ' +
+    'are illustrative sensitivity ranges, not fitted to data, so it is not a ' +
-    'not to forecast outcomes.',
+    'validated prediction of any real school. Use it to build intuition about ' +
+    'who to educate, not to forecast outcomes.',
   readingCaption:
     'Each dot is one student in the same simulated year group. Every scenario ' +
     'runs the identical school; only the education program differs.',