Reading-guide governance + newcomer onboarding paths — two entry artifacts + 3 reading paths + Headline-finding-block + interpretation pedagogy + pointer convention
Superseded on one or more axes by ADR-054061. The body below retains its original prose per the ADR-073 immutability rule; the corrected position lives in the superseding ADR. See the Decisions index to navigate.
ADR-053 — Reading-guide governance + newcomer onboarding paths
Status
Accepted (2026-05-18; landed in v1.0.4 patch alongside the stale-content sweep that surfaced the gap).
Context
The Quarto site landing page (index.qmd) is the highest- visibility surface for a reviewer who clicks into the live site root. EXECUTIVE_SUMMARY.md (added at v1.0.3 per NEXT_STEPS §1.7) is the parallel 1-page decision-maker artifact. Until v1.0.4 neither file’s content shape was governed by an ADR:
- ADR-030 locks Quarto HTML as the deliverable format (vs PDF) but does not anchor reading-flow architecture, landing-page content, or interpretation pedagogy.
- ADR-033 locks the GH release / reviewer-URL strategy (rehearsal + canonical + patch tag pattern) but does not anchor landing-page content.
- NEXT_STEPS.md §1.7 documents EXECUTIVE_SUMMARY’s intended role (“1-page decision-maker layer above the full WRITEUP”) but is a forward-looking work-list, not a governance ADR.
The v1.0.4 stale-content audit (this re-plan turn) surfaced that index.qmd had drifted materially:
- Silent on the actual headline finding. The pre-v1.0.4 index said “pooled IID + pooled OOD numbers per rung” — described WHERE to look, never WHAT the result is. A reviewer reading only the reading guide learned zero headline numbers.
- Stale ADR count. Index said
34+ ADRs; actual at v1.0.3 close was 52 (v1.0.4 lands ADR-053 → 53). - Stale make commands. Index said
make eval-from-hub RUNG=modernbert-lora+make smoke; canonical targets areRUNG=frozen-probe/RUNG=lora+make test-smoke. - Stale repo-identity URLs. Submission anchors used the pre-rename
prompt-injection-detection-submissionslug (repo renamed toprompt-injection-detection-prototypeat v1.0.0 Item 1; index.qmd was missed). - Stale Status section. Index said “At Phase 0-07 close, the spokes are skeletons; Phase 5 populates them” — Phase-0-time scaffolding language anchoring the reader in the wrong temporal frame; Phase 5 closed at v1.0.0.
- No reference to EXECUTIVE_SUMMARY.md. The 1-page artifact added at v1.0.3 was sidebar-listed but never promoted as the A1 Quick-skim step 1 — the natural cold-newcomer arc (1-page thesis → reading guide → deep-dive) was not surfaced.
The user explicitly named the governance dimension: “Does it conform to our initial guidance and/or does our ADRs need to be enriched?” + “We need in the guide to include a few analysis of the results and how to interpret them! how did we overlook this?” (2026-05-18).
The “how did we overlook this” framing is the load-bearing diagnosis: reading-guide architecture without an explicit governance anchor for what the guide must contain (results + interpretation, not just navigation) allowed the v1.0.x trail to drift into stale-scaffolding + results-silence. ADR-053 captures the lesson.
Decision
The reading-guide architecture is governed in 5 dimensions:
1. Two entry artifacts with distinct roles
EXECUTIVE_SUMMARY.md— 1-page decision-maker / time-constrained-reviewer entry. Thesis sentence + 4 headline claims + what-was-characterised + what-is-deferred + reading-path pointer + honest reading. ≤300 words public-facing. Third-person register. Personal voice + apology only appears in the parallel gitignoredSUBMISSION.mdcover letter (per.gitignore:35).index.qmd— Reading guide. Reviewer-who-clicks-into- live-site-root entry. Carries Results section + interpretation pedagogy section + 3 reading paths + submission anchors + Status section. ~150-200 lines.
These are NOT interchangeable. EXECUTIVE_SUMMARY is the distillation; index.qmd is the orientation.
2. Three reading paths canonical
- A1 Quick-skim (~15 min) — for the hiring-manager / executive read. Step 1 is always EXECUTIVE_SUMMARY (the pointer convention; see §5).
- A2 Audit (~60 min) — for the ML-researcher / due-diligence read. Includes a “Headline ADRs to read” curated sub-list so audit readers don’t have to grep through 53+ files.
- A3 Reproduce (~30 min CPU; $0) — for the engineer who wants the numbers to land on their machine. Anchors at the T0
make eval-from-hubreproducibility tier (per ADR-034); falls back to T1make test-smokefor no-network code- health verification.
These 3 paths are the canonical set; superseding ADRs may add A4+ paths or revise these, but cannot silently drop one.
3. Headline-finding-block-on-index is required
index.qmd MUST state the headline numbers up-front, in the first screen of the rendered page. The pre-v1.0.4 pattern (“see WRITEUP §5 for the table”) is banned; a reviewer reading ONLY the reading guide must learn the cross-family OOD finding + the AUPRC trio + the prevalence baseline. Numbers come from evals/bootstrap/marginal_cells.parquet (BCa CI, 10000 resamples) post-single-class-filter; placeholder bounds are banned.
4. Interpretation pedagogy on index.qmd is required
index.qmd MUST include “how to read these numbers” framing patterns. The reviewer should not need to assemble the framing themselves by clicking through to WRITEUP/eval-design.md §5. Canonical set of 5 patterns (subject to revision via superseding ADR):
- Prevalence baseline vs chance — AUPRC’s random-predictor floor equals the positive prevalence (not 0.5).
- Cross-family vs cross-source OOD — the OOD wall is on attack types absent from training, not source-distribution shift.
- Negative LoRA delta meaning — fine-tuning the head on the LODO direct-injection pool actively hurts cross-family OOD; the pretrained backbone carries what little OOD signal exists.
- ProtectAI v1 → v2 non-monotone — publication version monotonicity is not guaranteed under a cross-family slate; selection must be slice-aware.
- val → LODO threshold transfer fails — standard “tune on val, ship to prod” recipe under-quantifies operational FPR; dual-policy threshold (per ADR-025 + WRITEUP/threshold- policy.md §7) is the project’s response.
Pedagogy lives on index.qmd (not in EXECUTIVE_SUMMARY) because the reading guide is the first contact with the methodology; the executive summary is for readers who already know they want results-only. WRITEUP/eval-design.md §5 carries the deeper framing for A2 audit-path readers.
5. Pointer convention
index.qmd → EXECUTIVE_SUMMARY.md → WRITEUP.md → WRITEUP/*.md (8 spokes) → decisions/ADR-NNN-*.md
index.qmdMUST promote EXECUTIVE_SUMMARY as A1 Quick-skim step 1.EXECUTIVE_SUMMARY.mdMUST end with a reading-path pointer to WRITEUP for the full methodology narrative.- WRITEUP + spokes MUST cross-reference each other for sub-topic deep dives.
- Spokes MAY direct-reference ADRs for governance trail.
index.qmdMAY direct-reference ADRs in the curated “Headline ADRs to read” sub-list of the A2 audit path (the exception — without this, audit-path readers face the full 53-ADR ledger).
Consequences
Positive
index.qmdbecomes governance-load-bearing for newcomer onboarding. Future edits must preserve the 5 governance dimensions or supersede ADR-053.- EXECUTIVE_SUMMARY.md’s role is now ADR-anchored, not just NEXT_STEPS-anchored. NEXT_STEPS §1.7 gets a backref to ADR-053 in the v1.0.4 patch.
- The interpretation-pedagogy requirement names the oversight that allowed v1.0.x to ship a reading guide silent on the actual results. Future iterations cannot re-introduce that gap without an explicit supersession.
- The 5-pattern interpretation set is enumerated, so a reviewer audit of the reading guide has a checklist to verify against.
- Reviewer onboarding arc is explicit: 1-page thesis (EXECUTIVE_SUMMARY) → orientation + results + interpretation (index.qmd) → full methodology (WRITEUP + spokes) → governance trail (ADRs). No reviewer should need to reconstruct this flow themselves.
Negative
- Page weight on
index.qmd. The v1.0.4 rewrite added ~150 lines (Results table + 5 interpretation patterns); page is no longer terse-navigation-only. Mitigation: the 3 reading paths come AFTER the interpretation pedagogy so a reader who only wants navigation can skip past via the in-page table of contents. - Pointer-convention drift risk in future patches. If a WRITEUP spoke is renamed without updating
index.qmdcross-references, the reading guide’s curated link list goes stale. Partial mitigation via the v1.0.4make siterender-time link check (Quarto reports broken cross-refs). - Locks the 5 interpretation patterns. A future iteration that discovers a 6th pattern must either supersede ADR-053 or add the pattern via patch ADR (analogous to how ADR-050 Revisions 1 + 2 amended the rung slate). This is intentional — the rigidity is what prevents the silent-drift recurrence.
Neutral
- ADR-030 (Quarto deliverable) unchanged.
- ADR-033 (release strategy) unchanged.
- NEXT_STEPS.md §1.7 (EXECUTIVE_SUMMARY origin note) keeps its forward-looking framing but gains a “ADR-053 retroactively anchors this artifact’s role + interpretation-pedagogy requirement” reference line.
Alternatives Considered
B. Single entry artifact (only EXECUTIVE_SUMMARY OR only index.qmd)
Eliminate the duplication; pick one. Rejected because the artifacts serve distinct audiences: EXECUTIVE_SUMMARY for the 30-second decision-maker read; index.qmd for the reviewer orienting before clicking into deeper artifacts. The 1-page distillation + the orientation guide are not the same artifact even when they share content.
C. Defer the ADR to v1.0.5
Land the v1.0.4 stale-content sweep alone; write ADR-053 in a v1.0.5 governance patch. Rejected in /exploring-options batch 2 + 3 (2026-05-18): the v1.0.4 patch is already restructuring index.qmd to add the Results + interpretation-pedagogy sections; landing the governance ADR atomically with the implementation is cleanest (one tag, one release notes section, one commit graph entry).
D. Narrow ADR-053 scope (entry artifacts only)
Only lock the two-entry-artifacts dimension; let the 3-path / Headline-finding / interpretation-pedagogy / pointer conventions stay implicit. Rejected in /exploring-options batch 3 Q3 (full-scope option locked): a narrow ADR doesn’t prevent the recurrence the user diagnosed (“how did we overlook this”) — the interpretation-pedagogy requirement is the load-bearing dimension because that’s the gap that shipped at v1.0.0/v1.0.1/v1.0.2/v1.0.3.
E. Don’t write an ADR; treat the reading guide as documentation
Rejected because the project’s working contract (CLAUDE.md “Anti-patterns to avoid”: “Adding a methodology component without an ADR”) treats reader-facing methodology surfaces as governance-load-bearing. The reading-guide architecture governs how reviewers consume the methodology — that’s methodology-adjacent enough to warrant an ADR.
Links
- ADR-030 — Deliverable format: Quarto HTML site — locks the deliverable format that this ADR’s reading guide renders on.
- ADR-033 — GitHub release strategy — locks the 3-URL reviewer-submission set referenced from index.qmd’s submission anchors.
- ADR-050 — Rung-slate narrowing — sources the rung-slate framing the reading guide presents.
- ADR-052 — Full-FT OOD methodological reframing — sources the “negative LoRA delta” interpretation pattern.
- NEXT_STEPS.md §1.7 EXECUTIVE_SUMMARY origin — retroactively governed by this ADR.
- WRITEUP/eval-design.md §5 — the deeper AUPRC vs AUROC + interpretation framing that A2 audit-path readers continue into.
- WRITEUP/threshold-policy.md §7 — the dual-policy threshold characterisation referenced in interpretation pattern #5.