Squintly

Phone-first, viewing-condition-aware psychovisual data collection for zensim. Pre-registered study of how a condition-conditioned quality metric can close zensim's SROCC plateau on real web content.

Existing public IQA datasets (KADID-10k, TID2013, CID22, AIC-3) bake fixed viewing conditions into their labels: lab monitor, specified viewing distance, specified ambient. zensim trained on them plateaus around SROCC 0.82 on CID22, and the residual is structured by viewing conditions — DPR, intrinsic-to-device-pixel ratio, viewing distance, ambient, gamut. Squintly collects pairwise + threshold judgments with those conditions recorded as first-class data per trial, on the phones that actually deliver the bulk of web image traffic.

Two outputs, both consumed by zensim training:

1. Per-encoding latent quality θ_{s,e,c} (BT-Davidson scale, anchored to CID22 via shared anchors) — replaces zensim's training labels and adds a condition vector.
2. Threshold function q_threshold(c, k) for k ∈ {notice, dislike, hate} — what an encoder picker actually wants: "minimize bytes subject to q ≥ q_threshold(c, k)".

Status

Code	Rust+axum backend, Vite+vanilla-TS frontend embedded via rust-embed, SQLite via sqlx, deploy via Railway + Docker.
Live	https://squintly.imazen.org — all four flows work: rating (180 imazen26-v4-test sources (imazen-26 TEST split only) across 21 strata / 4320 encodings across libjpeg-turbo, jpegli, libwebp, libavif, served straight from public R2 — see DEPLOY.md §15), curator (903 R2 candidates), calibration, suggest. Confirm which build you are looking at with curl .../api/export/pareto.manifest.json.
Study	v0.2 protocol pre-registered in docs/STUDY.md — H1–H5 hypotheses, holdout discipline, decision rules for zensim v48 ship locked.
Jargon	New to IQA? Start with docs/GLOSSARY.md — every term, unit, method, and stat in this README, in plain English, with links to the underlying papers.
Phase	v0.2 finishing — ASAP active sampler wired into next_trial (§3 amplifier #1). Playwright e2e is green across 4 device projects incl. Z Fold 7 cover/inner (117 tests, with a per-screen layout-viewport guard in e2e/layout.spec.ts), plus a scripted demo-user audit (just audit) screenshotting every screen. Remaining: the same walkthrough on a physical phone + nightly per-observer η batch (#9).
Pilot	Planned: 100 sessions once v0.2 ships (§8 phases).

---

1. Quick start (local)

# Build the embedded frontend + the binary.
just build

# 1. Build a real corpus from imazen-26 and serve trials from it
#    (~10 min; 32 sources across all four size buckets, 4 real codecs).
#    See DEPLOY.md §15 for what the selection guarantees.
python3 scripts/build_demo_corpus.py --out ~/tmp/squintly-corpus --clean
cargo run -- --coefficient-path ~/tmp/squintly-corpus --bind 127.0.0.1:3030

# 2. Or point at any other coefficient SplitStore / HTTP viewer:
cargo run -- --coefficient-path /path/to/coefficient/benchmark-results --port 3030
cargo run -- --coefficient-http http://localhost:8081 --port 3030

# 3. Open http://localhost:3030 in any browser, on any device.
#    Calibrate (credit-card slider), pass the qualifier, rate trials.

# 4. Export when done (all TSV in zenanalyze schema):
curl http://localhost:3030/api/export/pareto.tsv      > pareto_human.tsv
curl http://localhost:3030/api/export/thresholds.tsv  > thresholds.tsv
curl http://localhost:3030/api/export/responses.tsv   > responses_raw.tsv
curl http://localhost:3030/api/export/unified.tsv     > unified.tsv

# 5. Smoke test (gates v0.2 done):
cargo test --test smoke

Deploy to Railway: see DEPLOY.md. Dockerfile builds the embedded TS, then the Rust binary, then ships a debian:bookworm-slim runtime — no Node at runtime.

---

2. The study

The study is pre-registered in docs/STUDY.md. One-paragraph summary:

Train zensim-v48 on squintly's (stimulus, viewing_condition_bin) JOD labels and see whether it beats v47-strict (the current production bake) on held-out squintly data by ≥ 0.05 SROCC, with the gain concentrated in the high-DPR / short-viewing- distance bins, without regressing on CID22 / AIC-3 / AIC-4. Threshold function q_threshold(c, k) must achieve R² ≥ 0.5 on held-out for at least k = notice. Decision rule for shipping v48 to zensim::Profile::A is locked in §3 of STUDY.md.

Methodology references: docs/methodology.md (squintly's internal methods), docs/participant-grading.md, docs/motivation-and-compensation.md. For the broader reproduction-grade methods reference (BT-MLE math, active-sampling info criteria, software inventory) see the zenpapers reference book.

---

3. How we amplify human effort — 15 levers

Ranked by ROI for converged-cells-per-human-hour. Status legend: ✅ done · 🟡 partial (impl exists, not fully wired) · 🔵 planned · ⚫ explicitly out of scope.

#	Lever	Status	Where	Plan
1	Active sampling (ASAP EIG) — pick next pair by expected information gain	✅	src/asap.rs, src/sampling.rs (select_pair_with_eig), src/handlers.rs (enhance_pair_with_asap), tests/asap_wire.rs	Wired into next_trial for non-golden pairs: on each pair request, refit BT-Davidson (σ_prior=1.0) over historical pair responses for the source and pick the highest-EIG adjacent non-trivial pair. Falls back to the random adjacent pair when comparisons < ASAP_MIN_OBS=8. Validate the ≥ 5× speedup-vs-random H4 prediction in pilot.
2	Viewing-conditions captured per trial (DPR, intrinsic-to-device, calibrated CSS-px/mm, ambient, distance, gamut)	✅	migrations/0001_init.sql, web/src/conditions.ts, web/src/calibration.ts	Display is a hard 1:1 device pixels minimum — never scaled down; anything larger is panned (docs/methodology.md §2). intrinsic_to_device_ratio is 1.0 by construction, and pan_*/visible_* (migration 0012) record how much was seen.
3	Staircase thresholds (Levitt 1971 transformed up–down, 79.4 / 70.7 / 50 % for notice / dislike / hate)	✅	src/staircase.rs (233 LOC)	tune step-halving cadence after pilot
4	Trivial-triplet filter — pre-filter pairs the metric ensemble unanimously predicts (≥ 95 % preference); humans never see them	✅	docs/methodology.md §3.4, src/sampling.rs	verify firing rate in pilot
5	Anchor reservation — CID22-style reference-pinned anchors at ~5 % rate align our scale across studies	✅	migrations/0006_v02_rigor.sql (corpus_anchors), docs/methodology.md §3.6	populate the table from CID22 source overlap
6	Onboarding calibration — codec-probe + credit-card slider for physical CSS-px/mm	✅	web/src/calibration-onboarding.ts, web/src/codec-probe.ts, migrations/0004_codec_support.sql	—
7	Honeypots — known-answer trials at ~1/30 rate; fail-two-end-session	✅	docs/methodology.md §3.9	tune fail rate to ≤ 5 % on T1 observers
8	Per-session grading A–F + session_weight ∈ {0, 0.5, 1.0, 1.5} multiplier at scale reconstruction	✅	src/grading.rs, migrations/0002_grading.sql	—
9	Per-observer reliability — nightly observer_grades aggregation (trial-weighted golden, session-count-weighted even-odd, geometric-mean composite weight, trusted-pool promotion at weight ≥ 0.70 ∧ n_trials ≥ 50); Crowd-BT η + pwcmp LOO log-likelihood queued	🟡	src/grading.rs::rebuild_observer_grades, hooked into a 24h tokio task at src/main.rs	implement the pwcmp LOO (dist_L > 1.5) + Pérez-Ortiz 2019 (δ_o, σ_o) ACR fits post-pilot once we have ≥ 50 sessions/observer signal
10	Engagement / retention — streaks, freezes, tier ladder (anon → email → passkey → researcher), leaderboard	✅	migrations/0003_engagement.sql, migrations/0005_auth.sql, web/src/auth-modal.ts	iterate UX based on drop-off telemetry
11	Phone-first single-stimulus + tap interaction — short attention sessions, no mouse, portrait orientation	✅	SPEC.md §"Target audience: phones", Playwright zfold7-cover/inner device profiles	—
12	Unified pairwise + ACR-rating scale (Pérez-Ortiz 2019 mixed BT-Davidson + ordinal model)	✅	src/bt.rs, src/unified.rs (solver fixed 2026-05-28 — see CLAUDE.md resolved-bug log), docs/methodology.md §5.5	fit_unified converges on consistent pair+rating evidence (regression-tested via unified_competitive_with_bt_only_on_heldout_pairs). pair_log_likelihood / rating_log_likelihood / total_log_likelihood shipped as the H3 evaluation entry point. The 2-nats/trial H3 gate itself is the pilot's job, not a unit test.
13	Predictive sampling (PS-PC) — offline classifier marks pair candidates predict vs defer using a metric ensemble (CVVDP / butteraugli / ssim2-gpu / zensim / PaQ-2-PiQ); humans only see defer	🔵	not yet	v0.3 add: classifier runs as a one-shot pre-flight, populates defer queue in sampling.rs. Expected: 8–22 % defer ratio at η ∈ {0.97, 0.995} (Mohammadi 2311.03850).
14	Boosted Triplet Comparison (BTC) — quadratic boosting h(d) = γ₁d + γ₂d² for high-fidelity near-JND sensitivity	🔵	not yet	v0.3 optional add for the q ≥ 80 band; expected ~3× discrimination (AIC-3 2410.09501)
15	Hybrid expert + crowd — route highest-EIG pairs to expert raters	⚫	not in scope for v1 (squintly is anonymous by design); zensim-level hybrid possible later via a separate study

The first 11 are done; #1 (ASAP→next_trial) and #9 (nightly observer_grades batch) both shipped recently. The remaining v0.2 work is the end-to-end smoke test on a real phone plus the deeper #9 follow-up (pwcmp LOO + Pérez-Ortiz 2019 per-observer ACR fits) once we have ≥ 50 sessions/observer to fit on.

---

4. Workflows

                         ┌────────────────────────────────────────┐
                         │ coefficient (image store)              │
                         │   HTTP or local SplitStore             │
                         └──────────────┬─────────────────────────┘
                                        │ src/coefficient.rs (consumer-only)
            ┌───────────────────────────┴────────────────────────────────┐
            │                                                            │
            ▼                                                            ▼
┌──────────────────────┐                                ┌──────────────────────────┐
│ Curator workflow     │                                │ Trial workflow           │
│ /api/curator/*       │                                │ /api/session, /api/trial │
│ src/curator.rs       │                                │ src/handlers.rs          │
│ Stream → Curate →    │                                │   sampler ─▶ next_trial  │
│ Threshold → Export   │                                │   response → grading →   │
│ → curated_manifest   │                                │   weighted BT-Davidson   │
└──────────┬───────────┘                                └──────────┬───────────────┘
           │ TSV upload                                            │ inserts
           ▼                                                       ▼
┌──────────────────────────────────────────────────────────────────────────┐
│ SQLite (sqlx) — single canonical store                                   │
│   observers · sessions · trials · responses · staircases · corpus_anchors│
│   curator_candidates · curator_decisions · curator_size_variants ·       │
│   curator_thresholds · suggestions · calibration_pool                    │
└──────────────────────────────────────┬───────────────────────────────────┘
                                       │
              ┌────────────────────────┼────────────────────────┐
              ▼                        ▼                        ▼
     ┌─────────────────┐    ┌─────────────────────┐   ┌──────────────────────┐
     │ Online (per     │    │ Nightly batch       │   │ Export TSVs           │
     │ session)        │    │ grading + Crowd-BT η│   │ pareto / thresholds /  │
     │ staircase + BT  │    │ (planned, §3 #8/#9) │   │ responses / unified    │
     │ online estimate │    └─────────────────────┘   │ → zenanalyze/zentrain  │
     └─────────────────┘                              └──────────────────────┘

Workflows in detail:

• Curator (docs/CORPUS_CURATOR_SPEC.md) — researchers iterate the corpus offline: stream candidates from R2/local, curate (include/exclude
  • size bucket), set per-source threshold q ∈ {30, 50, 70, 85, 95}, export curated_manifest.tsv. Frontend: web/src/curator.ts + curator-encoder.ts. Backend: src/curator.rs + src/variant_gen.rs.
• Suggestion — anyone can submit a content URL via /api/suggestions; researcher triages via /suggestions/{id}/{accept,reject}. Backend: src/suggestions.rs, src/suggestion_store.rs. Migration: 0008_suggestions.sql.
• Trial — observer arrives, opt-in calibration, qualifier (8 trials → ≥ 6 correct → qualifier_passed), main session of mixed Type-S + Type-P (sampler-chosen) with honeypots interleaved. Frontend: web/src/trial.ts. Backend: src/handlers.rs, src/sampling.rs, src/asap.rs (post v0.2-wiring).
• Auth — passwordless magic-link via Postmark for T1+ observers (src/auth.rs, migrations/0005_auth.sql). Observer aliases across devices (observer_aliases).
• Grading — per-session: golden-pass rate, straight-line max, honeypot failures → session_grade A..F → session_weight ∈ {0, 0.5, 1.0, 1.5}. Per-observer: trusted pool, skill score, per-observer reliability η (planned).
• Export — four TSVs in zenanalyze/zentrain schema: pareto.tsv (BT-Davidson per-encoding scores), thresholds.tsv (q_notice/dislike/hate per-source-per-condition), responses.tsv (raw per-trial), unified.tsv (joint BT + ACR fit).

---

5. Data storage

5.1 Canonical paths

What	Where	Format	Why there
SQLite DB	${SQUINTLY_DB:-/data/squintly.db}	sqlx-managed; 9 migrations applied at startup	Single source of truth for human responses; backups + Tower mirror per CLAUDE.md ML rules
Source images	coefficient store (HTTP or local SplitStore)	content-addressed by sha256	Squintly never writes back to coefficient; consumer-only
Variant encodings	served via /api/proxy/encoding/{id}; backed by coefficient	content-addressed	reuses coefficient's content-addressing
Curated manifest snapshots	uploaded R2 JSONL at pub-7c5c57fd3e0842f0b147946928891d40.r2.dev	one JSONL per snapshot	reproducible curator inputs
Export TSVs	downloaded by user; canonical location is whoever downloaded last	TSV (zenanalyze schema)	_MANIFEST.json per export, includes study_commit + db_snapshot_sha256 (planned)
Auth tokens	auth_tokens table	Postmark magic-link tokens, ephemeral	TTL'd; cleaned by cleanup_expired_tokens
Tower mirror	/mnt/tower/output/squintly-archive-<study-version>/	full SQLite snapshot + image manifest	per CLAUDE.md "Mirror canonical data to Tower NAS BEFORE any cleanup"

5.2 SQLite schema (current — 9 migrations)

Table	Purpose	Migration
observers	UUID + tier + calibration + email + skill_score + grading flags	0001, 0002, 0003, 0006
sessions	viewing-conditions snapshot, grade, weight, golden-pass-rate, codec-support	0001, 0002, 0004, 0006
trials	individual trial: source_hash, encoding id, condition vector, type S/P, staircase id	0001
responses	observer answer + dwell + swap_count + zoom + viewport snapshot at response time	0001
staircases	per-(source, condition_bucket) staircase state (reversals, current q, exit)	0001
corpus_anchors	CID22-style reference-pinned anchors for cross-study scale alignment	0006
source_flags	curator decisions on excluded/flagged sources	0006
calibration_pool + calibration_responses	shared calibration trials whose answers we know	0006
auth_tokens + observer_aliases	passwordless auth + multi-device merging	0005
curator_candidates / _decisions / _size_variants / _thresholds	curator workflow state	0007, 0009
suggestions	user-submitted content with workflow status	0008

DDL in migrations/. Schema is the source of truth for fields; this table is a navigation aid.

---

6. Statistical methods

6.1 Scale reconstruction (online + offline)

• Online: per session, BT-Davidson online estimate of θ (the BT score) via src/bt.rs::fit_session(); cheap, only over the current session's trial pool.
• Offline: full unified Pérez-Ortiz 2019 mixed BT-Davidson + ordinal-ACR with per-observer (δ_o, σ_o); cross-source comparability via corpus_anchors. Gaussian prior σ_prior = 1.0 (pwcmp standalone, arXiv 1712.03686; now in zenpapers corpus).

6.2 Threshold inference

For k ∈ {notice, dislike, hate}:

• Online: per-session staircase reversal-average estimate of q_k per (source, condition_bucket).
• Offline: GAM logistic P(rating ≥ k | q, c) = Φ((q − μ_k(c)) / σ_k(c)) with smooths on dpr, intrinsic_to_device_ratio, viewing_distance_cm, codec offsets, per-source random intercept. H2 in docs/STUDY.md pre-registers R² ≥ 0.5 for k = notice on held-out.

6.3 Per-observer reliability (Crowd-BT η)

• Hooks in src/grading.rs; full η + pwcmp leave-one-out per-observer log-likelihood is the TODO at src/grading.rs:340–343. Plan: nightly batch reading the last 30 days of trials, writing observers.skill_score + a per-observer η.

6.4 Confidence intervals

• Bootstrap over observers (1000 resamples) for θ and q_k. The observer is the unit of resampling because errors are correlated within observer (BT.500 + Pérez-Ortiz).
• Laplace / Hessian-inverse for active-sampling EIG (cheap, online); switch to bootstrap for the final reported CIs.

6.5 Significance tests

For metric-vs-metric (v48 vs v47):

• Δ-SROCC with bootstrap 95 % CIs over observers.
• Krasula different-vs-similar AUC.
• F-test on residual variance for nested model comparisons.
• Bonferroni across the 5 pre-registered hypotheses (α_individual = 0.05 / 5 = 0.01).

Full math + citations: docs/methodology.md §5–7 + the zenpapers reference book ch3-5_sampling_screening_cis.md.

---

7. Data sprawl — concerns + the rules we follow

ML projects accumulate sprawl faster than any other code; we apply the discipline from the global CLAUDE.md "ML Data Pipeline Discipline" section verbatim:

1. One canonical DB. ${SQUINTLY_DB} is the only store of human responses. No shadow exports, no "fast-iter" copies in /tmp. Backups via SQLite VACUUM INTO nightly to /data/snapshots/squintly-<iso>.db.
2. Every export carries build_commit. /api/export/{kind}.tsv ships a sibling /api/export/{kind}.manifest.json carrying build_commit (the git SHA the binary was built from, via build.rs), schema_version, exported_at, row_count, sha256 of the TSV body, and a source_query pointer at the Rust source that produced the rows. The TSV response also carries a Link: <…>; rel="describedby" header pointing at the sidecar. Without this, "is this export still valid?" becomes a forensic audit.
3. No source-image duplication. Squintly consumes the coefficient store; never copies images, never writes back. The R2 manifest at the public bucket is the canonical curator input.
4. Tower mirror BEFORE cleanup. Per CLAUDE.md: before deleting any local DB snapshot, sha256-verify a Tower mirror copy. The snapshot directory grows; at ~10 snapshots a one-time consolidate-to-Tower run is fine; never deletes without the mirror.
5. Block storage for anything > 30 KB. Curator manifests, exported TSVs, image archives all live under /mnt/v/output/squintly/<study-version>/ with a pointer file in this repo. Never commit images, never commit the DB.
6. Cleanup of stale agent worktrees. If multi-agent work creates sibling jj workspaces, jj workspace forget + rm -rf same-day per CLAUDE.md.
7. Dated experiment dirs OK ≤ 7 days; promote or archive by day 14. Any docs/explorations/<date>-<topic>/ either gets promoted into the main docs/ tree or moved to _archive/.

What we don't yet enforce automatically (tracking these as v0.2 exit gates):

• _MANIFEST.json emitted alongside every export TSV (/api/export/{kind}.manifest.json paired with the TSV; describedby Link header).
• Nightly Tower-mirror snapshot — main.rs auto-detects /mnt/tower and spawns a daily VACUUM INTO /mnt/tower/output/squintly-archive/<iso>.db task. Silently no-ops on Railway / CI / remote deploys where the mount is absent.
• Per-table row counts in a db_health table updated hourly (src/db_health.rs::refresh + migrations/0010_db_health.sql; hourly tokio task in main.rs).
• Curator R2 snapshot tag pinned in each session — manifest_snapshots table (migrations/0011_manifest_snapshots.sql) records (r2_public_base, manifest_path, manifest_sha256, body_bytes, n_candidates) per load; UNIQUE key on the three identity columns means re-loading an unchanged manifest is a cheap no-op. sessions.manifest_snapshot_id FK pins the latest snapshot at session creation so analysis six months later can join from any session to the exact candidate pool it drew from.

---

8. Phases

Locked sequence; the order matters because each phase depends on the previous one's data:

Phase	Goal	Done-when	Status
v0.1	Backend skeleton + trial routes wired; curator + variant generation; license surfacing	cargo test green + curator export round-trips	✅
v0.2	ASAP wired into next_trial; per-session grading active; export _MANIFEST.json; smoke test exercises the full session loop on a real phone	All 5 of SPEC.md §"Hard v0.1 is done when" + the 4 sprawl-enforcement checkboxes above	🟡 in progress
v0.3 — pilot	100 sessions across the existing curator corpus; H4 (active-sampling speedup) + H5 (phone-vs-desktop noise) measured; Crowd-BT η baseline	All five hypotheses' analysis pipelines green on pilot data	🔵
v1.0 — main	1500 sessions; H1–H3 evaluated; v48-cond zensim bake decision	The decision rule in docs/STUDY.md §3 executes — v48 ships or an honest-stop documenting the next hypothesis	🔵
v1.1+	PS-PC predict-fill turned on; BTC boosting for high-fidelity; per-observer η; mobile-specific (Galaxy Z Fold) sub-study	gated on v1.0 finding	🔵

Full study design + decision rules: docs/STUDY.md.

---

9. Contributing / open decisions

The amplifier table in §3 is the authoritative checklist of what's done vs planned. The next-most-impactful tractable chunks:

1. End-to-end smoke on a real phone (v0.2 exit gate). The Playwright e2e suite runs locally (just e2e; the repo has no CI workflow yet) but a human-on-Galaxy-Fold-7 walkthrough catches everything the headless browser misses.
2. pwcmp LOO + Pérez-Ortiz 2019 per-observer ACR fits (amplifier #9 deep dive). The nightly observer_grades aggregation already fills the trial-weighted golden pass rate, even-odd consistency, geometric-mean weight, and trusted-pool promotion; the LOO and per-observer ACR fits fill the pwcmp_* and sigma_acr / delta_acr columns once we have ≥ 50 sessions per observer.
3. Tower-mirror cron + db_health — the two remaining data-sprawl v0.2 exit gates.

Open decisions worth surfacing:

• PS-PC at v0.3 or v1.0? The metric ensemble (CVVDP / butteraugli / ssim2 / zensim / PaQ-2-PiQ) isn't yet wired into squintly's pair-selection. v0.3 is the natural slot.
• Anchor source. CID22 is the obvious anchor, but we don't yet have a content-overlap map; need to derive which of CID22's 250 originals overlap our curator corpus (zero, probably — different content pool — in which case we anchor via a controlled stimuli set instead).
• Recruitment. Are we self-recruiting friends + colleagues for the pilot, or going Prolific from day 1? Prolific has the screener depth but anonymous T0 observers should be a viable path on our own site.
• Galaxy Z Fold 7 sub-study. The Playwright device profiles exist; running an actual cohort on foldables would be a separate paper.

---

10. Deploy / ops

Production deploy uses Railway + a 1 GB volume mounted at /data. See DEPLOY.md for the full walkthrough.

Monitoring: stdout JSON logs (Railway captures them); per-endpoint latency in tracing spans. No third-party telemetry.

---

11. License & attribution

• Code: Apache-2.0 OR MIT (dual).
• Per-image content: runs under each source corpus's license (per src/licensing.rs registry — 7 policies surfaced on the welcome screen + per-trial badge). See docs/motivation-and-compensation.md.
• Derived score tables (pareto.tsv, thresholds.tsv, unified.tsv): planned for CC-BY-4.0 release after study close, conditional on consent posture (see docs/STUDY.md §8).

---

See also

• SPEC.md — full v0.1 spec (data model, schema, threshold/BT math).
• docs/STUDY.md — formal pre-registration: research questions, hypotheses, design, analysis plan, decision rules.
• docs/GLOSSARY.md — layperson glossary of every term, unit, method, and statistic this README and STUDY.md use, with hyperlinks to all the underlying research papers, datasets, and tools.
• docs/methodology.md — methods reference internal to squintly.
• docs/participant-grading.md — session-grade + weight policy.
• docs/motivation-and-compensation.md — recruitment economics + ethics posture.
• docs/CORPUS_CURATOR_SPEC.md — curator workflow.
• docs/HANDOFF.md — agent / contributor handoff notes.
• External: zensim (what we feed), zenanalyze (TSV consumer), coefficient (image store we consume), zenpapers reference book (methodology citations).