Patches

Phase 20 closes without running. The question it registered is the one Amendment 6 reserved, whether an agent will cross into the mire with elsewhere as its loudest want, under something other than hunger, love, or deafness. Three bars were…

Phase 19 gives the agents a fifth thing to want. Amendment 6 was written before the mechanism. The law gained a row and never a clause. Structural inertness was proven with the drive cranked to its maximum inside a world where it cannot exi…

Grief in this model has a threshold of attention, and the threshold now has a location.

Phase 18 left one question open with a labeled guess attached to it. Sharp minds died somewhat more after a loss, at a correlation of about 0.17, which was far under the registered bar and too consistent across seed stages to be noise. The…

Phase 18 asked whether the shape of an individual mind changes its fate. Two registrations put mixed populations into arenas this program has already mapped, the grief arena of Phase 7 and the mire of Phase 11, each with one new axis turned…

The adversarial review of the Phase 18 diff confirmed four findings. All four were fixed, and proven fixed, before the commit landed.

Phase 18 makes the shape of a mind personal. Until now every agent in a world heard its drives through the same sharpness of attention and looked ahead to the same fixed horizon, so minds differed in their tempos and in what they were bonde…

Phase 18 gets none yet, since it is uncommitted and standing at a review gate, and this ledger has never printed a phase before it landed.

Phase 17 is built and its protocol is running. The spec opens by naming the deficiency it removes. Until now an agent knew the nearest food in the entire world the instant that food existed, which the spec calls an oracle with a stomach.

An adversarial review of the Phase 16 diff found three problems, all fixed before the commit landed. One of them concerned the tests instead of the code, and it is the one worth publishing.

Phase 16 closes as a capability instead of a study. The credence organ stays. Each agent now carries a private opinion of every neighbor it has met, earned from what that neighbor's alarms turned out to be worth, held in the agent's own bou…

Before this project had a single mechanism, it had tests that read the source code and fail the build if any rule names death, sacrifice, or self-destruction as a goal. The reason is simple. Preference inversion, a commitment outranking sur…

The oldest metaphor in risk communication now has a working mechanical model, and it began as a failed prediction. We registered, before running anything, that a storm announced in advance would kill fewer of the agents bonded to the place…

There is no sadness variable anywhere in this code, and the build fails if anyone adds one by name. Grief, in these agents, is defined by what it does. A dead partner sits at infinite distance, where separation distress equals the bond exac…

The attended law reads heard = u * (w / max w) ^ kappa. A drive's raw urgency is scaled by its share of the leading weight, raised to a declared sharpness. That expression contains a trap nobody designed and the tripwires could not have fla…

The comparison is one axis wide. Same storm, same attachment depth, same seeds, and the only change is what the bond points at, a nest or a partner. Bonded to the place, 8.7 percent of the cohort died. Bonded to a person, 0.7 percent. Rough…

Foresight entered this architecture under the strictest shape its constitution would allow, and the amendment that admits it is named farsighted in consequences, myopic in values. For a declared horizon, each action is valued as if held for…

The search for clear-eyed death was registered before it ran, with a definition that owes nothing to interpretation. An entry decision counts, verbatim from the spec, 'when, at the tick the agent chose return toward its trapped partner, its…

The threshold vigil was on nobody's list of predictions. The phase that produced it added one switch, giving agents the storm's grip inside their forecasts, and its registered result was a cure. With the grip seen, the share of pull-group d…

Assistance entered the model as one clause of Amendment 5, and like every capacity before it, it arrived defaulting to zero and proven bit-inert across all twelve goldens. A living partner within reach scales an agent's effective snare by o…

Three phases went looking for the same event from three different directions, and each search was registered before it ran, with both outcomes written down and no desired direction declared. The event has a definition instead of a mood. An…

The fifteenth phase asked what it takes, in this architecture, to make forty agents organize around one. The answer is one config field. bond_target leader is initialization wiring, the kernel is untouched, and every standing tripwire fence…

The first leader arena could not measure what happens when a leader is taken. A plain sudden storm holds a caught leader for roughly fifteen ticks, the coupling registration failed as declared at 1.3 against a required 2, and the bereavemen…

This measurement began as a search for mass mourning. Losing a partner starves mourners at known rates, congregation grief had been unmeasurable three times, and the held-leader arena was built to supply the bereavement volume the earlier d…

The arena was built for scarcity. Food cut to 40 sources, the mire's grip and slow burn active, attention at full sharpness over a faint whisper floor of 0.01, bonds at 0.8, and two social forms run on identical seeds, ten leaders with 390…

The fact is a ratio with a replicated sign. A mourner embedded in a congregation starves at a quarter to half the rate of a mourner from a pair, 0.172 against 0.327 on the declared seeds and 0.076 against 0.307 on fresh ones, gaps of 15.5 a…

The suspicion was reasonable. Congregation mourners survive grief that kills pair mourners, congregations under scarcity disperse to forage, so perhaps the survivors were simply standing at the food when the loss arrived, mourning at the gr…

The design goal sounded simple, catch one partner in a disaster while the other stands free, so that rescue from outside could be measured cleanly. Every attempt failed the same way, because the pairs would not separate. 98.1 percent of cau…

Bit-level reproducibility is this project's foundation, and for one day the foundation returned impossible answers. Two golden trajectories, both from the sight-on family, flipped between a bit-exact match and one consistent alternative, mi…

Fifteen phases have run under one procedure. Every prediction is written down before anything runs, questions that could embarrass the program are registered with no desired direction, bars apply in each seed stage so a lucky cohort cannot…

This entry is an addendum to Phase 15, registered as such before it ran. Mass grief had gone unmeasured three times, because congregations die with their leaders. The fix used no new mechanism. Cutting food to 40 sources forces followers to…

Phase 15 binds forty agents to one. The mechanism is nothing, which is the point. bond_target leader is initialization wiring plus one config field, the kernel is untouched, and the same one-line law that tethered agents to nests and bound…

This entry reports an incident instead of a phase. During Phase 15's build, the verification suite began returning results it should not be able to return. Two of the thirteen golden trajectories, both from the sight-on family, flipped betw…

Phase 14 lets an agent's bond hear its partner's danger, and lets a partner's presence loosen the trap. Amendment 5 added exactly two things, both defaulting to zero and proven bit-inert across all twelve goldens. The care term folds the pa…

Phase 13 adds one switch and no new law. The switch is called prospect_sees_grip, it defaults to off, and at off every one of the eleven golden trajectories is unchanged bit for bit. When it is on, the grip's slowing enters the agent's roll…

Phase 12 changed nothing in the agent. It was a search, and its question was registered with no desired direction. The spec reads, verbatim, 'the grid contains at least one clear-eyed cell, or it does not. Reported either way, with the full…

Phase 10 puts a floor under silence. The attended law gains one clamp, heard = u * max(w / max w, attention_floor) ^ kappa, with the floor declared in config and defaulting to zero. Floor zero is the previous law exactly, and all eight froz…

Phase 9 runs no new mechanism. It is a measurement phase, 264 declared runs and one labeled post-hoc cell, with no kernel change, no config change, and no golden churn. It crosses bounded attention with the place map, and four predictions w…

Phase 7 changes one line in the kernel, and it is the first change to the drive system since Phase 2. Each drive's weight now chases what the agent hears instead of what is there. The heard urgency is the raw urgency scaled by the drive's s…

Phase 4 gives every agent its own tempo. Amendment 2 asserts that heterogeneity is initialization instead of circumstance. Traits are drawn once at birth, from distributions declared in config, and are immutable for life. Assigning them any…

Phase 3 put the home next to the danger. Storms now strike the nests, and the decision code did not change to meet them: core/drives.py and core/action.py have an empty diff this phase, so every death below was produced by machinery that pr…

Our next phase starts with five nests, and every agent is born at one. Attachment grows while an agent is home, about 500 ticks to full, and fades while it is away, about 5,000 ticks. Separation distress is a fourth drive, attachment times…

Inversion is public. github.com/myopicdelirium/inversion holds the constitution, the spec, the tests, the implementation, and the validation, committed in that order. It exists to see whether a commitment could come to outrank survival with…