38 KiB
Project status and next steps
Updated 2026-09-26. The repository contains an engineering foundation and a deterministic headless proof. It does not yet contain cellular gameplay or a persistent simulation server. The founding vision remains unchanged.
Completed implementation
- Vision, architectural pillars, risks, open-question register, architecture diagrams, biological contracts, engineering constitution and 47 ADRs. Proposed/deferred choices remain distinct from accepted proof decisions.
- M0–M4 task breakdowns with dependencies, acceptance evidence and risks, plus M5–M14 gates in ROADMAP.md.
- C++23 CMake/Ninja library and CLI with no graphics, networking, database or Kubernetes dependency. Versioned fixed ticks, stable entity IDs, component lifecycle and explicit per-entity seeded randomness.
- Bounded canonical snapshots and restore, plus replay verification. An independent reference fingerprint and Release-active tests cover lifecycle, malformed data, overflow, seeded properties and CLI save/load.
- Formatter, static analysis, sanitizer presets, codec fuzz entry point, benchmark, structured host logging and metrics output. Benchmark compiler metadata now resets numeric formatting after printing its hexadecimal state hash.
Delivery and verification
Git remote is https://git.aeternalisgames.com/dingo/AEvolution.git. The maintainer
added a working Gitea/Harbor workflow; run 1399 on 8ab48d3 successfully verified
and published. The foundation update extends that single workflow with Linux
analysis/sanitizers/fuzzing and runtime smoke verification. Images target
registry.aeternalisgames.com/aeternalis/aevolution.
Run 1402
passed all jobs for source commit 4c2cc04: Linux checks, separate TSan and image
smoke/publish. That verified image was previously pinned. The maintainer's
subsequent GitOps commit 20e609d selects the run 1399 image at
sha256:0c6c6e1ac28de2ce0648fd0992ca694963e303ef5c9414fe5ac668937136800d;
the M1.6 change preserves that independent deployment choice.
Windows checks also pass; full commands and intermediate CI failures are recorded
in the validation report. M0's engineering proof is verified; dev cluster
execution is the remaining delivery check, not a claim of persistent-world readiness.
The maintainer confirms Argo is configured and deployment additions will sync when a pod workload is ready. This workspace has no kubeconfig, so completed dev Job execution is not independently verified. This is not a blocker for M1. The repository supplies a restricted dev Job and bootstrap examples; image publication does not demonstrate a Kubernetes rollout or world recovery.
See VALIDATION.md for actual commands/results and
deploy/README.md for image promotion and application setup.
Local INITIAL_BRIEF.md and OpenAI_ChatLog.md are maintainer-provided untracked
material and are not automatically included in commits.
Next reviewable increments, in order
M1.5 progress: the opt-in row/EnTT4 experiment now checks canonical equality, restore continuation, retired identities and allocator balance, and records seven-sample sparse/dense/churn results. Windows Release, analysis and ASan checks passed, as did Linux Release, analysis and ASan/UBSan in Gitea run 1404. All 56 sample fingerprints match across platforms. ADR-0005 retains rows for the bounded proof; the general production ECS remains a later measured choice.
| Priority | Work | Completion evidence |
|---|---|---|
| 1 | M5.7 local publication and compaction | Verified snapshot/journal replacement, recovery and injected filesystem failures |
| Research | Improve aggregate fidelity and measure Linux/target-host costs | M4.6 found no qualifying width; see the aggregate budget report |
| 2 | Revisit the ADR-0031/0032 reproduction cost model | Reproduction is affordable only at the genome schema's floor; decide whether that is intended before M4 builds cohorts on it |
| Delivery gate | Add/update the dev workload when a deployable increment needs it; Argo is ready | Observe Job/pod completion separately from image publication; no World0 changes |
M1.6 now implements the sequential scheduler contract: exact-tick admission, bounded ordered lifecycle commands, receipts and atomic rollback on rejection. Windows Release, clang-tidy, ASan and formatting checks pass. Permutations across 32 seeds, restored continuation, duplicate/late/future inputs, limits and exhaustion are covered. The no-input golden and snapshot v1 are unchanged. Gitea run 1407 passed all Linux checks, ECS, TSan and image jobs; evidence is recorded in the validation report. Nonempty batches copy world state; this favors a clear transaction boundary for the proof, not a measured realtime performance claim. No thread pool or gameplay input protocol is introduced.
M1.7 implements read-only periodic neighborhood queries with stable IDs and inclusive integer-distance boundaries. A nine-image oracle checks 2,304 seeded queries across initial, scheduled and restored worlds. Boundary and full-capacity tests supplement the oracle. The benchmark records uniform/clustered worlds up to 100,000 entities. This is a linear correctness baseline, not a production spatial index. See the spatial contract and validation report for scope/evidence. Native Release/ASan/analysis/format checks passed, as did all jobs in Gitea run 1409. All 72 Linux benchmark result checksums and match counts equal Windows.
M2.1 adds an independent finite-inventory accounting library (ADR-0024). Typed picojoules and C/N/P/other femtograms transfer atomically across bounded local reservoirs. Opening totals are checked, repeated action IDs rejected, and failures leave both balances and action watermark unchanged. One million seeded transfers are checked against independently maintained balances and recomputed totals. No organisms, biochemical rates, external replenishment, world snapshot format or deployment changes are included. Conservation does not establish valid chemistry or reusable heat; process restrictions are M2.2 work. Native Release, clang-tidy, ASan and formatting passed; Gitea run 1411 passed all Linux checks, ECS, TSan and image publication. Validation details are recorded in the engineering validation report.
M2.2 implements a synthetic finite-source producer process owning its ledger. It captures energy, sends conversion losses to heat, and funds whole material recipes from reserves and finite nutrients. Fractional rates and efficiency carry exact remainders; tick/remainders and balances commit together. Tests cover hand-calculated outcomes, 32 seeded reference/replay runs, source and nutrient exhaustion, invalid tick/config rejection and u64 arithmetic limits. No organism, genome, database, process snapshot or deployment is implied by this fixture. Native Release, analysis, ASan and format checks passed, and Gitea run 1413 passed all Linux checks, ECS, TSan and image jobs.
M2.3 adds a shared feeding community with one producer and up to 16 consumer capabilities. One ledger owns all process resources, heat and waste. Ordered intake-capped feeding, producer progression and terminal depletion commit as one tick; a rejected request leaves the entire community unchanged. Tests cover predation, resource contention, terminal actors, batch order/replay and conservation. No spatial/perception validation, organism biology, persistent checkpoint or deployment is included in this fixture. Native Release, analysis, ASan and formatting passed; all four Gitea run 1415 jobs passed.
M2.4 adds terminal events with preserved corpse inventory, followed by a shared bounded decomposition phase. Energy moves to heat, elemental matter returns to nutrients, and next-tick producer growth can use those nutrients. Corpses retain their process ID and cannot resume activity. Invalid ticks roll back death and decomposition with all earlier phases. Default zero decay preserves earlier fixture behavior. This is not death-cause biology, microbial chemistry, corpse scavenging, persistent lineage history or a database-backed world. Native Release, analysis, ASan and formatting passed, and all four jobs in Gitea run 1417 passed.
M2.5 adds funded budding (ADR-0028). A bounded list of trusted reproduction intents runs after feeding and before decomposition, so a consumer can fund offspring from what it ingested and a process terminated in that tick cannot reproduce. Parent eligibility is fixed on entry to the phase, so a child born in a tick cannot itself be a parent until the next one. The intent names a withdrawal from the parent; a fixed construction fraction endows the child and the remainder becomes real heat and waste. Reserved slots are pre-allocated at creation, so a birth never grows the ledger and a full community rejects rather than funding a smaller child. Process identities are never reused and a parent reference plus origin tick outlive both organisms. A child duplicates its parent's configuration because no genome exists yet; this is duplication, not heredity. Default zero capacity leaves every M2.4 result and the particle goldens unchanged. Native Release, analysis, ASan and formatting passed; Linux CI for this increment has not run yet. This is not a reproductive strategy, fecundity rule, mutation model or funded founder admission.
M2.6 closes M2 with a deterministic ecosystem driver (ADR-0029). It owns one
community and one seeded stream, holds no inventory and emits only the intents
the earlier increments already accept, so no simulation contract changed. Two
limits shaped it: retained identities and capped slots mean one community is not
a population, and with no maintenance cost an uneaten consumer never dies. The
harness therefore measures many bounded episodes, with mortality from predation
and a configured lifespan delivered as a host terminal fact. The driver probes a
discarded copy each tick so it never submits a batch the community rejects; a
rejected tick is reported as a driver defect rather than absorbed as ecology.
evo-sim --ecosystem reports seeded episode ensembles: 3,000 episodes advanced
1,093,007 ticks, all reaching extinction, with deaths reconciling exactly against
founders plus births and conserved totals identical at every tick. That
demonstrates arithmetic and turnover, not ecological plausibility; strategy
constants are chosen inputs and must not become gameplay by having been measured.
Native Release, analysis, ASan and formatting passed; Linux CI has not run yet.
M3.1 opens M3 with the immutable genome record (ADR-0030) in a new evo::biology
library that nothing else links yet. A genome is validated once at construction
and exposes no mutator, so an invalid one cannot exist at runtime; decode
applies identical checks, so a decoded record is never less validated than a
constructed one. Schema 1 fixes the seven v0.1 capability kinds with units,
inclusive ranges, required flags and mandated developmental edges as schema facts
rather than record data. Identity is a nonzero world-scoped genome ID distinct
from every runtime identity, a nonzero origin event and zero or one parent, with
room reserved for recombination. The encoding is canonical, bounded and exact:
strict ascending order, a stored digest that is recomputed on decode, and a
trailing checksum, with the byte layout in schemas/genome-v1.md. No genome
affects any simulation; the M2 fixtures still duplicate founder strategies.
Development, phenotype, mutation and heredity remain M3.2 to M3.5. Native
Release, analysis, ASan and formatting passed; Linux CI has not run yet.
M3.2 derives a phenotype from that record (ADR-0031). derive_phenotype and
plan_development are pure integer functions of genome, growth and environment,
so a phenotype is recomputed from retained inputs rather than cached. Growth and
damage are runtime state that never reach the genome: damage changes the
phenotype and leaves the content digest identical, which is the risk this
increment was scoped against. The genome's development version selects a rule
set and an unknown version fails closed. Version 1 declares every constant:
the body is a cube of uniform declared density, membrane mass follows surface and
thickness, maintenance follows mass plus a membrane upkeep term, and uptake scales
with surface, so size trades uptake against maintenance. A capability the genome
does not carry yields zero. A plan advances growth by at most the rule step,
reduced to the largest step the available material funds exactly, and reserves
nothing: no inventory moves until M3.3. Omitted physics is a documented model
limitation, not a modifier, and the constants must not become tuning by having
been written down. Native Release, analysis, ASan and formatting passed 18/18;
Linux CI has not run yet.
M3.3 gives capability a cost (ADR-0032). metabolic_budget returns body
composition, maintenance, movement, construction, the per-tick intake limit and
the reproduction investment, in accounting types; metabolism is the single
translation layer and the genome record stays free of those types. Each budget
comes from a different phenotype quantity, so one parameter moves some and not
others: a thicker membrane costs more matter to build, more energy to hold and
more to haul, while leaving uptake, speed and assimilation identical. That
non-uniformity is asserted directly, because a capability that became a flat
multiplier is the failure this increment was scoped against. Composition is
declared per compartment, and splits give the remainder to the last channel so
parts sum back to the whole and a derived body can be spent without a rounding
leak. processes::consumer_for is the first path from a genome into the M2
fixtures, with opening inventory left to the caller so no matter is invented; a
derived consumer is shown driving a real feeding tick with the ledger conserving.
No existing fixture reads genomes yet and no organism spends any budget. Native
Release, analysis, ASan and formatting passed 19/19; Linux CI has not run yet.
M3.4 proposes successors (ADR-0033). A static catalog of four effects over the seven capabilities is filtered by the M3.1 validation itself, so eligibility is not a second rulebook: shedding a capability another declares as a prerequisite fails without a rule written for it. Candidates are drawn without replacement by rejection sampling on an explicit stream, so every eligible operator is exactly as likely as any other, and the result is returned in canonical order with no score anywhere in the record. The draw reads only the parent digest and a recorded context, so the same context always returns the same candidates and a retry cannot reroll, while a new event index earns a new draw. Selection regenerates the offer and refuses a changed context outright rather than answering it with a different mutation. A candidate reports what ADR-0032 would charge its successor and reserves nothing. The uniformity test was confirmed to fail against a draw that took the first eligible operators. Candidate count, mutation cost as a game rule, inherited versus acquired expression, NPC mutation and recombination stay open; no organism mutates. Native Release, analysis, ASan and formatting passed 20/20; Linux CI has not run yet.
M3.5 keeps what M3.4 produces (ADR-0034). LineageArchive has no removal
operation at all, so retention is a property of the interface rather than a rule
someone has to remember, which is the only form of "keep extinct genomes
indefinitely" a later change cannot quietly undo. Admission requires the parent to
be present, so ancestry never dangles and the parent graph is acyclic by
construction; a duplicate identity is refused rather than overwritten. Organism
and genome identity stay separate: a death is terminal for the organism and
touches nothing about the genome, and ancestry, descendants and divergence all
resolve after every carrier has died. A lineage with no living carrier is reported
absent, never extinct, because terminal extinction needs a reconciled watermark
covering suspended members, eggs and in-flight transfers that this type cannot
see; absence is also reversible. Capacity is bounded and admission fails loudly,
since refusing to record is not deleting. NPC mutation stays undecided: the
archive records an origin event and never an actor. Nothing is serialized, so
durable retention and migration remain M7 and M8. Native Release, analysis, ASan
and formatting passed 21/21; Linux CI has not run yet.
M3.6 closes M3 by making the M3 pieces meet (ADR-0035). strategy_for derives a
StrategyConfig from ADR-0032's budget, so a strategy is what the genome already
implies rather than a second description of the same organism, and a parent
reproduces once it holds what it would invest. Founders are variants of one
ancestor diverged by ADR-0033's draw before any episode starts and admitted to
ADR-0034's archive with their parent links; every birth and death is recorded as
it happens, so the diversity a run reports has a provenance. Nothing mutates
during an episode, because autonomous divergence would settle NPC mutation, which
stays open. evo-sim --ensemble runs the grid and reports each cell's seeds in
two halves that are never merged. ADR-0029's driver is untouched and the M2.6
fingerprint still holds.
Two findings are recorded rather than smoothed over. Resource level never reaches
strategy derivation: founders are derived fully grown, so their construction
transition is empty and ADR-0031's per-channel available cannot move intake,
assimilation or investment; the resource level acts only as the nutrient pool
bounding producer growth. And the resource axis is not monotonic, so no test
asserts a direction. A sweep peaks near 8000 femtograms and falls away on both
sides, because ADR-0026 lets only one consumer bite a target per tick: the
community's income is capped at one producer bite regardless of the pool, scarcity
drives predation that concentrates mass into individuals that reach the threshold,
and abundance spreads the same income too thin to fund a birth. Separately, the
fixture's reproduction investment is pinned to the genome schema's floor because
intake scales with surface while investment scales with mass; at any larger
fraction nothing can earn a birth within a lifespan and both axes measure nothing.
That is a finding about the ADR-0031/0032 cost model, not a fixture detail.
Review of 89ad95b found three defects, all now fixed with regressions confirmed
to fail without their fix. The harness recorded a death for every organism still
alive when an episode stopped, so a one-tick run reported no extinctions while
writing terminal records for founders that had not died; ending observation is not
death, survivors now stay living in the archive and their watched time is summed
apart from completed lifetimes. A draw that produced a genome without the
reproduction capability derived a zero investment, which ADR-0029's driver refused
outright, so seeds 10, 12 and 15 failed the whole run; the driver now runs such a
strategy and proposes no birth for it, because filtering those variants out would
have made the measured set the fertile subset chosen by a guard. And a half's
thermal shortfall was assigned rather than accumulated, so a field documented as a
total reported one episode's reading. The M2.6 fingerprint 2cd4a15f9fbe9949 is
unchanged, since only configurations the driver previously rejected behave
differently; the ensemble CLI pin was repinned because the reported facts changed.
Native Release, analysis, ASan and formatting passed 33/33; Linux CI has not run
yet.
Before biology, resolve the relevant heredity and energetic contracts rather than treating proof particles as organisms. Exact game tick rate, NPC mutation, logout treatment, species/extinction rules and multicellular transition remain open decisions. Renderer/networking remain later milestones. Deployment availability does not complete M7/M8: fenced ownership, durable outcomes, backup restore and world-version migration require their own implementation/fault tests.
M4.1 opens M4 (ADR-0036) by defining the tier 2 record and who may act on it, and
nothing else: there is no collapse transaction, no population advance, no
materialization and no habitat geometry. A Cohort is validated on construction
and never mutated afterwards. Age classes each carry their own count and their own
reserve, so the record holds a distribution rather than the mean the milestone's
named risk warns about; within a class members are indistinguishable, which is the
record's declared resolution and a stated loss rather than a hidden one. No
accessor returns a mean. Totals are summed from the classes instead of stored
beside them, following the lineage archive's rule that a second tally is a thing
that can disagree. Lineage is deliberately not a field: ancestry is resolved from
the genome through the archive, so a cohort cannot contradict it or invent
parentage from membership. Residuals are integer numerators over a declared scale,
strictly below one member, and are never added to a count.
Ownership is enforced rather than described. CohortRegistry admits at most one
record per habitat and genome, so ECOLOGY_MODEL's "no inventory is spendable in
two tiers" is unrepresentable rather than forbidden, and replacement is refused
under a different worker, an older epoch, or a changed habitat or genome. Epochs
make a stale writer fail locally and say nothing about fencing between workers,
which stays M7 and M10. The approximation bound is carried and never computed:
M4.5 measures resolution equivalence, and carrying a bound is not meeting one.
Nothing is serialized, so schemas/population-v1.md pins field meaning and
invariants rather than a byte layout. All nine construction and registry guards
were confirmed load-bearing by removing each one and observing the suite fail.
Native Release, analysis, ASan and formatting passed 34/34; Linux CI has not run
yet.
Review of 1c8b292 found two defects, both fixed with regressions confirmed to
fail without their fix. Replacement checked the ownership epoch but nothing that
ordered two writes inside one: a record installed at tick 1,000 holding 50
picojoules could be replaced by its own earlier tick-900 record holding 100,
restoring inventory the installed one had already spent. Cohorts now carry a
revision that strictly increases within an epoch, and the interval boundary is
checked independently, so neither advancing a revision nor re-acquiring ownership
rewinds simulated time. Separately, the implicit move emptied a cohort's age
classes while keeping its identities, and the registry then accepted through a
move exactly what create rejects by name; the type is now copy-only, as ADR-0030
already made the genome for the same reason. That precedent existed and should
have been applied when the type was written.
M4.2 moves individuals into that record (ADR-0037), and does only that: there is
no population advance, no materialization back into individuals, no transfer
between habitats and no persistence. Collapse is two phases over one sample.
plan_collapse reads the members once and returns an immutable plan that already
carries a validated cohort; commit_collapse applies exactly that plan and
derives nothing, so ECOLOGY_MODEL's rule that recovery replays a committed sample
rather than sampling again is a property of the types.
A controlled, suspended, dead or unsettled member refuses the whole plan rather than being skipped. That is the answer to adversarial selection: filtering would let a caller obtain a plan by omitting whoever was inconvenient, and the result would be indistinguishable from an honest conversion. Ages are carried into declared classes and a member fitting none refuses the plan, so a conversion cannot quietly move anyone to a different class or discard one for having nowhere to go. Reserves are summed per class, never redistributed.
Conservation runs through the ledger rather than beside it: one atomic transaction of one leg per member, each moving that member's entire reserve, so the conserved total is unchanged by construction. Dual representation is closed from both ends. A member must be emptied exactly, and a reservoir holding anything other than what was sampled makes the plan stale rather than partly applicable. The destination is proved available before anything moves, so no rejection can strand inventory in a reservoir nothing owns, and every refusal leaves the ledger and the registry as they were. Committing one plan twice is refused without moving inventory again.
Two limits are stated rather than implied to be covered. The function cannot tell whether the members it was given are all the members there were, so completeness is a caller obligation. And emptying a reservoir is not withdrawing the right to use it: the ledger has no removal, so nothing prevents a later transfer crediting that reservoir again. Durable conversion sequences, escrow and replay across a process boundary are M7 and M8. All ten construction and commit guards were confirmed load-bearing by removing each and observing a failing run. Native Release, analysis, ASan and formatting passed 35/35; Linux CI has not run yet.
Review of 1a59b63 found one defect, now fixed with a regression confirmed to
fail without it. Commit checked that the destination reservoir existed and that
the registry slot was free, but not that the reservoir was empty, so a conversion
into an occupied reservoir succeeded and mixed inventory it never sampled into the
cohort's backing store. The reported case reproduces exactly: the ledger holds 67
picojoules against a cohort recording 60, so the cohort's reserve understated what
backed it and the surplus belonged to nobody. A destination holding anything is
now refused, a cohort's reserve is the whole of its reservoir rather than a part
of it, and the success test asserts that equality directly.
M4.3 advances a cohort over one declared interval (ADR-0038) and deliberately does not model one. ECOLOGY_MODEL keeps population maths an experiment and asks for several candidates to be compared, so choosing one here would turn a recommendation into a locked rule. A model proposes fluxes; this checks that the declared result is exactly what they imply, for counts and for every inventory channel.
Nothing is clamped. A loss larger than what is there refuses the interval and names whether it was the population or the biomass, so a caller can subdivide and retry, and subtraction is checked against what is present rather than performed and repaired. That is the answer to the milestone's named risk: a clamp is how an interval a model could not integrate becomes a plausible looking result, and how a false extinction arrives. Counts carry whole members and owed fractions together, so a population below one member is a number rather than a rounding decision. Biomass has no birth term at all, so a birth cannot bring inventory in and newborns are necessarily funded from what the cohort already holds; immigration draws from a reservoir that already held it, so a late transfer is not new production either. Conservation runs through the ledger as one atomic transaction, and an interval in which nothing moves submits none.
Two amendments to M4.1's record were forced rather than chosen. A cohort now records the reservoir backing it, because an advance cannot otherwise know where the inventory is and passing it alongside would let a caller advance one cohort against another's store, which comparing balances cannot detect between equal reserves; the registry refuses two cohorts sharing a reservoir. And a cohort may be vacant, through a separate factory so vacancy stays deliberate, because a population declining to exactly zero would otherwise be unrepresentable and the caller would have to clamp. A vacant cohort is not an extinction: the record, its genome, its habitat and its provenance all remain and it stays installed.
Limits are stated rather than implied covered. Accounting is not fidelity: closure
is checked, stability is not, so a long interval whose arithmetic closes and whose
ecology is nonsense is accepted. The declared age structure is not verified to be
what ageing would produce, because that needs the within-class distribution the
record deliberately discards. Emigration escrows into a reservoir nothing settles
with, since TransferPopulation does not exist. All fourteen guards, including
both clamp mutations, were confirmed load-bearing by removing each and observing a
failing run. Native Release, analysis, ASan and formatting passed 36/36; Linux CI
has not run yet.
M4.4 turns whole members of a cohort back into identified individuals (ADR-0039), and does nothing else: no transfer between habitats, no persistence and no resolution-equivalence claim, which is M4.5. Materialization is instantaneous, at the cohort's current tick, because advancing time is M4.3's interval.
The named risk is that re-materialization becomes free healing, and it has two separable sources. The arithmetic one is closed outright: a drawn member receives its class share by integer division rounded down, because rounding up would create biomass, and a class drawn empty hands over everything it held because a class with no members may keep no reserve. A residual is never drawable, so an expected fractional density cannot become an individual. Age is taken as the oldest the class admits rather than the youngest or a seeded choice within it: the exact age is unrecoverable, so the question is which direction to be wrong in, and every other answer lets an adversary who picks when to convert roll back towards youth. A round trip can therefore only age someone, and oscillation is strictly lossy.
The second source is the declared resolution itself and is stated rather than mitigated. Members inside one class are indistinguishable by construction, so collapsing a starving member with a well-provisioned one and materializing both returns two average members; the starving one has been healed at the other's expense. The tests demonstrate both halves of this rather than only the flattering one: a round trip is exact when each class holds one member, and a coarse class conserves the total while averaging the individuals. The mitigation is the caller's choice of classes when it collapses, not anything this code can recover.
Materialization produces exactly the record a collapse consumes, so the two compose and are tested against each other instead of being described as inverses. Eleven guards were confirmed load-bearing by removing each and observing a failing run, including the two that are the risk itself: rounding shares up, and drawing ages from the bottom of a class. Native Release, analysis, ASan and formatting passed 37/37; Linux CI has not run yet.
Review of 380d609 found two defects, both fixed with regressions confirmed to
fail without them. Materialization treated an empty reservoir as an available one,
so it credited a store that another installed cohort owned while leaving that
cohort's record unchanged: the same inventory in two places, reached from the
opposite direction to the M4.2 defect. Destinations are now checked against the
registry's ownership as well as their balance, on both the collapse and the
materialize side. The collapse side had the worse form of it: the shared-reservoir
rule would have refused the install only after the transfer had run, stranding the
members' biomass in a reservoir nothing accounted for, so whichever registry step
will follow a transfer is now proved to succeed before it.
The second was a functional gap the oscillation test had been hiding by discarding its registry each cycle. A full draw leaves a vacant record, which ADR-0036 retains because absence is not extinction, and collapse then refused that habitat and genome as occupied, making a full materialization a one-way door out of a habitat. Collapse now repopulates a vacant record in place, continuing its ownership and revision series rather than restarting them, and refuses a record that still holds a population. The test keeps one registry throughout and asserts the revision continues.
M4.5 closes M4 by measuring what aggregation costs (ADR-0040). It adds no library code: one rule is written once and applied at both resolutions, and a one-member-per-class control proves the two runs are the same rule before any divergence is attributed to resolution. Without that control a measured gap could equally be a transcription mistake, and nothing afterwards would separate them; a deliberate one-unit drift in the aggregate rule is caught by it.
The result is reported rather than smoothed into a tolerance. At class widths of ten and fifty the aggregate population survives ten thousand ticks and tracks the individual run within 375 parts per thousand on population and 43 on energy. At widths of twenty-five, one hundred and two hundred it dies out entirely while the individual run sustains seventy-five members: a false extinction caused by resolution alone, because a whole class starves the moment its average cannot pay upkeep where individuals would have starved a few at a time and the survivors would have kept the shared per-capita share up.
The collapse is not monotone in class width. Twenty-five destroys the population while fifty sustains it, so a resolution cannot be chosen by assuming degradation worsens smoothly as classes widen, and a width that works is no evidence that every narrower width works. The tests assert that non-monotonicity, because it is surprising and because the warning written around it would otherwise be able to become wrong silently. They also assert the runs disagree, so equivalence can never be read as identity, and the word is not used for what was measured.
Conservation was exact at every tick on both sides of every run, including the three whose populations died out. Four of five runs conserve perfectly and three are ecologically worthless, which is exactly ECOLOGY_MODEL's point that conservation is insufficient evidence.
The comparison also found a real defect in M4.3, which is what a validation
milestone is for. plan_advance checked affordability with arrivals added before
losses were taken, but emitted the ledger legs with the outflows first, so an
interval that was affordable as a whole underflowed leg by leg and was refused at
commit. A fifty-tick interval shedding 10,480 against 6,950 held while drawing
13,080 in reproduced it. Arrivals are now the first leg, matching the order the
arithmetic was checked in, with a regression covering it. Native Release, analysis,
ASan and formatting passed 38/38; Linux CI has not run yet.
M4.6 - aggregate budget review (2026-09-26)
Completed test-only shared model, CPU/retained-memory benchmark, and 15-scenario
scarcity sensitivity report. The partial tests/fixtures/resolution_model.hpp
was integrated with existing M4.5 controls, including contested parent allocation.
No production library, simulation format, deployment or gameplay rule changed.
See the report and ADR-0041. None of the five
widths passes the proposed fidelity gate across supplies 240/300/360; automatic tier
selection remains disabled. The next implementation milestone is M5's sparse voxel
prototype; aggregate-model improvement and Linux/target-host measurement remain
prerequisites for a production tier policy. See VALIDATION.md for checks actually run.
M5-M9 planning expansion - 2026-09-26
ROADMAP.md now breaks M5-M9 into 42 reviewable tasks, each with module boundaries, dependencies, completion criteria, tests and risks. These are planned gates, not implementation or deployment claims. M8 dev/stage recovery is separate from World0 launch; M9 client integration can use dev/stage without premature World0 promotion. M4.6's failed fidelity gate remains explicit. That expansion identified 5.1 as the next task; it is now implemented as recorded below. No library, service, cluster or dependency changed in this update.
M5.1 - bounded local voxel contract (2026-09-26)
Implemented dependency-free evo::world coordinate and cell-sample utilities,
ADR-0042, and
LOCAL_VOXEL_CONTRACT.md. Signed floor mapping handles
negative chunk seams; all public operations reject invalid bounds before arithmetic.
The experiment uses a 1m cell pitch and explicit edge candidates 8/16/32/64; it
selects no storage layout, renderer, spherical topology or production world extent.
Logical fill describes volume fraction, not surface shape or consumable inventory.
Tests independently enumerate chunk intervals, exhaust 299,520 local ordinals, check metric/domain corners and hostile signed/unsigned extrema, and validate fill/material consistency. Existing simulation goldens/formats are unchanged. No world data or service is deployed. Next: M5.2's measured layout/edge comparison; actual chunk allocation and generator work follow its decision. See VALIDATION.md for checks.
M5.2 compact layout evidence
Compared three representations, four edges and five fixtures with seeded edit/oracle checks. Evaluation records memory, costs, encoding estimates and limits. ADR-0043 selects 16-cube uniform/palette/dense geometry. No production container, generator, codec or deployment was added. Next: M5.3.
M5.3 deterministic abiotic base
Added pure seed/address/version generation, a bounded inert material table and 16-cube logical output. Contract specifies rejection, compatibility and storage boundaries. Twelve independent reference goldens and signed seam/query-order tests freeze v1. No gameplay, persistence or deployment was added. Next: M5.4 sparse active chunk ownership.
M5.4 bounded active chunk ownership
Added fixed-slot, budgeted ChunkStore and immutable uniform/palette/dense CompactCells. Revision tokens reject stale and foreign reads. Clean chunks regenerate; dirty pins refuse eviction. Ownership contract documents memory/lifetime limits and the absence of cell edits or durability. Next: M5.5 accounted terrain edits.
M5.5 transactional terrain excavation
Added up to 64 cuts across four loaded chunks, atomically transferring occupied volume into finite per-material debris. Revision/digest/full-payload retry checks, 64-command retention and peak allocation backpressure are explicit. The edit contract records excavation-only scope and persistence obligations. No mass conversion or durable acknowledgement is claimed. Next: M5.6.
M5.6 snapshots and edit journals
Added bounded canonical world snapshot v1 and edit journal v1, world/frame and base version checks, full geometry/debris/retry restore, semantic validation and detached all-or-nothing replay. Codec contract and schemas define memory/trust limits. Dirty flags remain set; no file publication or durable ACK is implemented. Next: M5.7 local publication, compaction and recovery.