How kaelvalen/relic works
RELIC is a Python research implementation of revocable per-fact knowledge updates for frozen language models. Its central workflow trains a low-rank adapter for a fact, checks divergence, stores accepted entries, and routes queries to at most one entry or the unchanged base model; answers retain entry provenance. The tree also contains confirmatory evaluation, a text live-learning sandbox, and a VLM mirror sandbox. Edges follow README-described behavior; sampled excerpts do not expose enough of the runtime implementation to assert additional call-level wiring.
Main entry points: router.py Entry router · serving.py Answer serving · pilot.py Confirmatory pilot · model.py Frozen model
Components 23
RELIC Runtime
- Frozen model
relic/model.py - Adapter training
relic/adapters.py - Entry ledger
relic/ledger.py - Entry router
relic/router.py - Answer serving
relic/serving.py - Divergence metrics
relic/metrics.py
Confirmatory Study
- Fact generation
experiments/sandbox/live_learning/confirm/facts.py - Confirmatory pilot
experiments/sandbox/live_learning/confirm/pilot.py - Routed evaluation
experiments/sandbox/live_learning/confirm/routed_eval.py - Locality evaluation
experiments/sandbox/live_learning/confirm/locality_eval.py - Router calibration
experiments/sandbox/live_learning/confirm/router_v4_calibrate.py - Adapter comparison
experiments/sandbox/live_learning/confirm/adapter_only.py
Text Sandbox
- Learning stream
experiments/sandbox/live_learning/text/stream.py - Learning mechanisms
experiments/sandbox/live_learning/text/mechanisms.py - Text runner
experiments/sandbox/live_learning/text/run.py - Ledger bandit
experiments/sandbox/live_learning/text/ledger_bandit.py
VLM Sandbox
- VLM harness
experiments/sandbox/live_learning/vlm/vlm_core.py - VLM ledger
experiments/sandbox/live_learning/vlm/vlm_ledger.py - VLM controller
experiments/sandbox/live_learning/vlm/vlm_controller.py
Other components
- User
- New fact
- Query
- Answer + provenance
Connections 22
- User → New fact: submits
- New fact → Frozen model: rewrites with
- Frozen model → Adapter training: trains from
relic/adapters.py - Adapter training → Divergence metrics: checks canaries
- Divergence metrics → Entry ledger: accepts if below limit
relic/ledger.py - User → Query: asks
- Query → Entry router: routes
- Entry ledger → Entry router: provides entries
relic/ledger.py - Entry router → Answer serving: selects entry or abstains
- Answer serving → Frozen model: answers with
- Answer serving → Answer + provenance: returns provenance
- Fact generation → Confirmatory pilot: supplies facts
experiments/sandbox/live_learning/confirm/pilot.py - Confirmatory pilot → Routed evaluation: evaluates store
experiments/sandbox/live_learning/confirm/routed_eval.py - Confirmatory pilot → Locality evaluation: evaluates locality
experiments/sandbox/live_learning/confirm/locality_eval.py - Router calibration → Entry router: calibrates routing
experiments/sandbox/live_learning/confirm/router_v4_calibrate.py - Adapter comparison → Adapter training: compares adapters
- Learning stream → Text runner: provides events
experiments/sandbox/live_learning/text/run.py - Text runner → Learning mechanisms: runs mechanisms
experiments/sandbox/live_learning/text/run.py - Ledger bandit → Learning stream: consumes stream
experiments/sandbox/live_learning/text/ledger_bandit.py - VLM controller → VLM harness: uses harness
experiments/sandbox/live_learning/vlm/vlm_controller.py - VLM controller → VLM ledger: uses ledger
experiments/sandbox/live_learning/vlm/vlm_controller.py - VLM ledger → VLM harness: uses core
experiments/sandbox/live_learning/vlm/vlm_ledger.py