Adds the P5a recall-precision predicate that was listed as NOT_COVERED in
the L10 continuity lane. Closes the gap in the schema-as-proof discipline
(CLAUDE.md): every predicate must have both a *_holds test (real soak) and
a *_bites mutation test so a passing lane cannot silently miss the violation
it nominally catches.
Changes:
- runner.py: ProbeRecord dataclass + probe_at/verify_probes_at params on
run_soak(). Registers a field state (float32 bytes, matching vault's
_exact_index dtype) at a named turn and recalls it against the vault at a
later turn — including after a reboot, which is the cross-reboot claim.
- predicates.py: evaluate_p5a_recall_precision — fails if any ProbeRecord
has rank=None or rank>top_k, or if no cross-reboot probe was recorded.
- report.py: wires P5a into build_report(); probe registered at turn 1,
verified at reboot_turn+2 (intentionally before the vault's
null_project auto-reproject cycle at store_count=20, which would destroy
all CGA inner-product scores — documented as a real finding, deferred to
a follow-up increment); NOT_COVERED is now empty ().
- contract.md: P5a row in the predicate table + reprojection-boundary
scope note.
- test_l10_continuity.py: 4 tests — holds (real 6-turn soak across reboot)
+ 3 bites (rank=None, no cross-reboot probe, empty probe_records).
Key finding: vault.null_project() fires every vault_reproject_interval=20
stores and produces CGA-orthogonal versors (inner product → 0.0 with the
original), completely destroying both exact-match and ranked recall. This
is a long-horizon vault stability issue, recorded here rather than silently
avoided. The P5a probe window is constrained to the pre-reproject interval
to keep wrong=0 intact while documenting the gap.
Build evals/l10_continuity/, the empirical gate between the two L10 targets
(T-resume: provable same-life resume; T-experience: continuous experiencing
field-life). Drives the REAL turn loop (ChatRuntime + CognitiveTurnPipeline)
over a deterministic in-vocab corpus, with reboot and orphan-crash legs, and
evaluates falsifiable predicates over recorded evidence. Additive only; no
existing file touched; read-only over the runtime; no serving-path import.
Predicates (each with a *_holds real-soak test AND a *_bites mutation test, per
the CLAUDE.md schema-as-proof discipline):
- P1 closure: versor_condition < 1e-6 every turn (green guard).
- P2a determinism: two independent runtimes -> byte-identical trace_hash.
- P2b reboot transparency (the diagnostic): a reboot never alters pre-reboot
turns (hard guard); post-reboot transparency is MEASURED and today FALSE --
the mechanical proof that Shape B (ADR-0146) discards the lived field/vault,
i.e. "many lives sharing a checkpoint". A pinned test flips if persistence is
ever added, forcing a doc update so the gap can't close silently.
- P3 bounded resources: vault grows linear-bounded/monotonic (RSS recorded).
- P4 crash recovery: two recoveries from one checkpoint converge (determinism)
+ commit-point/ARIES force boundary (recovered turn_count == committed) +
atomic-write survives mid-os.replace kill (ADR-0156).
- P5b anchor stability (T-experience crux): field anchors without COLLAPSE
(dist_to_anchor not -> 0) or FREEZE (turn_movement not -> 0); the long-horizon
test of the sanctioned _session_anchor_pull (alpha=0.05). Thresholds measured.
- P5c coherence: surfaces stay non-empty and not collapsed to one output, over
more than one corpus cycle.
- P5a recall precision@k: recorded as not_covered (needs a held-out probe set).
report.py assembles the panel into a structured report with a hardware-stable
deterministic_digest (trace_hash sequence + verdicts; excludes RSS/wall-clock)
as the freeze handle. Run: python -m evals.l10_continuity [n_turns] [reboot_turn].
24 tests pass; adversarially reviewed across 4 lenses (bite-discipline,
invariant/trust-boundary, honesty/determinism, correctness) before landing.
