plan: WP-0004 — adaptive cost-quality routing (todo)
Some checks failed
CI / test (3.10) (push) Has been cancelled
CI / test (3.11) (push) Has been cancelled
CI / test (3.12) (push) Has been cancelled

Draft the workplan that extends the static RoutingPolicy (WP-0003) with
a quality observation ledger, a BaselineGrader (ClaudeCodeAdapter as the
default oracle), an AdaptiveRoutingPolicy that picks the cheapest
adapter clearing a per-task quality floor, and a sampled
ShadowingAdapter for production observation collection.

Scope is explicit: ship primitives only. Task-type taxonomy, quality
thresholds, baseline choice, and re-grading cadence stay with the
consumer. infospace-bench is the named first consumer; consumer wiring
deferred until T01-T03 land.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
2026-05-17 17:17:07 +02:00
parent 66dfc7cf06
commit deade6ad76

View File

@@ -0,0 +1,166 @@
# LLM-WP-0004 — Adaptive Cost-Quality Routing
**status:** todo
**owner:** llm-connect
**repo:** llm-connect
**created:** 2026-05-17
**depends-on:** LLM-WP-0003 (RoutingPolicy primitive)
## Purpose
Provide reusable primitives that let a consumer route each task to the
cheapest model whose observed output quality clears a per-task bar, with
the local Claude Code session (`ClaudeCodeAdapter`) available as the
baseline-quality oracle. The current `RoutingPolicy` (LLM-WP-0003) is
static: rules and cost caps are hand-authored. This workplan adds the
observation, grading, and adaptive-selection layer that learns *which*
model is good enough for which task type.
Demand signal: `infospace-bench` is about to scale from one-chapter
smoke runs to multi-chapter and full-book infospace generation across
multiple workflow stages (summarise, extract entities, extract
relations, evaluate). Each stage has very different quality / cost
trade-offs, and the consumer needs llm-connect to pick the right model
per stage instead of hard-coding one model for the whole run.
## Scope guardrails (read this before adding tasks)
llm-connect ships *primitives*, not consumer policy.
In scope here:
- Data models for quality observations and grading results
- A reusable observation ledger (persistent, append-only, file-backed)
- A `BaselineGrader` that pairs a baseline adapter with a candidate
adapter and emits a structured quality delta
- A small built-in grader catalogue (exact-match, embedding similarity,
LLM-as-judge wrapper)
- An `AdaptiveRoutingPolicy` that extends `RoutingPolicy` by consulting
the ledger to pick the cheapest adapter whose observed quality for a
task type still clears a configured threshold
- A shadow-mode wrapper adapter for collecting observations in
production without changing caller behaviour
Out of scope (belongs in consumer repos):
- Task-type taxonomy (callers name their tasks)
- Quality thresholds per task type (callers set their own bars)
- Choice of baseline (callers wire whichever adapter they trust)
- When to re-grade (callers decide; this repo just exposes ledger TTL
and refresh helpers)
- Cost accounting for billing or budgets beyond a per-call estimate
## GAAF notes
All additions are Functional-layer per GAAF-2026. Core stays untouched.
Each new module gets a functional contract doc under
`contracts/functional/`. Maturity on release: Beta — `infospace-bench`
is the first known consumer; the API may shift before any second
consumer (`inter-hub`, `markitect`) adopts it.
## Tasks
### T01 — Quality observation data model + ledger
| ID | Title | Priority | Status |
|-----|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------|--------|
| T01 | `QualityObservation` dataclass: `task_type`, `adapter_id`, `model_id`, `cost_usd`, `quality_score` (0..1), `latency_ms`, `tokens_in`, `tokens_out`, `baseline_adapter_id`, `recorded_at`, `tags` | high | todo |
| T02 | `QualityLedger` append-only JSONL store with file-locked writes, configurable path, simple query helpers (`by_task_type`, `recent`, `mean_quality`) | high | todo |
| T03 | TTL helpers: `prune_before(timestamp)` and `is_stale(observation, max_age)` so callers can refresh observations without re-reading the whole ledger | medium | todo |
| T04 | Functional contract doc for the ledger schema and the field semantics of `quality_score` | medium | todo |
| T05 | Tests: round-trip, concurrent writes, query helpers, TTL, malformed-line resilience | high | todo |
### T02 — Baseline grader
| ID | Title | Priority | Status |
|-----|----------------------------------------------------------------------------------------------------------------------------------------------------|----------|--------|
| T06 | `GradingResult` dataclass: `quality_score`, `notes`, `grader_id`, `baseline_response`, `candidate_response` | high | todo |
| T07 | `BaselineGrader` protocol: `.grade(baseline_adapter, candidate_adapter, prompt, run_config)``GradingResult`; built-in concrete `PairedGrader` runs both calls and delegates to a `Judge` | high | todo |
| T08 | `Judge` protocol + three built-ins: `ExactMatchJudge`, `EmbeddingSimilarityJudge` (uses an embedding adapter), `LLMJudge` (uses a third adapter with a fixed rubric prompt) | high | todo |
| T09 | Functional contract doc covering judge bias caveats (length bias, format bias, position bias for `LLMJudge`) | medium | todo |
| T10 | Tests: each judge against canned inputs, grader emits stable result with both responses preserved, deterministic seed for `LLMJudge` rubric | high | todo |
### T03 — Adaptive routing policy
| ID | Title | Priority | Status |
|-----|--------------------------------------------------------------------------------------------------------------------------------------------------|----------|--------|
| T11 | `AdaptiveRoutingPolicy` extends `RoutingPolicy`: given `task_type` + `quality_floor` + `ledger`, returns the cheapest adapter whose observed mean quality clears the floor over a configurable window | high | todo |
| T12 | Tie-breaking: when two adapters meet the floor, prefer lower observed cost; if still tied, prefer the explicitly-preferred adapter from the underlying static rules | medium | todo |
| T13 | Cold-start behaviour: when no observations exist for a `(task_type, adapter)` pair, fall through to the static `RoutingPolicy.resolve` result so the system stays usable on day zero | high | todo |
| T14 | Functional contract doc; document the trade-off between sample size and freshness | medium | todo |
| T15 | Tests: floor enforcement, tie-break, cold-start, window-size effect, fallback chain | high | todo |
### T04 — Shadow-mode observation wrapper
| ID | Title | Priority | Status |
|-----|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------|--------|
| T16 | `ShadowingAdapter` wraps a candidate adapter; on each call, also invokes the baseline adapter (sync or via a thread pool), grades, and appends to a `QualityLedger` | medium | todo |
| T17 | Sampling: caller-configurable fraction (`shadow_rate=0.1` means grade one call in ten) so production load is not doubled | medium | todo |
| T18 | Failure isolation: shadow errors never affect the candidate response returned to the caller; failures are logged but not raised | high | todo |
| T19 | Functional contract doc | low | todo |
| T20 | Tests: candidate response always returned even when baseline raises, ledger gets exactly `shadow_rate × calls` entries (within tolerance), sync vs async modes | high | todo |
### T05 — End-to-end example + integration test
| ID | Title | Priority | Status |
|-----|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------|--------|
| T21 | Example script: route a small fixture batch through three candidate adapters (one OpenRouter cheap, one OpenRouter mid, `ClaudeCodeAdapter` as baseline), grade each, populate ledger | medium | todo |
| T22 | Integration test with mocked adapters covering: cold-start → static fallback → first observations → adaptive selection converges to the cheapest qualifying adapter | high | todo |
| T23 | Brief consumer-integration guide in `docs/` showing how `infospace-bench` (or any caller) wires task-type-per-stage into the adaptive policy | medium | todo |
## Risks and open questions
- **Judge bias.** `LLMJudge` has known biases — length, position, format,
self-preference when the judge model is the same family as a
candidate. The contract must document these and recommend pairing
with at least one non-LLM judge for calibration.
- **Baseline cost in shadow mode.** `ClaudeCodeAdapter` is not per-call
billed (it shells out to a local subscription session), but every
shadow call still consumes wall-clock and rate-budget. Sampling is
load-bearing, not optional.
- **Non-stationarity.** Provider model updates, prompt changes, and
template edits all silently invalidate prior observations. Plan for
a `prompt_fingerprint` tag on observations so the ledger can be
filtered to a coherent regime.
- **Scope creep.** "Pick a model" is one decision; "decide whether the
task is worth doing at all" is another. The latter is consumer
policy. Keep this workplan firmly on the former.
- **Privacy.** Observations contain prompt and response text by
default. Add a `redact: Callable[[str], str]` hook on the ledger
writer so sensitive callers can store hashes / digests instead.
- **API-vs-CLI baseline parity.** A consumer that grades against
`ClaudeCodeAdapter` (CLI) but later switches to a Claude API adapter
may see quality drift that's actually transport drift. Document this.
## Exit criteria
- `QualityLedger` round-trips observations and exposes the documented
query helpers
- `BaselineGrader` produces deterministic `GradingResult` objects for at
least one non-LLM judge and one LLM judge given canned inputs
- `AdaptiveRoutingPolicy.resolve(task_type, quality_floor=0.8)` returns
the cheapest adapter whose mean quality over the configured window
clears the floor, with a documented cold-start fallback
- `ShadowingAdapter` never alters the candidate response and respects
the sampling rate within statistical tolerance
- End-to-end example produces a ledger with at least three adapters per
task type and the integration test shows convergence to the cheapest
qualifying adapter
- Functional contracts published for the new data models, the grader
protocols, and the adaptive policy
## Consumer-side follow-up
`infospace-bench` will need a small companion workplan (`IB-WP-NNNN`) to:
- Replace direct `OpenRouterAssistedGenerationAdapter` use with a
task-type-tagged route through `AdaptiveRoutingPolicy`
- Define its task-type taxonomy (`summarize-source`, `extract-entities`,
`extract-relations`, `evaluate-entity`, `synthesize-report`)
- Pick a baseline adapter (most likely `ClaudeCodeAdapter`) and a
quality threshold per stage
- Wire the shadow-mode wrapper for the first multi-chapter run so the
ledger fills up while real generation proceeds
That workplan should be drafted after T01T03 of this workplan land, so
that the consumer-side wiring is anchored in a stable llm-connect API.