Propose mutations without proposing a best build

M3.4 adds the candidate-generation experiment GENOME_MODEL describes: a
static catalog of four effects over the seven capabilities, drawn from a
recorded context with an explicit stream.

The named risk is a hidden best build or offerings biased toward what
would help, so the defence is structural rather than asserted. Nothing is
ranked: candidates are drawn without replacement by rejection sampling,
making every eligible operator exactly as likely as any other, and the
result is returned sorted by operator ID. No score exists in the record,
position carries no meaning, and nothing about an organism's situation
makes a helpful mutation more likely to be offered.

Eligibility is not a second rulebook. An operator is eligible when the
genome it would produce passes the M3.1 validation unchanged, so
prerequisites, required capabilities, ranges and the acyclic development
graph are enforced once, where they were defined. Shedding a capability
another declares as a prerequisite fails that check with no rule written
here to catch it. Range boundaries are refused rather than clamped, since
a clamp would make two operators agree and offer a no-op as a choice.

The draw reads the parent digest and the context and nothing else, so the
same context always returns the same candidates and a retry cannot
reroll; a new event index is what earns a new draw. select_mutation
regenerates the offer and refuses unless it still matches, so a changed
context is rejected rather than answered with a different mutation.

Cost is not invented: a candidate reports the successor's own maintenance
and mass under ADR-0032, and reserves nothing.

The eligible set is pinned by hand from the schema rather than observed,
and the uniformity check is confirmed to fail against a draw that takes
the first eligible operators. The rejection sampling that removes modulo
bias is correct by construction, not test-verified: at this scale the
bias is one part in 10^18. That is recorded rather than claimed.

No organism mutates and no fixture calls any of this. Candidate count,
mutation cost as a game rule, inherited versus acquired expression, NPC
mutation and recombination all remain open, as GENOME_MODEL requires.

Native Release, clang-tidy analysis, ASan and format checks pass 20/20.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
dingo
2026-09-23 15:27:21 -04:00
co-authored by Claude Opus 5
parent fd6faeb7f8
commit 5497f4d74c
11 changed files with 788 additions and 6 deletions
+7 -1
View File
@@ -83,7 +83,13 @@ composition, maintenance, movement, construction, intake and reproduction
investment, each derived from a different phenotype quantity so one parameter
moves some budgets and not others. `consumer_for` is the first path from a genome
into the M2 fixtures, and mass splits conserve exactly so a derived body can be
spent through the ledger. No existing fixture reads genomes yet; that is M3.6.
spent through the ledger. No existing fixture reads genomes yet; that is M3.6. M3.4 now proposes successors
(ADR-0033): a bounded operator catalog whose eligibility is the M3.1 validation
itself, drawn uniformly without replacement from a recorded context and returned
in canonical order with nothing ranked. Selection regenerates the offer and
refuses a changed context rather than substituting. Candidate count, mutation
cost as a game rule, inherited versus acquired expression, NPC mutation and
recombination stay open, and no organism mutates yet.
## M3 — Genome v0.1
+20 -3
View File
@@ -7,7 +7,7 @@ persistent simulation server. The founding vision remains unchanged.
## Completed implementation
- Vision, architectural pillars, risks, open-question register, architecture
diagrams, biological contracts, engineering constitution and 32 ADRs.
diagrams, biological contracts, engineering constitution and 33 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](ROADMAP.md).
@@ -61,8 +61,8 @@ for the bounded proof; the general production ECS remains a later measured choic
| Priority | Work | Completion evidence |
|---|---|---|
| 1 | M3.4 seeded mutation candidates | Bounded reproducible candidates conditioned by recorded context, with prerequisite, cost and fixed-seed tests |
| 2 | M3.5 heredity and divergence history | Reproduction persists parent/child links and genome identity survives death |
| 1 | M3.5 heredity and divergence history | Reproduction persists parent/child links; genome identity survives death; extinct genomes are never deleted |
| 2 | M3.6 genome ecology ensembles | Strategies compared under resource and temperature gradients, with conservation and a diversity report |
| 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](architecture/SCHEDULER_CONTRACT.md):
@@ -208,6 +208,23 @@ 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.
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
+111
View File
@@ -0,0 +1,111 @@
# ADR-0033: Seeded mutation candidates over a bounded operator catalog
## Status
Adopted for M3.4 as the candidate-generation experiment GENOME_MODEL describes.
It settles none of the rules that document leaves open: candidate count, context
window, mutation costs as a game rule, inherited versus acquired expression, NPC
mutation and recombination all remain undecided. Heredity history is M3.5.
## Context
M3.3 gave capability a cost, but every genome is still authored by hand. Before
lineages can diverge, something has to propose a successor, and GENOME_MODEL is
specific about how: a constrained catalog of operators over capability records,
filtered by prerequisites, sorted by stable operator ID, sampled from a recorded
genome and context with an explicit stream, with no wall clock, no `std::hash`
and no implementation-defined distribution.
The named risk is a hidden deterministic best build, or offerings biased toward
what would help. Either would turn mutation into a progression tree wearing a
biological costume.
## Decision
Add `evo::biology::mutation_candidates` and `select_mutation` over a static
catalog. Each operator has a stable ID and one of four effects on one capability:
adjust its parameter up or down, acquire a capability the genome lacks, or shed
one it carries. The adjustment step is one rule rather than a constant per
capability: a hundredth of that capability's declared range, at least one.
Eligibility is not a second rulebook. An operator is eligible when its own
precondition holds and the genome it would produce passes the M3.1 validation
unchanged. Prerequisites, required capabilities, ranges and the acyclic
development graph are therefore enforced exactly once, where they were defined.
Shedding a capability another one declares as a prerequisite fails that check by
itself, with no rule written here to catch it.
Nothing is ranked. Candidates are drawn without replacement from the eligible set
by rejection sampling on a SplitMix64 stream, so each eligible operator is exactly
equally likely, and the result is returned sorted by operator ID. Position carries
no meaning, there is no score anywhere in the record, and nothing about an
organism's situation makes a helpful mutation more likely to be offered. Whether
offerings should ever be environment-conditioned is a game-design question
GENOME_MODEL leaves open, and this implementation does not open it.
The draw is seeded from the recorded context alone: the parent's content digest,
an explicit seed and stream, the eligible event index, the organism's growth and
environment, and the requested count. Two calls with the same context return
identical candidates, so a retry cannot reroll; a new event index is what earns a
new draw. Because the functions read nothing but the genome and that context,
events elsewhere in a world cannot move a candidate set.
A candidate reports what the successor would be and what it would cost, and
reserves nothing. Cost is not invented here: it is the successor's own maintenance
and mass under ADR-0032, derived from the same rules as the parent's, so a
candidate carries a consequence rather than a price this ADR made up.
`select_mutation` takes the candidates that were offered, regenerates them from
the context, and refuses unless the whole list still matches. A changed context is
rejected outright rather than silently answered with a different mutation. Only
then is the immutable successor built, with a caller-allocated identity, the
parent recorded, and M3.1 validating it once more.
## Alternatives Considered
Scoring or ordering candidates by any derived quantity would be the best-build
failure directly, however it was labelled. Conditioning the draw on what an
organism needs would make mutation anticipate its environment, which GENOME_MODEL
explicitly refuses to claim. Re-implementing prerequisite checks in the catalog
would let the two rulebooks drift. Clamping an adjustment at a range boundary
would make two operators produce one result and offer no-ops as real choices.
Plain modulo on the stream would bias the draw very slightly toward low operator
IDs, which is exactly the accusation this increment must be able to refute.
Letting selection substitute a still-valid candidate when the context moved would
mean an organism mutating into something nobody was shown.
## Benefits
Divergence becomes possible without a progression tree. Every offer is legitimate
by the same validation that defines legitimacy, uniformity is exact rather than
asymptotic, retries are stable, and a changed context fails loudly. The cost
shown is the cost the rest of the system would charge.
## Costs
Four effects over seven capabilities is a small catalog, and the shared
range-fraction step means a capability's granularity follows its range rather than
any biology. Eligibility runs validation once per operator per call, which is
cheap here but is a full genome construction each time. Regenerating the whole
list to select is deliberate duplicated work.
## Risks
Do not read a candidate's reported maintenance or mass as a recommendation, a
score or a price; it is what ADR-0032 would charge either way. Uniform offering is
not a claim about real mutation rates, and the catalog is not a claim about which
mutations are possible in biology. Nothing here decides whether NPCs mutate,
whether acquired state is ever inherited, what a mutation should cost a player, or
how often events occur. No organism mutates yet, and no fixture calls any of this.
## Consequences
Test that the same context always yields the same candidates and a new event index
yields a new draw, that offers are uniform across the eligible set over many seeds
and cover it completely, that the returned order is always canonical, that every
offered candidate produces a genome M3.1 accepts, that ineligible operators are
never offered, that shedding a declared prerequisite is refused, that boundary
adjustments are refused rather than clamped, that selection rejects an unknown
candidate and any changed context, and that a successor records its parent and
survives encoding. No genome schema, ledger, fixture, renderer or deployment
changes.
+1
View File
@@ -25,6 +25,7 @@ Each record has the eight required sections.
0030 adopts the immutable genome v1 record, validation and canonical encoding.
0031 adopts deterministic development, derived phenotype and bounded plans.
0032 derives non-uniform metabolic budgets and the first genome/ledger path.
0033 adopts the bounded mutation catalog and uniform seeded candidate draws.
Every dependency adoption records exact version/commit and integrity hash,
transitive license obligations, maintenance/security review, owner and removal
+43
View File
@@ -188,3 +188,46 @@ budgets differing by orders of magnitude make every tradeoff invisible. That is
legibility choice for tests, not tuning, and must be re-derived before balancing.
A budget is a derived quantity, not a rule: nothing here decides starvation,
death or repair, and no organism spends any of it yet.
## Implemented M3.4 mutation candidates
[ADR-0033](../adr/0033-seeded-mutation-candidates.md) implements the
candidate-generation experiment this document describes, and only that. Four
effects over the seven capabilities give a static catalog with stable IDs:
adjust a parameter up or down by a hundredth of its declared range, acquire a
capability the genome lacks, or shed one it carries.
Eligibility is not a second rulebook. An operator is eligible when the genome it
would produce passes the M3.1 validation unchanged, so prerequisites, required
capabilities, ranges and the acyclic graph are enforced once, where they were
defined. Shedding a capability another declares as a prerequisite fails that check
without a rule written for it.
Nothing is ranked. Candidates are drawn without replacement by rejection sampling
on an explicit SplitMix64 stream, so every eligible operator is exactly as likely
as any other, and the result is returned sorted by operator ID. No score exists in
the record, position carries no meaning, and nothing about an organism's situation
makes a helpful mutation more likely to be offered.
The draw reads the parent's content digest, a seed, a stream, an eligible event
index, growth, environment and the requested count, and nothing else. The same
context always returns the same candidates, so a retry cannot reroll; a new event
index is what earns a new draw. `select_mutation` regenerates the offer and
refuses unless it still matches exactly, so a changed context is rejected rather
than answered with a different mutation. A candidate reports the successor's own
maintenance and mass under ADR-0032 and reserves nothing.
| Contract | Status after M3.4 |
|---|---|
| ValidateGenome | Implemented for schema 1, and reused as the eligibility filter |
| PlanDevelopment | Implemented: a proposal, narrowed by the composition layer |
| ApplyDevelopment | Half: the transfer still does not exist |
| DerivePhenotype | Implemented for rule version 1 |
| GenerateMutationCandidates | Implemented: uniform, canonical, context-recorded |
| SelectMutation | Implemented: revalidates and refuses a changed context |
Uniform offering is not a claim about real mutation rates, and the catalog is not
a claim about which mutations are possible in biology. Candidate count, context
window, mutation cost as a game rule, inherited versus acquired expression, NPC
mutation and recombination all remain open, exactly as stated above. No organism
mutates, no fixture calls any of this, and heredity history is M3.5.
+43
View File
@@ -583,3 +583,46 @@ chosen for legible magnitudes, not balance. Golden `d61ad5bf0c928c76`, ecosystem
fingerprint `2cd4a15f9fbe9949` and the frozen genome fixture are unchanged. No
deployment manifest or image pin changed. Linux validation is pending its first
pushed run; no Gitea result is claimed for any increment since `b01a824`.
## M3.4 seeded mutation candidates (2026-09-23)
Native Release, analysis and MSVC ASan each passed 20/20 CTests; formatting
passed. Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan>
-Action check` and debug `-Action format`. `mutation-contracts` is the one new
CTest case, in the existing `evo::biology` library. No existing suite changed.
The eligible set for the founder fixture is pinned by hand rather than observed:
nine operators, derived from the schema by the rules that make them eligible.
Energy conversion and membrane can adjust either way, movement, ingestion,
reproduction and thermal tolerance can be acquired, target size can adjust up but
not down because its 9,999,999 step would fall below the minimum from 8,000,000,
and no required capability can be shed. Over 2,000 seeds the offers are exactly
that set and nothing else.
Uniformity is checked to within a fifth of each operator's exact share, about six
standard deviations for 8,000 offers across nine operators. That test was
confirmed to fail when the draw was replaced by taking the first eligible
operators, which is the shape a biased offering would take. The rejection
sampling that removes modulo bias is correct by construction rather than
test-verified: over nine bins with 64-bit draws the bias it removes is about one
part in 10^18 and no test at this scale could observe it. That is recorded rather
than claimed as covered.
Stability is checked both ways: the same context always returns the same
candidates, while a new event index, seed, stream or growth state returns a
different draw. Selection is checked to refuse an operator that was never offered,
a moved event index, a changed growth state, a tampered offer and a different
parent genome, each with its own error. Boundary operators are checked to be
absent rather than clamped at both ends of a range, and a capability another one
declares as a prerequisite is checked to be unsheddable while the dependent
capability itself remains sheddable. Every offered candidate is selected in turn
and its successor checked to validate, to record parent 7, and to survive a
decode; the reported maintenance and mass are checked to equal what
`metabolic_budget` charges that successor.
No organism mutates and no fixture calls any of this. Candidate count, context
window, mutation cost as a game rule, inherited versus acquired expression, NPC
mutation and recombination remain open and are not settled here. Golden
`d61ad5bf0c928c76`, ecosystem fingerprint `2cd4a15f9fbe9949` and the frozen genome
fixture are unchanged. No deployment manifest or image pin changed. Linux
validation is pending; no Gitea result is claimed since `b01a824`.