LimFlex
Evidence · Version 1.3

Different AI paths.
A stable judgment boundary.

The strongest current evidence is structural: across three tested provider/model families and two Fresh passes, AI action paths varied while the external qualification boundary remained substantially more stable.

LimFlex Wrapper connecting specialized AI and operational tools with a shared Judgment Terrarium
Fresh product qualification
192/192

Structural transitions admitted

Across three provider/model families under one frozen external qualification framework.

32 cases × 3 providers × 2 Fresh passes.
Strict combined outcome
191/192

Strict qualified outcomes

Structural validity plus required semantic-coordinate coverage.

The single residual remained visible rather than being repaired or hidden.
Two-pass continuity
96/96

Admission agreement

The external admission boundary agreed across both Fresh passes.

Exact action agreement was lower: 87/96.
Strict two-pass continuity
95/96

Strict-outcome agreement

Qualified outcomes remained highly stable even when model action paths differed.

Internal same-program evidence; not population-level reliability.
The important contrast: exact action paths agreed in 87/96, while admission agreed in 96/96. LimFlex is not freezing reasoning; it is preserving a qualified boundary for what still holds.
Across Different AI Systems

One external continuity surface instead of one model's hidden state.

GPT-5.6 Sol, Claude Sonnet 5 and Gemini 3.8 Flash produced the 192 Fresh candidate slots. The same external qualification semantics admitted all 192 structural transitions.

Shared and inheritable

The tested result supports maintaining qualified judgment outside any single AI model so it can be inspected, inherited and requalified by another system.

3provider/model families in the Fresh programme
25/25lineage-hop survival in companion durability evidence

Protected continuity

Companion Terrarium tests preserved protected state and authority boundaries while judgment lineage continued.

0protected-state drift
0authority leakage
Lightweight Qualification

The qualification core is tiny beside provider latency.

Measured median transition qualification

0.287 msmedian qualification-core time
4.762 smedian provider latency

≈ 0.0060% of median provider latency.

Claim boundary

This is a measurement of the deterministic transition-qualification core. It is not a claim that total deployment, evidence gathering, organizational governance, integration or review overhead is 0.0060%.

Selective Requalification

Update only what changed. Reuse what still holds.

The repeated supporting evidence preserved the same architecture-native work pattern in both Fresh runs.

Less recomputation

144 → 68recomputed coordinates
−52.8%fewer recomputations vs fixed recomputation baseline

More reuse

72 → 148reused coordinates
2.06×reuse multiplier
Version durability: 1000/1000 historical versions preserved with zero collateral mutation in companion deterministic evidence.
Replay · Audit · Recovery

A judgment can remain inspectable after the conversation is gone.

Deterministic durability

1000/1000exact replays
1000/1000historical-version preservation

Recovery and continuation

10/10Terrarium recovery checks
12/12same-RAW requalifications with zero provider re-calls in prior recovery evidence
The Terrarium preserves prior state, actual state relation, lineage, version, unresolved coordinates and authority boundaries so a proposed change is not automatically treated as a qualified change.
Evidence Boundary

Strong internal evidence. Deliberately bounded claims.

Version 1.3 makes the corrected prospective Fresh product qualification the primary evidence and keeps earlier mechanism, recovery and durability results as supporting evidence rather than pooling them into one global score.

SupportedA stable external qualification boundary across the tested provider/model variation.
SupportedReplay, versioning, selective requalification and recovery in the reported internal tests.
Not establishedPopulation-level reliability, all-model generalization or real-world outcome correctness.
Not establishedIndependent external replication, multi-vendor JEZ interoperability, security certification or legal conformity.