Skip to content

docs(learning): ADR 0025, place memory and brain capacity were tried and made no difference - #142

Merged
CodeByBryant merged 5 commits into
mainfrom
feature/place-memory
Nov 13, 2026
Merged

CodeByBryant merged 5 commits into
mainfrom
feature/place-memory

Conversation

@CodeByBryant

Copy link
Copy Markdown
Owner

Summary

The brains half of "smarter brains and a more complex world". Two claims were stated before the runs and tested with an identical-size control and the observatory (ADR 0024):

  1. Capacity does not matter. Config only, 8 seeds x 30,000 ticks: hidden 8 / 16 / 32 and memory 4 / 8 give population 61.4 / 62.9 / 62.1 / 63.2, and nothing on the observatory (plans, response types, intake shares) moves beyond the noise.
  2. A remembered feeding site helps foraging. I built it (a tenth input group, recall: the place of the last bite of real food as freshness, body-frame bearing and proximity; observation 48 to 52, weights 514 to 546, hash schema 13; the prior steered gently toward it), proved it sound (12 new tests with negatives; with the sense off, founder brains equal to the prior and mutation and learned inheritance frozen, 8 scenarios give identical snapshots through births, and the independent Python hasher agrees on the version 13 vectors), then measured it. It does not help: against the control with the four inputs zeroed (same 546-weight brain), the instinct-wired version made each creature eat 1 to 4 percent more but cost 13 to 14 percent of the population (50.4 against 58.5, and 48.1 against 55.9 on a second seed set) and 2 of 16 worlds died out against 0 of 16; the version that only gives the information, with no instinct, is indistinguishable from the control (recall input weight 0.065 after 30,000 ticks). The stated criteria (no extinct world, population not lower, starved share down 5 points or intake up 10 percent) all missed, so the mechanism was reverted in this branch (commit a741b65 in the history) and the result is recorded.

What is kept because it stands on its own: the observatory's inputGroupOf now throws for an observation slot with no group instead of silently counting it as threat (found while adding the recall group), with a test; and docs/learning/brain.md said the observation had 40 numbers where there are 48.

ADR 0025 has the full tables, the honest limits (one remembered site, a hand-wired return instinct, one world, 8 to 16 seeds; why the population falls is a hypothesis, not a finding) and a recommendation for what to try next: three null results in a row (the jaw's persistence, capacity, place memory) say the limit is the world, not cognition, so a changing, less visible environment is more likely to show a gain than another cognition feature.

Type of change

  • Feature
  • Bug fix
  • Refactor
  • Performance improvement
  • Tests
  • Documentation
  • Build/CI
  • Breaking change

Simulation impact

  • No simulation behavior changed
  • Changes organism behavior
  • Changes genetics or inheritance
  • Changes learning
  • Changes ecology or resources
  • Changes persistence or save compatibility
  • Changes determinism

Determinism

  • Uses the seeded random source
  • Does not use Math.random() in simulation code
  • Does not use wall-clock time for simulation behavior
  • Determinism tests were added or updated
  • Not applicable

The net diff against main touches only docs and the observatory's inputGroupOf (a read-only analysis function, never hashed). The mechanism and its regenerated goldens were added and then reverted inside this branch, so all 50 golden entries are the current main ones.

Testing

Commands run:

pnpm format:check
pnpm lint (simulation, contracts, config)
tsc --noEmit (simulation src and tests)
vitest run tests/observatory   69 of 69
Before the revert, on the mechanism: simulation unit suite 1292 of 1292, contracts and config suites, UI 65 of 65, web 55 of 55, the golden suite regenerated for hash schema 13.

Additional tests:

tests/probe/place.probe.ts and the arms in the ADR (control, full wiring, steering only, information only; seeds 1 to 8 and 9 to 16) are in a741b65's history; capacity probes use tests/probe/evolution.probe.ts with EVO_CONFIG.

Performance impact

  • No expected performance impact
  • Improves performance
  • May reduce performance
  • Benchmark included

Save compatibility

  • No save-format impact
  • Save format changed
  • Migration added
  • Existing saves tested

Screenshots or recordings

Not applicable (no UI change).

Review notes

  • Second negative-result record in a row (docs(predation): ADR 0023 step four, a mouth that breeds true was tried and made no difference #141 was the mouth). Both are deliberate: the code was removed so main carries only what measured out, and the numbers stay in the ADRs.
  • The equivalence harness detail worth knowing: a multi-cluster world differs between versions only through the shifted genetics stream, because founders of different species differ and so the blend of their children depends on the draws; a one-cluster world was identical.

Generated by Claude Code

…(ADR 0025, hash schema 13)

A new input group, recall (observation 48 to 52, weights 514 to 546): the world
position of the creature's last bite of real food as freshness, body-frame
bearing and proximity. It fades over learning.placeMemory.horizon, is forgotten
when the creature stands at the site without a bite, and a carcass never sets
it. The innate prior steers gently toward it with nothing else in view. The
observatory gets a recall input group (and an unmapped slot now throws instead
of counting as a threat) and two recall situations in the response battery.

Equivalence: with the sense off, founder brains equal to the prior, mutation and
learned inheritance frozen, eight scenarios give identical snapshots to the
previous commit through births (a multi-cluster layout differs only via the
shifted genetics stream; one cluster is identical). All 50 goldens regenerate.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cn5qp4xBHxLnTrwReNXcf8
…ast fed (ADR 0025, hash schema 13)"

This reverts commit a741b65.
…and made no difference

Capacity (hidden 8/16/32, memory 4/8) moves nothing on the observatory. A
remembered feeding site (built, proven equivalent with the sense off, then
measured against an identical-size control) does not help: population down 13
to 14 percent with the instinct wired, identical to the control without it, two
of sixteen worlds extinct against none. Reverted; the record keeps the numbers.

Kept because they stand on their own: the observatory's inputGroupOf now throws
for an unmapped observation slot instead of counting it as a threat, and
brain.md said 40 inputs where there are 48.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cn5qp4xBHxLnTrwReNXcf8
Copilot AI balanced review requested due to automatic review settings November 13, 2026 04:27

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

CodeByBryant and others added 2 commits November 13, 2026 06:41
… budget

long-run-chunking's two 40,000-tick tests ran under the global 180 s testTimeout.
On Ubuntu CI shard 2 the first took 156.5 s (87 percent of the budget) and the
second 78.4 s; on a runner about 40 percent slower it needed about 219 s and timed
out on PR #142, after the same code had passed. The claim they assert (the same
state hash) is unchanged; the budget is 420 s, 2.7 times the nominal slowest.
The ecology-stability tests already carry a 300 s budget.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cn5qp4xBHxLnTrwReNXcf8

Copy link
Copy Markdown
Owner Author

CI on d910a58 (after main was merged in) failed one job: Determinism (ubuntu-latest, shard 2), long-run-chunking.test.ts > running in two uneven chunks matches a single run of the same total, at scale, timed out at the 180 s global testTimeout. It is a timeout, not a hash mismatch, and nothing in this PR is on that test's path (docs plus the observatory's inputGroupOf).

Evidence: in the earlier green run on the same code the test took 156.5 s, 87 percent of the budget. In the failing run the runner was about 40 percent slower (the neighbouring ecology-stability files went from 243 s to 342 s and 229 s to 320 s), so it needed about 219 s. It had no explicit budget; the ecology-stability tests already have 300 s.

Fix pushed: an explicit 420 s budget (2.7x the nominal time of the slowest) on the two heavy chunking tests. What they assert is untouched. The file passes locally (3 of 3).


Generated by Claude Code

@CodeByBryant
CodeByBryant merged commit a798843 into main Nov 13, 2026
20 of 21 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants