Model character cognition separately from system intelligence

This commit is contained in:
Meta Vens 2026-05-04 19:45:00 +01:00
commit b8f671b843
6 changed files with 63 additions and 0 deletions

View file

@ -101,6 +101,35 @@ The sequential runtime still owns:
The projector owns the translation from Ghostlight's internal state language to
character-local operating prose.
## Character Cognitive Ceiling
The response model may be very capable. The character should not inherit that
capability by default.
Projection must carry a character-local cognitive ceiling: how quickly the
character infers, how much social indirection they notice, how well they
manipulate attention, how abstractly they reason under stress, what education or
culture has trained them to see, and where their blind spots live. A Lucent
crisis professional may operate with alarming metacognitive precision because
that is part of the institution's ecology. A tired dock worker, sheltered child,
panic-struck survivor, credulous fan, or blunt military specialist should not
suddenly perform the same level of social chess just because the underlying
model can.
The projector should therefore render:
- strategic acuity and social-reading ability
- education, professional training, and cultural scripts
- stress effects on reasoning
- blind spots, taboos, and habitual misreads
- whether the character can explain their own tactic, only feel toward it, or
act it out without understanding it
This is separate from narrator intelligence. A coordinator or narrator can help
the reader understand a maneuver after it happens. That explanatory prose must
not be smuggled into the character's private state or future action capacity
unless reviewed mutation says the character actually learned it.
## Current Limit
The first projector is deterministic and deliberately simple. It uses threshold

View file

@ -190,6 +190,15 @@ Glue prose is part of the artifact. It is not canonical state, but it is useful
training signal for how scenes are bridged, how pressure is restated without
exposition sludge, and how a procedural branch becomes readable narrative.
Coordinator and narrator intelligence are allowed to exceed character
intelligence. This is how opaque high-context scenes become legible without
flattening the characters. If a Lucent negotiator, influencer, or feed operator
is performing layered reputation warfare, the coordinator may explain the
maneuver in reader-facing prose, but the explanation is not automatically
available to the character who performed or witnessed it. Character cognition
still comes from projected state: intelligence, social acuity, training,
cultural scripts, stress, and blind spots.
The coordinator may propose state changes, scene transitions, and tool calls. It
does not bypass deterministic gates. It does not silently decide hidden facts. It
does not mutate another organ's private output after review. It conducts the

View file

@ -167,6 +167,14 @@ constraints, route access, public categories, and resource pressure; let actors
choose from local state; then fold the observed action into reviewed state.
Elastic strings, not rails. Little puppet theater, but with receipts.
Character-intelligence lesson: keep the machinery as smart as possible, but do
not make every character inherit frontier-model metacognition. Characters should
reason, manipulate, misread, explain themselves, and regulate stress only to
the level supported by their stat sheet, culture, role, experience, and current
state. The coordinator/narrator may translate high-context maneuvers for the
reader, especially in Lucent-style reputation warfare, but that explanation is
not automatically character knowledge.
Coordinator seam correction: the Lucent character responders were sandboxed
no-fork agents, but the coordinator was Codex-authored and then captured as
artifacts. Future training-shaped loops should sandbox coordinator turns too.

View file

@ -56,6 +56,16 @@ Apply these principles:
- Clarity under invented terms: invented factions, institutions, technologies,
and social movements need enough in-text context for a new reader to infer
what they represent in the current scene.
- Legible social maneuvering: if characters are operating with advanced
strategy, metacognition, or cultural tactics, the reader needs enough
narrator/coordinator framing to understand what changed and why it matters.
Do not require every reader to be as socially fluent as the smartest
character in the scene.
- Character cognitive fit: smart prose is not the same as smart characters.
Check whether each character's inference speed, manipulation skill,
self-awareness, education, and strategic patience fit the state implied by
the artifact. High-IQ social chess can be correct for some cultures or roles,
but should not leak into every character by default.
## Aetheria Tonal Range
@ -113,6 +123,8 @@ Review the artifact against these questions.
- Are invented terms explained by context before they matter?
- Does the story rely on knowing what a faction represents before the prose
gives the reader usable cues?
- Are socially complex moves translated for the reader when needed, without
forcing characters to explain tactics they would not consciously articulate?
- Does the opening establish place, time, body affordances, work hierarchy, and
local stakes without drowning the reader?
@ -165,6 +177,8 @@ Review the artifact against these questions.
- Is the chosen tone appropriate to the character, scene, and Aetheria source
material, or is the artifact defaulting to dry technical severity because the
machinery is more comfortable there?
- Are any characters performing above their modeled intelligence, social
training, emotional regulation, or stress capacity?
### Tonal Fit

View file

@ -53,3 +53,4 @@
{"ts": "2026-05-04T16:09:04+00:00", "type": "implementation", "status": "ok", "note": "Extended lucent-hostage-feed-v0 through safe-line confirmation and non-contact protective handoff. Key Ghostlight machine lesson: the coordinator should nudge with game-ergonomic affordances and state gates rather than force outcomes; actors choose locally within elastic constraints and observed actions fold through reviewed mutation."}
{"ts": "2026-05-04T17:14:29+00:00", "type": "architecture", "status": "ok", "note": "Coordinator seam correction: future training-shaped loops should sandbox coordinator turns as their own promptable organ. Codex acts as meta-coordinator, preparing inputs, launching/reviewing workers, labeling interventions, and wiring structured state; raw coordinator training data must not be silently collapsed into omniscient chat steering."}
{"ts": "2026-05-04T18:00:02+00:00", "type": "maintenance", "status": "ok", "note": "Dedicated Perfect Machine memory pass: distilled global Epiphany from origin narrative into operational doctrine, added project-local persistent-state stewardship rule, and sharpened Ghostlight map around sandboxed coordinator seam and current fixture truth."}
{"ts": "2026-05-04T18:44:45+00:00", "type": "architecture", "status": "ok", "note": "Character cognition seam: Ghostlight machinery should be maximally smart, but characters must reason and socially maneuver only to the level supported by their modeled intelligence, culture, role, experience, stress, and state. Coordinator/narrator prose may translate opaque high-context tactics for reader legibility without granting that explanation to characters."}

View file

@ -38,6 +38,7 @@ hard_boundaries:
- Canonical agent state and per-agent perceived state must remain distinct.
- Characters act from local knowledge, local perception, local pressure, and bounded affordances.
- Aetheria prose must support tonal range and ordinary life, not only crisis pressure. Wit, warmth, domesticity, ritual, boredom, wonder, and quiet dread are valid state-grounded modes when they fit the scene.
- Character intelligence is modeled state, not a default model capability. Make the Ghostlight machinery as smart as possible, but project each character's cognition, metacognition, manipulation skill, inference speed, education, and blind spots only to the level supported by their stat sheet, culture, experience, stress, and current state.
- Listener misread is first-class; intent and interpretation are not the same object.
- Human-facing appraisal is a first-class target; VoidBot must be able to read people, form grounded opinions, and revise those opinions over repeated contact.
- Generated action is an event proposal; deterministic gates and reviewed mutation decide what becomes state.
@ -77,6 +78,7 @@ live_architecture:
- Future coordinator turns should be sandboxed as their own promptable worker outputs, with exact visible input, raw output, parsed coordinator artifact, review labels, and interventions preserved. Codex acts as meta-coordinator: it scopes, launches, reviews, repairs, and wires Ghostlight organs, but should not silently collapse the trainable coordinator organ back into omniscient chat.
- Glue prose is training signal, not canonical truth.
- Coordinator prose should choose an intentional tonal mode for the scene and protect breathing room when the story needs routine, levity, domestic texture, wonder, or slow dread before consequence.
- The coordinator/narrator may be smarter than the characters for reader service: explain opaque maneuvers, stakes, institutional games, and social tactics in accessible prose without retroactively making the acting characters more self-aware than their state allows.
- The coordinator nudges by affordance, not decree. Prefer concrete game-ergonomic levers such as safe lines, export locks, security holds, handling categories, visible posture scripts, route access, resource scarcity, clocks, and public proof objects over authorial pressure toward a desired ending.
branch_compiler:
- The branch compiler turns fixture briefs, lore digests, scene/world state, actor-local branch candidates, reviewed consequence packets, fold plans, and optional visual replay requirements into playable Ink plus training sidecars, separate visual plan artifacts when illustrated replay is in scope, and compiler notes.