docs(improve-02-brief): prompt for the next improve branch — E03/E04/E05 strict re-runs

Self-contained prompt to paste into a new chat for
improve/02-strict-eval-e03-e04-e05. Mirrors the structure of the
original improve/01 brief but measurement-only (no prompt
iteration). Covers:

- Branch creation (from improve/01 tip, not main).
- Required reading order (E02 SUMMARY's strict-eval section as the
  methodology reference; the helper scripts in
  apps/temporal/src/scripts/; the schema-aware evaluator's one-of GT
  support; cross-experiment table to update at the end).
- Per-experiment flow: pre-flight → trigger → poll → errors file →
  conditional GT format-variant promotion → re-run if promotion
  happened. With actual command lines.
- What to update per experiment (benchmark-run.json, SUMMARY.md
  strict-eval section, errors-for-gt-cleanup.md, cross-experiment
  table at the end).
- What NOT to do (no seed config changes, no prompt iteration, no
  paid benchmarks per format-variant change).
- Cost estimate ($5-18 depending on promotion re-runs).
- Convergence criteria + report-back checklist.

Use case: copy-paste into a new chat at the start of improve/02 work
so the next agent has full context without re-discovering the
infrastructure. Same template can be cloned for future improve/<NN>
branches that re-measure other engines or apply other cross-engine
changes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: 2 people

experiments/briefs/improve-02-strict-eval-e03-e04-e05.md

@@ -0,0 +1,244 @@
+# Prompt for `improve/02-strict-eval-e03-e04-e05`
+
+Copy this into a new chat to re-run E03/E04/E05 against the cleaned dataset
++ strict evaluator. Measurement-only — no prompt iteration. All three
+experiments go on one branch (`improve/02-...`).
+
+---
+
+Start a new improvement-iteration branch off the current tip of
+`improve/01-strict-eval-and-mistral-tune`. The cross-experiment
+infrastructure landed on `improve/01` is complete: strict evaluator
+(`apps/shared/prisma/seed.ts:2044-2062`), one-of GT array support
+(`apps/temporal/src/evaluators/schema-aware-evaluator.ts`), dataset move
+from `samples-mix/private` → `public`, GT format-variant promotions on
+sin/date/phone, and helper scripts in `apps/temporal/src/scripts/`. E02
+is at canonical pass_rate 0.925 / f1.median 0.972 / matchedFields.median
+69 of 74 — strongest result on record for E02.
+
+This phase re-runs E03, E04, and E05 against the same cleaned-GT +
+strict-rule state so the cross-experiment table reflects strict
+metrics for all five engines. Measurement-only — no prompt iteration.
+
+BRANCH: create `improve/02-strict-eval-e03-e04-e05` from
+`improve/01-strict-eval-and-mistral-tune` (NOT from main — we want
+every E01-E05 provider + the improve/01 evaluator + dataset changes).
+Naming continues the `improve/<NN>-<short-desc>` convention so future
+improve branches stack.
+
+START BY READING (in this order):
+
+1. `experiments/results/02-mistral-doc-ai-azure/SUMMARY.md` — the
+   "## Strict-equality re-evaluation + improvement loop (improve/01)"
+   section establishes the methodology and the canonical pre/post
+   comparison table.
+2. `experiments/POST_BENCHMARK_FOLLOWUPS.md` item 1 — the cross-
+   experiment rollout plan; this branch handles E03/E04/E05.
+3. `experiments/results/05-vlm-ocr-hybrid/SUMMARY.md` cross-experiment
+   table — the source of truth for cross-engine numbers. You'll update
+   E03/E04/E05 rows here at the end.
+4. `experiments/results/02-mistral-doc-ai-azure/iteration/CHANGELOG.md`
+   round-3 and round-5 entries — engine-ceiling diagnosis pattern. E03
+   also uses a Foundry deployment (Azure Content Understanding, not
+   Mistral Document AI) — keep an eye on whether CU's annotation pass
+   has the same OCR-markdown-only limitation that the Mistral SKU has.
+5. `apps/temporal/src/scripts/` — note the helper scripts available:
+   - `trigger-experiment-benchmark.ts <slug>` — POSTs the run, returns
+     the run id. Loads `TEST_API_KEY` from env without leaking it.
+   - `poll-experiment-run.ts <runId> <slug>` — polls until terminal and
+     writes `experiments/results/<slug>/benchmark-run.json`.
+   - `dump-errors-for-gt-cleanup.ts <slug>` — generates the per-sample
+     mismatch table at `experiments/results/<slug>/iteration/errors-for-gt-cleanup.md`.
+     Supports `--known-hard "id1,id2"` (defaults to `"81 blank,81 coffee"`).
+   - `promote-gt-format-variants.ts <slug> [--write]` — auto-promotes
+     sin/date/phone GT scalars to one-of arrays where the engine reads
+     form-as-written. Dry-run by default.
+   - `setup-cu-defaults.ts` — one-time CU defaults patch (E03 only).
+   - `preflight-vlm.ts <deployment>` — checks env + 1×1 PNG strict-mode
+     round-trip (E04).
+   - `preflight-hybrid.ts <deployment>` — same plus Azure DI probe (E05).
+6. `apps/temporal/src/evaluators/schema-aware-evaluator.ts` — note the
+   evaluator now accepts arrays as GT values (any-match across one-of
+   alternates). All five matching rules support it.
+7. `data/datasets/samples-mix/public/` — the cleaned dataset.
+   `sin` / `spouse_sin` / `date` / `spouse_date` / `phone` /
+   `spouse_phone` GTs for 23 samples are already one-of arrays
+   accepting Mistral's form-as-written variants. E03/E04/E05 may
+   produce DIFFERENT format variants (CU and gpt-5.4 may normalise
+   differently than Mistral); `promote-gt-format-variants.ts` is the
+   tool to absorb those too if they show up.
+
+PREREQUISITES (confirm at session start):
+
+- Backend running with the cleaned dataset uploaded to blob storage.
+  If the backend has been restarted since `improve/01` landed but
+  WITHOUT `FORCE_RESYNC_LOCAL_DATASETS=true`, the
+  `samples-mix-public` prefix may not be on blob storage yet — the
+  first benchmark trigger will fail at materialization. Recovery:
+  restart backend once with `FORCE_RESYNC_LOCAL_DATASETS=true` on
+  the env, drop the env var on the next boot.
+- Azure deployments still live from previous work:
+  - `gpt-5.2` capacity 100 (E03)
+  - `gpt-5.4` capacity 100 (E04 + E05)
+  - `text-embedding-3-large` (E03)
+  - Azure DI resource (E05)
+  - CU defaults patched (E03 — run `setup-cu-defaults.ts` if you've
+    rotated the resource)
+- TEST_API_KEY auto-loaded by trigger script from
+  `apps/backend-services/.env` (or the override file). Don't read
+  the value yourself; the script handles it.
+
+KNOWN-HARD SAMPLES (treat as floor, not signal):
+- `"81 blank"` and `"81 coffee"` are low-resolution / obscured forms.
+  Every engine in the stack scores poorly on them. Mistakes are
+  EXPECTED. The `dump-errors-for-gt-cleanup.ts` script tags them
+  ⚠️ KNOWN-HARD by default. Document any movement on these samples
+  but don't iterate against them.
+
+PER-EXPERIMENT FLOW (run independently; each is ~5-30 min wallclock):
+
+### E03 — Azure Content Understanding + gpt-5.2
+
+```bash
+# Optional pre-flight (only needed if CU resource was rotated since
+# the fuzzy-era canonical run); script is idempotent:
+cd apps/temporal
+npx tsx -r tsconfig-paths/register src/scripts/setup-cu-defaults.ts
+
+# Trigger:
+npx tsx -r tsconfig-paths/register \
+  src/scripts/trigger-experiment-benchmark.ts 03
+# → captures run id
+
+# Poll (~15-30 min wallclock; gpt-5.2 generative is the bottleneck):
+npx tsx -r tsconfig-paths/register \
+  src/scripts/poll-experiment-run.ts <runId> 03-content-understanding
+# → writes experiments/results/03-content-understanding/benchmark-run.json
+
+# Errors file for GT cleanup review:
+npx tsx -r tsconfig-paths/register \
+  src/scripts/dump-errors-for-gt-cleanup.ts 03-content-understanding
+```
+
+If the errors file lists ≥ 10 sin/date/phone format-variant mismatches,
+run `promote-gt-format-variants.ts 03-content-understanding` (dry-run
+first). If the proposals look right, apply with `--write` and re-run
+the benchmark — that's one extra paid run but locks in the format
+recovery. Promotions are idempotent + deduped against E02's existing
+one-of arrays.
+
+### E04 — gpt-5.4 VLM-direct
+
+```bash
+# Pre-flight:
+cd apps/temporal
+npx tsx -r tsconfig-paths/register \
+  src/scripts/preflight-vlm.ts gpt-5.4
+
+# Trigger + poll (~6-10 min wallclock):
+npx tsx -r tsconfig-paths/register \
+  src/scripts/trigger-experiment-benchmark.ts 04
+npx tsx -r tsconfig-paths/register \
+  src/scripts/poll-experiment-run.ts <runId> 04-vlm-direct
+
+# Errors:
+npx tsx -r tsconfig-paths/register \
+  src/scripts/dump-errors-for-gt-cleanup.ts 04-vlm-direct
+```
+
+Same conditional promote-GT step as E03.
+
+### E05 — gpt-5.4 VLM + OCR hybrid
+
+```bash
+# Pre-flight:
+cd apps/temporal
+npx tsx -r tsconfig-paths/register \
+  src/scripts/preflight-hybrid.ts gpt-5.4
+
+# Trigger + poll (~5-10 min wallclock; 4:33 was the fuzzy-era run):
+npx tsx -r tsconfig-paths/register \
+  src/scripts/trigger-experiment-benchmark.ts 05
+npx tsx -r tsconfig-paths/register \
+  src/scripts/poll-experiment-run.ts <runId> 05-vlm-ocr-hybrid
+
+# Errors:
+npx tsx -r tsconfig-paths/register \
+  src/scripts/dump-errors-for-gt-cleanup.ts 05-vlm-ocr-hybrid
+```
+
+Same conditional promote-GT step.
+
+WHAT TO UPDATE per experiment:
+
+1. **`experiments/results/<slug>/benchmark-run.json`** — overwritten by
+   the poll script.
+2. **`experiments/results/<slug>/iteration/errors-for-gt-cleanup.md`** —
+   generated by `dump-errors-for-gt-cleanup.ts`.
+3. **`experiments/results/<slug>/SUMMARY.md`** — add a section
+   `## Strict-equality re-evaluation (improve/02)` near the top
+   (mirror E02 SUMMARY's structure). Include:
+   - Run id of the canonical strict run.
+   - The 3-column comparison table:
+     `| | Fuzzy@0.85 (historical) | Strict (no GT cleanup) | Strict + GT cleanup (canonical) |`
+     ... pass_rate, f1.median, f1.mean, precision.mean, recall.mean,
+     matchedFields.median, falsePositives.mean.
+   - One-paragraph retrospective: which samples regressed under strict,
+     whether GT cleanup absorbed format variants (and how many), any
+     engine-specific quirks surfaced.
+   - If GT cleanup happened, list the format-variant categories absorbed
+     (e.g. "CU normalises SINs with hyphens — 12 sin/spouse_sin
+     promotions added").
+4. **Cross-experiment table in `experiments/results/05-vlm-ocr-hybrid/SUMMARY.md`**
+   — once all three are done, replace E03/E04/E05 column metrics with
+   the new strict numbers. Drop the per-row `(fuzzy)` tag. Update the
+   footnote to reflect that all 5 engines are now strict-evaluated.
+   Note which engine benefited most from GT cleanup if there's a
+   clear winner.
+
+WHAT NOT TO DO:
+
+- Don't change `apps/shared/prisma/seed.ts` evaluator config (already
+  `exact` on improve/01).
+- Don't change the workflow JSONs except for the `targetLocalDataset`
+  if it doesn't already point at `samples-mix-public` (it should — the
+  `improve/01` rename touched all 5).
+- Don't change the workflow JSONs' `parameters.documentAnnotationPrompt`
+  / `parameters.fieldDescriptions` / equivalent CU analyzer config.
+  This branch is measurement-only — no prompt iteration.
+- Don't run paid benchmarks on E01 or E02 (out of scope here).
+- Don't run paid benchmarks per format-variant change — measure once,
+  promote GT if obvious, re-measure once. Cost-discipline.
+
+COST EXPECTATION:
+
+- E03 (gpt-5.2 generative on 40 samples): ~$3-5 per run
+- E04 (gpt-5.4 vision on 40): ~$1-2 per run
+- E05 (gpt-5.4 vision + DI on 40): ~$1-2 per run
+- Worst case (each engine needs a second post-promotion run): $10-18
+  total. Plan accordingly.
+
+CONVERGENCE (done when):
+
+- `benchmark-run.json` from this branch exists for E03, E04, E05.
+- Each has a "Strict-equality re-evaluation (improve/02)" section in
+  its SUMMARY.md with the comparison table + retrospective.
+- Each has `iteration/errors-for-gt-cleanup.md` generated.
+- E05 cross-experiment table is updated for all 5 engines strict.
+- Tests pass (CI=true): `cd apps/temporal && npx jest src/experiment-03 src/experiment-04 src/experiment-05 src/ocr-providers/azure-content-understanding src/ocr-providers/vlm-direct src/ocr-providers/vlm-ocr-hybrid`.
+- Lint + tsc clean (pre-commit hooks gate this).
+- Commits on `improve/02-strict-eval-e03-e04-e05`. One commit per
+  experiment is fine, OR a single bundled commit per the
+  "all-on-one-branch" framing. Final cross-experiment-table commit
+  at the end.
+
+REPORT BACK AT END:
+
+- Per-experiment delta vs fuzzy-era (one line each).
+- Updated cross-experiment table (paste it).
+- Total cost spent (rough).
+- Any engine that hit a similar Foundry-style ceiling to Mistral
+  (E03 in particular — CU's annotation pass on Foundry could have
+  the same OCR-markdown-only limitation).
+- Whether any of E03/E04/E05 now beat E02 on the cleaned-GT
+  strict numbers.