Wiki · Evidence & Verdicts
G4 allostasis — horizon-limited, does NOT clear on the current model (2026-07-11)
How to read this page
Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.
Eighty-seven dated pages: receipts, pre-registrations, handoffs, validation records and review verdicts. A receipt is written at the moment a piece of work was checked. It names what was claimed, the commit and the seed, what was actually run, and the outcome in one of a small set of controlled words. Then it names what the work did not achieve. That last part is what makes it a receipt rather than an announcement. A pre-registration is the same discipline run in advance: the conditions that would count as a pass and the conditions that would falsify the claim are written down before the run, so neither can be adjusted once the numbers arrive.
That is why so many small dated stubs are an audit trail rather than noise. No one of them is meant to be a good read. The value is in the sequence and in the dates, because you can watch a prediction be registered, then the run happen, then the verdict land — sometimes against the prediction. Pages here record a falsified result, a rejected fix, a retracted overclaim, and a green receipt that turned out not to be reproducible from the commit that carried it. A record that carried only successes would be worth a good deal less than this one.
A gentle way in is to read a pre-registration first, so the shape becomes familiar, then a result page, then one of the corrections. This section sits off the main navigation on purpose: it is the record you check the rest of the site against, not the place to begin.
What it is not: documentation, and not a summary. Nothing here has been tidied in hindsight. Every entry reads as of its date, a later entry may overturn an earlier one, and the presence of a page is not a claim that its result stood.
Your browser cannot switch reading levels, so the document itself is shown.
Precise — the source document
This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.
Finding: the registered G4 allostasis gate (depth-5 forage-trigger ≥ depth-1 + 1 — deeper planning forages earlier/higher = anticipatory regulation) does not clear on the live metabolism model, at any phase or depth. Honest FAIL-as-designed; recorded so the gate is not silently skipped.
Evidence: runs/g4_sweep.exs (deterministic; mix run --no-start). Forage-trigger energy bin:
| phase | depth-1 | depth-3 | depth-5 | depth-7 | depth5−depth1 |
|---|---|---|---|---|---|
| 0 | bin 2 | bin 2 | bin 2 | bin 2 | 0 |
| 1 | bin 2 | bin 2 | bin 2 | bin 2 | 0 |
| 2 | bin 2 | bin 2 | bin 2 | bin 2 | 0 |
Depth-independent: the agent forages at "ok" (bin 2, the setpoint peak) regardless of lookahead. There is no anticipatory (depth-dependent) foraging.
Root cause (structural, matches metabolism.md §12): at upkeep 0.04/tick, going from "ok" (bin 2) to
"empty" is ~15 ticks; the depth-5/7 beam horizon (5–7 ticks) does not reach the depletion cliff, so the
planner cannot see the danger to forage against it. The setpoint-C already triggers reactive foraging at the
setpoint; deeper planning adds nothing because the cliff is out of view. The §12 reduced demo showed G4 can
separate only for work_bonus ≳ 4.0; the live setpoint-C map does not provide that separation.
What would clear G4 (the fix — a structural change, deferred to the hierarchy program): a slow-context /
L2 signal that carries the depletion pressure into the plan (so anticipation does not require the cliff to be
inside the beam horizon), or a deeper effective horizon. hierarchy2.ex exists but is "NOT wired into the
live decide path" — wiring an interoceptive slow-context is the concrete G4-clearing work, part of the
tail-up developmental-ladder / A4 program.
Consequence for "one more gate tonight": G4 is not clearable without that structural change, so the next
clearable registered gate is the strengthened G5a interoceptive-regulation gate (drive-on vs drive-severed
control) — see docs/receipts/metabolism_regulation_gate*.
Claim fence: mechanism only; this is a planning-horizon property of a model, never experience.
sha256 0f0db0ca8afb16bb — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This page records a gate that does not pass, and says so on purpose. The idea under test was that planning further ahead should make the system top itself up earlier, which would be anticipation rather than reaction. Across several planning depths, it made no difference at all. The reason turns out to be structural: the danger the plan would need to see lies further ahead than the plan can look, so looking deeper adds nothing. The result is written down rather than quietly skipped, and the page says what kind of change would be needed to clear it.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 0f0db0ca8afb16bb
Clear — written for this website, not the source document
An adverse result, recorded deliberately so that a gate does not silently disappear from the list. The claim under test was that deeper planning would trigger the top-up behaviour earlier. It does not, on the current model, at any of the depths or phases tried.
The evidence is a small deterministic sweep, and its table is flat. The trigger sits in the same bin whatever the planning depth, and the difference between the deepest and the shallowest is zero in every row.
The root cause is offered as structure rather than as a guess, and it matches an existing design note. At the drain rate in use, the fall from comfortable to empty takes far longer than the planner can see ahead. The cliff is outside the window, so a deeper plan has nothing extra to react to, while the existing reactive trigger already fires at the setpoint. A reduced version described in that same note had shown the effect can separate only under a much stronger bonus than the live map provides.
The page then names what would clear the gate: carrying the pressure into the plan through a slower signal, so that anticipating does not require the danger to sit inside the planning window, or else a longer effective horizon. The module for that exists but is not wired into the live path, so this is deferred work rather than a fix waiting to be applied. A closing note names a different gate as the next one that can be cleared, and a stated limit keeps the finding to mechanism: a property of a planning horizon in a model, never experience.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 0f0db0ca8afb16bb