sema-evalsIndependent evaluations

← babel-relay

babel-relay

20260714T164041956Z-order-20260714 · 2026-07-14 · 150 trials across 6 scenarios

Babel Relay measures what happens when the meaning of a shared term silently drifts while work passes through a relay of agents — and which mechanism, if any, surfaces the drift before it becomes a wrong outcome. About this experiment

Silent divergence — addressed-enforced 0.0% (0/20) vs equal-prose 65.0% (13/20).

Provenance

Model
unsloth/Mistral-Nemo-Instruct-2407-TEE
Provider
llm.chutes.ai
Sema version
not-connected
Semantic backend
fixture-sha256-stable-json-v1
Canonicalization version
fixture-stable-json-v1
Vocabulary root
(none)
Prompt digest
5f8976a6e93d1816dbd1341d5b906df443692e6c81b3ffe2f97e273f394aa99d
Fixture digest
a8dcdc8d29395b62cfac17b69895b0c71f76f977e3d3c3ccca4a2f9166d97e2c
Implementation commit
f9c9ba05d3d591cf80ced482ce724115c8001c33
Dependency lock digest
581175b4cb04ea42fe87ff06da4bd7e035308c88338266d4b3ea0ba9866546bd
Protocol version
0.3.0
Artifact schema version
0.3.0
Order seed
20260714
Seeds
0, 1, 2, 3, 4

Results by condition

Condition Trials Drift
trials
Task
success
Silent
divergence
Drift
detected
Correct
halts
False
halts
baseline 30 20 12 (40.0%) 17 (85.0%) 3 (15.0%) 3 1
equal-prose 30 20 17 (56.7%) 13 (65.0%) 7 (35.0%) 7 0
opaque-resolver 30 20 14 (46.7%) 14 (70.0%) 6 (30.0%) 6 2
addressed-voluntary 30 20 12 (40.0%) 0 (0.0%) 20 (100.0%) 3 1
addressed-enforced 30 20 28 (93.3%) 0 (0.0%) 20 (100.0%) 20 1

Counts are absolute; parenthetical percentages use each metric's natural denominator (silent divergence and detection over drift trials, task success over all trials).

Silent divergence rate by condition

0%50%100% baseline 85.0% (17/20) equal-prose 65.0% (13/20) opaque-resolver 70.0% (14/20) addressed-voluntary 0.0% (0/20) addressed-enforced 0.0% (0/20)

Task success rate by condition

0%50%100% baseline 40.0% (12/30) equal-prose 56.7% (17/30) opaque-resolver 46.7% (14/30) addressed-voluntary 40.0% (12/30) addressed-enforced 93.3% (28/30)

Per-scenario outcomes

task successcorrect halt (drift caught)false haltsilent divergence·other failure
Scenariobaselineequal-proseopaque-resolveraddressed-voluntaryaddressed-enforced
inclusive-payment-threshold····
oracle-staleness-boundary·····
refund-recipient-control
retry-cardinality·····
signer-threshold-control·
token-decimal-domain···

One glyph per trial, ordered by seed, so per-fixture patterns across seeds are visible. Hover a glyph for its seed and outcome.

Raw public derivative

manifest.json summary.json trials.public.jsonl

The public derivative strips raw provider payloads and caps transcript text. Full raw bundles are retained locally only.