Warden plan

28 - CR-16 (pending): Small-projector agreement

← eval suite index


tier: hermetic requires: []


28 - CR-16 (pending): Small-projector agreement

Pending stub — GATED criterion. The precondition below is not built; this plan reports the criterion as pending (skip, never fail) until it lands. When it does, replace this stub with a real plan in the shape of 25-cr-18 / 26-cr-25 / 27-cr-20.

Criterion implemented (verbatim from the register)

Register: ~/.studio/master.withrobin.ai/project/rfc.one/criteria-register.md (status there: PENDING OWNER ACCEPTANCE; accepted by the 2026-08-31 proxy screen).

### CR-16 — Small-projector agreement - Statement: Projection can be done by a small fine-tuned model in agreement with the frontier projector at a tenth of the cost. - Metric: r_s or κ between small projector (fine-tuned on frontier bootstrap) and frontier projector, per Dimension, with confidence interval; per-projection cost ratio. - Dataset/inputs: Held-out sample per Dimension of up to 200 Claims (≥30 per Domain). - Threshold: Agreement meeting §5.4's Measurable thresholds on ≥3/4 of active Dimensions, at ≤1/10 the frontier projector's per-projection cost. - Drop condition: Below threshold keeps projection on the larger model, making c_π the binding constraint on active-Dimension count (lower K, higher W(m) in CR-14's ledger). - Source: Anchor §10.6. - Tier: gated — requires Dimensions, projection machinery, and a trained small projector (note: one of the anchor's two fine-tuned components, in tension with a no-training posture).

Named precondition

Precondition: requires Dimensions, projection machinery, and a trained small projector. Not built at HEAD 7f379974 (verified 2026-08-31: no 'claims', 'dimensions', or 'initiatives' tables in server/src/db/schema.ts, and no conflicts_with edges anywhere in server/src or packages/). The trained small projector has no code marker to probe (one of the anchor's two fine-tuned components, in tension with a no-training posture — the register carries that flag).

Skip guard

set -uo pipefail
source "$WARDEN_LIB/assert.sh"
cd "${PROJECT_ROOT:-$(git rev-parse --show-toplevel)}"

SCHEMA=server/src/db/schema.ts
MISSING=""
grep -q "'dimensions'" "$SCHEMA" || MISSING="$MISSING dimensions-tables"

if [ -n "$MISSING" ]; then
  warden_skip "CR-16 — Small-projector agreement" "pending: requires Dimensions, projection machinery, and a trained small projector — gated criterion, precondition not built (probes missing:$MISSING)"
else
  warden_skip "CR-16 — Small-projector agreement" "schema probes now pass — upgrade this pending stub to a real plan before CR-16 can bind (probes are heuristic, not proof)"
fi

Shape (note for the next author)

Pending stub, hermetic on purpose: the guard reads the checkout only and emits skip in every branch, so this plan can never red a run — the criterion surfaces in every summary as pending with its precondition named. The probes are heuristic activation bells, not proof: quoted table names (and conflicts_with where the contradiction check is a precondition) flip the detail text when the machinery lands. The real plan must quote the criterion verbatim, turn its thresholds into warden_pass/warden_fail gates, and follow the conventions the three runnable plans set: machine-local fixtures (25-cr-18), NLI probe and versioned judge prompts (26-cr-25), data-keyed guard activation (27-cr-20).