#Canticle Updates · Lab Notes · 2026-07-10

update / state snapshot

SEAM — Where We Are

Snapshot 2026-07-10 · HISTORY#374 · 374 entries across 85 days of development. Measurement integrity landed; the score picture is honest; the remaining lever is answerer strategy.

╭─ Score progression ─╮

cat1 / cat3 → target 0.80

80% 65% 50% 35% 20% target 0.80 Old knee Broad +Prompt v1 +DeepSeek TARGET
cat1 (entity QA) cat3 (inference) —— target 0.80
╭─ Holdout funnel ─╮

What happens to 82 holdout questions

Evidence classifier (PR#371) on the corrected holdout baseline — the first honest inventory of where answers land.

Answered correct
35
Open-domain
14
Uncertain (1 turn)
30
Retrieval miss
3
Clean answerer miss
0

Key insight: the old v1 classifier claimed 33 answerer_miss cases. evidence/1 finds zero — every non-correct case had gold scattered across turns or was open-domain.

╭─ 3-PR program ─╮

cat1/cat3 → 0.80 program

PR 1 · Measurement Integrity LANDED #371

Evidence classifier (evidence/1) — fixed crude context_recall precision + the ≥0.5 failure classifier. v1 fields preserved byte-identical.

PR 2 · Judge + Rejudge LANDING #372–374

Judge/2 prompt fixes alias/abbrev under-scoring + groundedness axis. Replay harness for 82 stored answers (~$0.007 max). Judge version plumbing fixed in #374.

PR 3 · Answerer Strategy NOT STARTED

The actual lever. All prior answerer work was diagnostics. This is where the score has to move.

2 of 3 PRs landed · zero paid spend so far · all paid calls operator-gated

╭─ Levers ─╮

Tried & exhausted vs still open

Lever Result Verdict
Retrieval tuning recall barely moves compact→broad EXHAUSTED
Prompt v1 (include all) +0.041 cat1 / +0.027 cat3 dev; over-gen on holdout MIXED
Stronger answerer (DeepSeek-v4-pro) cat1 0.689 vs anchor 0.705 — noise margin WASH
Prompt v2 (precision) 0/17 cases changed at all NULL
Incomplete-search fix Only confirmed fixable pattern (2/11 cases) OPEN

4 of 5 meaningful levers exhausted. Only incomplete-search + the strategic 0.80-achievability question remain.

╭─ Key numbers ─╮

Holdout + project stats

0.689 cat1 best holdout
0.429 cat3 best holdout
−0.111 cat1 gap to 0.80
−0.371 cat3 gap to 0.80

374 HISTORY entries
~$0.59 total paid spend
82 holdout cases
0 clean answerer misses
╭─ The problem ─╮

Post-measurement correction

What we thought (v1)

  • 33 answerer_miss cases — answerer failing to use available evidence
  • cat3 is “generation-bound” — retrieval fine, answerer weak
  • Fix answerer prompt/model → scores go up
  • Context_recall ≥ 0.5 = gold is available

What we know now (evidence/1)

  • Zero clean answerer misses — gold scattered across turns
  • Gold-label incompleteness is real
  • Context_recall false-positives on generic tokens
  • Multi-turn entity coreference: same person → different entity ID per turn

Classifier shift (of 82 cases)

correct
35
v1 miss
33
e/1 miss
0
v1 ret.miss
14
e/1 ret.miss
3
uncertain+
44

The “answerer-bound” framing rested on classifier false confidence. Real bottleneck: scattered evidence, gold incompleteness, task-design tension.

╭─ Options ─╮

What’s next

A. Complete the 3-PR program

Finish PR#136 merge → paid rejudge with judge/2 (~$0.007) → design PR 3 answerer strategy from corrected evidence. Highest EV, lowest cost.

B. Manual classify remaining 44

Only 11/44 rigorously hand-checked. Completing this sharpens the fixable-error rate and ceiling before more spend. ~2–4h operator time, $0.

C. Incomplete-search pattern fix

Only confirmed fixable pattern (2 of 2 real errors): multi-pass/re-scan or structural entity-grounded scoring. Unscoped, uncosted.

D. Fix failure_class threshold

context_recall ≥ 0.5 is too loose. Affects trust in miss counts on every future run. Small, free, defensive.

E. Is 0.80 achievable here?

Given gold incompleteness and multi-turn scatter, can answer-quality tuning alone reach 0.80 — or does the path need per-turn entity resolution first? Operator call.

╭─ Activity cadence ─╮

HISTORY entries · last 30 days

06-11
06-12
06-13
06-14
06-15
06-16
06-17
06-18
06-19
06-20
06-21
06-22
06-23
06-24
06-25
06-26
06-27
06-28
06-29
06-30
07-01
07-02
07-03
07-04
07-05
07-06
07-07
07-08
07-09
07-10

2026-07-08 spike = paid diagnostic + DeepSeek infra + run-record + numpy cache in one session. Steady 3–5 entry/day otherwise.

╭─ Topics ─╮

What we’ve been working on

history
263
verify
249
status
148
benchmark
124
protocol
100
audit
75
retrieval
69
security
50
docs
45
locomo
40

Top topics by HISTORY entry count. Recent shift toward judge / answerer / measurement.