A NaN or negative rate (a NaN DrainTrait made Resolve NaN and Tick threw from
PoseBehaviour.Write after the drive had half-stepped) is now a construction
error. DriveParameters.Validate owns the checks, WillMendsAt and WillCurve
included. Addresses finding nan-drive-throws-after-step.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Layer 3 (Breaks, WillMendsAt, Will with its Attention and Consciousness
slots) and layer 4 (EntryResolve, Effort, InvoluntarySeconds, the strength
rule) in docs/lodur-target.md, and the README's one sentence.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A goal that is voluntary (past its InvoluntarySeconds) writes each muscle at
HoldStrength x min(1, Will^WillCurve + leak x FlinchGain); an involuntary one
writes HoldStrength. PoseBehaviour.Write takes the weight first and throws
outside [0, 1]; targets (Blend) are unchanged. DriveParameters.WillCurve
(default 0.25) is validated in the Actor constructor.
Body tests re-pinned, with the reason:
- seed-1 replay: break tick 1770, switch tick 1782 and 16 flinches are
unchanged; the arm's track and world hash moved (hand e283e3615b6b2cbc ->
a0fafe85c80e98f7, world cd1902f3dd65af64 -> 83155f2abf0e7e91) because the
hold now weakens as resolve falls. Renamed from "pre_brain_base".
- "the movement is the leak": a leak-free arm now also sags as the will
fails, and loses the last flinch (1750 ms) because it leaves the box sooner.
The leak is isolated with WillCurve 0.001 (weight ~1), the claim unchanged;
a new test pins the sag itself.
The golden table is untouched.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
GoalSpec.EntryResolve sets resolve on the tick the goal takes over (then the
mend check runs); Effort is resolve drained per second to sustain the goal,
paid only once it is voluntary and scaled by (1 - Pull); InvoluntarySeconds
is the onset that runs without will (+infinity throughout). Brain.Validate
refuses out-of-range values naming the goal. The pain-box fixtures mark their
release goal involuntary throughout. The behaviour weight arrives next.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Resolve keeps evolving through a break. A break fires on each tick resolve
reaches zero while the will is not broken (counted in Breaks); a broken will
mends when resolve climbs back to DriveParameters.WillMendsAt (default 0.25).
WillBroken is now "broken now"; DriveQuantity.Breaks is the one-way "has
broken". Will = Resolve x Attention capped by Consciousness, the two being
slots that read 1.0 until lodur-startle-attention and lodur-physiology.
Reflex suppression and the flinch gate read Will, not a WillBroken special
case. The pain-box fixtures gate their release on Breaks >= 1. The golden
table is untouched and still byte-identical.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A goal's minimum dwell now gates only decisions whose target does not strictly outrank it; a strictly
higher-priority goal preempts after the latency. GoalSpec.When, PoseBehaviour.Keys and FlinchPose read back as
read-only views (the Actor's path keeps its own arrays). Infinite dwell is refused. Tests pin the exact
boundary tick, the dwell index, the rounding rule, time-in-goal from takeover, preemption both ways and that
writing through a getter is impossible.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
GoalSpec.MinDwellSeconds (default 0) holds a goal until it has been active that many ticks, so a condition
hovering at its threshold no longer chatters. Brain, GoalSpec and PoseBehaviour copy the lists and arrays
they are given. New tests pin every drive quantity a condition reads, Below's strictness, the one-tick latency
floor and negative refusal, that the pending goal's behaviour is not written during the latency, that a
behaviour's phase restarts from the time in the goal, and the dwell and copy rules. Defaults leave the
golden table and the connected world hash unchanged.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The Actor takes a Brain (goals with id, priority, pull, conditions and a behaviour) and switches to the
highest-priority eligible goal after the Brain's deliberate latency, firing GoalChanged. Lodur no longer
has a hold/withdraw rule: Goal, Actor.Goal, TickEvents.Withdrew, ObjectivePull and
WithdrawalParameters.LatencySeconds are deleted; resolve drain takes the active goal's pull.
HoldAgainstReflex becomes PoseBehaviour (a pose held or a loop of keys). The pain box is a two-goal
Brain whose golden table (420 rows) and connected world hash are unchanged.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Office emitters are sensed at the head. RegionCount is 3; the connected pain box still replays the base world hash, break tick and hand track.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Break tick, deliberate switch tick, flinch count, hand track hash and world hash recorded at 5d0fec8, so the Brain cut can show it drives the body with the same bits.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Arousal rises under load and decays when quiet, drains resolve beside pain and strain
(scaled by pull and trait), and resolve recovers only when pain and arousal are both low.
Actor.Tick gains a trailing emitter span; the pain box's expressions keep their operand order
and the 420-row golden table passes unchanged.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Per channel the largest emitter intensity at the sense region's point goes through the same
threshold, sensitisation and conduction delay as noxious input; the channels saturate as
1 - prod(1 - afferent). Without senses the loop does nothing and the pain box is untouched.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The record is the one transduction for every channel from the sensory-load cut on, noxious or not.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Soul lows on cut-nervous-layers.s2: shake-test-has-no-upper-bound (0.06 and
0.15 now fail at 7.7 and 14 cm), golden-tsv-eol-unpinned,
golden-generator-trusts-head.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ymir 0b0321a makes the SoA world public and locks world create/destroy. The ragdoll's
Build takes the caller's Box3DSoaWorld and Lodur still never creates one. Deleted
AssemblyInfo.cs (InternalsVisibleTo), NativeWorld.cs and the collection attributes; first-slice.md
now records that no DLL rename is needed under MinGW.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
At the shipped StrainShake 0.03 the hand's height moves over 3 cm peak to peak per second
in seconds 8-13; at StrainShake 0 it moves under 1.5 cm. NativeWorld.cs now names
lodur:finding:cut-nervous-layers.s1.ymir-concurrent-world-create and says to remove it
once Ymir's lock lands.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
pain-box-golden.tsv holds BadDaySimulator 9f59764's PainBoxSimulation for 7 intensity/pull
settings x seeds 1-60: outcome, break and end ms, flinch count and a hash of the per-tick
stimulus/pain/strain/resolve/goal stream. The Actor reproduces it exactly.
tools/pain-box-golden regenerates it from that commit. New tests kill 'resolve recovery
off' and 'flinch reported after the break'.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Imagination's map for the nervous-layers cut, committed from the scratchpad so it is
durable.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
lodur-target.md becomes the r3 text: body on Ymir, layer type names and tick order,
one-physics-world restated with the game as owner, core tests build without Ymir.
README status drops 'layers 1-4 not coded'. first-slice.md records the MinGW Windows
build (DLL prefix workaround) instead of 'unrun'.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The Actor drives the ragdoll (pelvis welded to a static stool) through strengths and
targets only: the hand stays in the stimulus box until the will breaks, flinches pull it
away from a leak-free control, the broken will takes it out within 300 ms, same seed gives
the same events and world hash. Classes that build native worlds share one xunit
collection: two live Box3D worlds on different threads crashed the host (0xC000001D).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A hand is the tip of its lower-arm capsule (centre + rotation * (0, -0.14, 0)); the
ragdoll has no hand body. Allocates nothing. The visibility change (HumanoidRagdoll,
Muscle, Limb public; InternalsVisibleTo deleted) waits on Ymir's public SoA world.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
References only Lodur.Core. Covers conduction delay, sensitisation, containment,
reflex gain set from the previous tick's resolve, suppression cost, broken will,
ordered break times, stand-in output, zero allocation, no shared state, and parity
with the pain box (seed 1: will breaks at 14730 ms, withdraws at 14830 ms, 16 flinches).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Layers 1-3 and a layer-4 stand-in move in from the pain box (BadDaySimulator
PainBoxSimulation.cs at 9f59764), keeping its double arithmetic and in-tick order.
Actor owns one character's nervous state and runs perception, reflex, drive, behaviour
once per tick from caller-owned spans into caller-owned spans; it allocates only in its
constructor and never touches physics. Quat gains Nlerp.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Lodur.Core: engine-free PD law with per-muscle strength. Lodur.Body.Ymir: 13-body, 12-joint ragdoll
driven through the world's torque lanes from caller-owned strength and target arrays. Tests: stand, fall,
rest, per-muscle strength, torque conservation, no allocation, same seed gives the same state hash.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Licences (quoted from the NVIDIA Open Model Agreement), training-data
provenance, runtime needs, pose format, tracker sketch, verdict.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>