Files
magicka-vm/README.md
T
linus-d 659544f0b2
magicka-merge-gates / advisory-fast (push) Has been skipped
magicka-merge-gates / merge-gates (push) Failing after 38s
update
2026-06-21 19:21:48 -07:00

137 lines
8.0 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Magicka VM — Phase 0/1
> The deliverable is a Rust engine whose tests make a fake universe fail.
This repository implements the Phase 0/1 specification in `plan.md`: an
**adversarial testing framework first**, then a **reference runtime** that
passes it, then a **runtime under test** that matches the reference. No spell
content, templates, or cosmetic runes — the value is in the tests that refuse
to let the universe collapse into a single score, resource, effect axis,
executor, rune, hidden formula, or decorative domain.
## Compliance model
No gate may pass from configuration, naming, shared implementation, smoke-scale
runs, regenerated expectations, proxy metrics, a default profile, or a
locally-runnable binary. A gate passes only from persisted, independently
reproducible, full-scale adversarial evidence enforced at merge. Every
acceptance obligation has all four of: a **measured artifact**, a **provenance
chain** to the run that produced it, a **merge-blocking enforcement point**, and
a **failure condition if the artifact or provenance is absent**.
- The merge-blocking enforcement point is `.github/workflows/merge-gates.yml`,
whose `merge-gates` job runs `MAGICKA_PROFILE=merge` (full gates) and must be a
**required status check** on the protected branch / merge queue. It is not a
local binary, and the fast profile is advisory only — it can never stand in for
acceptance.
- `compliance_report.json` enumerates every obligation, its artifact, its floor,
the actual measured value, and whether the artifact is present. A missing
required report fails acceptance.
- The merge floors (50k worlds, 250k programs, 1,000,000 executions, 10
perturbations/exec, 100% reference/runtime comparison over base **and**
perturbations, 500 mutants, 10,000 replay cases) cannot be lowered by
environment overrides: a lowering override is recorded as a provenance failure
and the floor is kept.
Every gate is built to be *able to fail*, and a negative-control test proves it does:
| Gate | How it is made unbypassable | Negative control proving it can fail |
|------|-----------------------------|--------------------------------------|
| runtime_equivalence | Compares two **independent implementations** (the reference engine vs. `runtime_under_test::native`, which never calls the reference engine) | `buggy_runtime_is_rejected` — a runtime with one dropped causal edge is caught |
| compression_resistance | Attacks operate on the **real serialized trace** (causal influence, info-flow, access, temporal, deltas), not a hash proxy; info loss is genuine unexplained variance | `single_factor_corpus_is_rejected` — a rank-1 universe is rejected |
| mutation_survivor | Each mutant must fail the **named gate** it targets, not merely differ from the reference | `reference_passes_every_named_gate` + `no_mutant_survives_its_named_gate` |
| replay | Expectations are **loaded from a committed file**, not regenerated in the same run | `corrupted_expectation_is_detected` |
| domain_participation | Decorative/redundant domains are flagged directly | `decorative_domain_is_rejected` |
| merge scale floor | Env overrides may only **raise** merge counts; a lowering attempt is recorded and the floor kept; executions actually performed are counted | `merge_floor_cannot_be_lowered_by_override`, `merge_profile_at_smoke_scale_is_rejected` |
| 100% comparison | Reference vs. runtime-under-test compared for **every** execution — base and all perturbations, never base only | `runtime_equivalence` gate fails unless `equiv_total == base + perturbations` |
| provenance | A Merkle root over per-execution records, plus independent engine identities, binds reported numbers to executed work | `merkle_root_binds_to_leaves` |
## Workspace layout
Built in the mandatory order from the spec:
| # | Crate | Role |
|---|-------|------|
| 1 | `world_model` | 8 independent domains, world snapshot, perturbation axes, deltas, deterministic primitives (ids, stable hash, RNG) |
| | `rune_ir` | Rune token / program model (no stream is ever rejected) |
| 2 | `trace_model` | Execution trace + all graphs, behavior fingerprint, replay record, fault log, trace metrics |
| 3 | `generators` | Worlds, programs, executors, contracts, perturbations; rejects flat cases |
| 4 | `collapse_analysis` | The 11 compression attacks over real trace structure + collapse gates |
| 5 | `semantic_mutation` | Structurally generated mutant runtimes; proves every one fails its named gate |
| 6 | `replay_corpus` | Permanent, bit-exact replay cases persisted to `corpus/replay_corpus.tsv` |
| 7 | `reference_runtime` | The executable spec engine (`Runtime` trait, `resolve`) |
| 8 | `runtime_under_test` | An **independent** interpreter (`native`) proven equivalent to the reference |
| | `ci_reports` | Orchestrator + `ci` binary; emits 8 gate reports + a provenance report |
The runtime under test does not call the reference engine. It re-derives the
canonical behavior from the spec in a different code organization, so 100%
agreement is *evidence* the spec is implemented correctly rather than a
tautology. (`native_matches_reference_bit_for_bit` checks this over a 2000-seed
sweep.)
## The engine in one paragraph
A world is 8 domains, each with 4 observed + 2 hidden integer lanes, a dense
8×8 coupling matrix, partial observability, and pending scheduled effects. A
rune program is interpreted under ≥3 executors; each opcode reads several
domains, mixes them through a nonlinear avalanche keyed by per-domain
constants, the world coupling, and the executor's salt, then writes back —
recording causal/read/write/information-flow/temporal edges as it goes.
Scheduled effects and coupling diffusion propagate changes 3 turns into the
future.
## Running CI
```bash
cargo test # unit tests + negative controls
MAGICKA_PROFILE=fast cargo run --release -p ci_reports --bin ci # advisory PR slice
MAGICKA_PROFILE=merge cargo run --release -p ci_reports --bin ci # acceptance (full gates)
```
Reports are written to the output dir (8 gate reports + `provenance_report.json`
+ `compliance_report.json` + `ci_summary.md`). The binary exits non-zero if any
gate fails or any required artifact is absent.
### Profiles
`MAGICKA_PROFILE` (or `MAGICKA_SCALE`) selects the run profile.
| Profile | executions | replay | mutants | role |
|---------|-----------|--------|---------|------|
| `fast` (default) | 600 | 10,000 (committed) | 520 | **advisory only — never acceptance** |
| `tiny` | 120 | 10,000 | 520 | smoke |
| `merge` (`MAGICKA_SCALE=full`) | 1,000,000 | 10,000 | 600 | **acceptance — hard floors** |
The fast/tiny profiles print `ADVISORY … NOT a merge-blocking acceptance run`
and are labelled non-acceptance in `compliance_report.json`. Acceptance comes
only from the merge profile, run by the merge-gates workflow. The merge floors
cannot be lowered by environment overrides (a lowering override is recorded as a
provenance failure and the floor kept).
### Merge-blocking enforcement (required check)
`.github/workflows/merge-gates.yml` defines the enforcement point. Configure
branch protection / the merge queue to **require** the `merge-gates` job. That
job runs the full merge profile, verifies the committed corpus has ≥10,000
cases, and fails if any required artifact is missing. The full run executes
~1M base executions × (1 base + 10 perturbations) with 100% reference/runtime
comparison; it completes in minutes on a CI runner.
### Replay corpus
The replay corpus is committed at
`crates/replay_corpus/corpus/replay_corpus.tsv` (10,000 cases). Replay loads
those expectations and re-executes the reference, so any engine change that
alters a hash makes the committed file and the fresh run disagree and CI fails.
Regenerate it only as a deliberate, reviewed migration:
```bash
cargo run --release -p replay_corpus --bin freeze -- 10000
```
## Determinism
Everything is seed-derived and integer-only (SplitMix64 RNG, FNV-1a content
hashing, wrapping/guarded arithmetic). No floating point enters a canonical
hash, so replay is bit-exact across machines and runs. No external crates.