Does it reproduce the books it cites?

1. Worked-example suite (runs on every commit; red blocks merge)

Second reads: examples 08–12 are second-pass checked (2026-09-21) against PyMuPDF 200 dpi renders of the owner's Turton 5th ed., read independently of the first pass's pypdf text layer — every statement input, index value, exponent, factor, printed intermediate and total matches (0 corrected, 0 unreadable; each file carries a checked block, rows in docs/transcription-log.md). The two known failures (11 and 12) stand: the render confirms the book substitutes (P + 1) into Eq. 7.10 while its text defines P in barg. Examples 01–07 and the TEA/back-test cases keep their earlier status.

Intermediate-value checks (packages/library/test/textbook-checks.test.ts): Turton Example 7.11 pressure factor at 100 barg = 1.383 (reproduced exactly from the transcribed Table A.2 coefficients); Example 7.10 Cp0 = $25,328 by equation vs $25,000 read from the chart by the book (1.3 %, chart rounding).

TEA worked examples (Phase 4, FR-28/FR-31):

All three known failures are recorded with the reason quantified in the example file, not silently accepted: tea-01 misses the published LCOH by −4.7 % (the OpenPyTEA case and computeTea diverge on a modelling convention the file names); tea-02, once npv_reference: "project_start" matches Turton's own discounting zero point (start of construction), still misses the book's NPV by +42.8 % — a residual fully accounted for by three quantified treatment differences the book's cash flows include and TeaBasis cannot represent: the $10 MM land purchase and its recovery, a loss-year tax credit the book takes in year 2, and the $10 MM salvage value the book taxes as income where the engine adds it untaxed. tea-03 misses T&S Example 9.4 by −5.3 % for one reason only — tax paid a year in arrears — and the file re-discounts the book's tax stream to recover the printed $146.40 MM exactly. All three run as ordinary tests, not skips — see "known failures" below.

Known failures, made honest (engine_anchor): every known_failure worked example (seven total: 04-turton-example-7-11, 11-turton-example-7-13 and 12-turton-example-7-14 above, bt-02-usda-dry-mill-ethanol in §3, tea-01, tea-02, tea-03) now carries engine_anchor: { value, tolerance_pct: 0.01 } — the engine's own value on 2026-09-21, not a published figure — and runs as a plain test asserting two things: the result stays outside the published/target band (so the tool never silently "fixes" the known gap) and the result stays within 0.01 % of the anchor (so a change to the miss itself is caught, not just presence of a miss). Anchors: 04-turton-example-7-11 direct 135724.2481170268; 11-turton-example-7-13 direct 7194921.146171735; 12-turton-example-7-14 direct 1056759.5880168518; bt-02-usda-dry-mill-ethanol total 140033875.12125257; tea-01-openpytea-cs1-smr tea.lcop 3.7324296969383304; tea-02-turton-ch10 tea.npv 24445572.894993633; tea-03-ts3-example-9-4 tea.npv 138683835.8273631. Before this, a known failure ran as it.fails — it caught a miss appearing or disappearing but not a miss changing size; the anchor closes that gap.

Textbook suite: 12 against a target of 10 (cases 01–12 above; nine within ±1 %, three — 04, 11, 12 — quantified known failures with engine anchors). Turton Chapter 7 read in full from the owner's copy (2026-09-20): 7.10 and 7.12 share 7.11's chart-read C_p^0 (7.10 would miss by +1.2 %; 7.12 also needs the Fig. A.18 stainless F_M = 2.73, not held); 7.13 and 7.14 added 2026-09-21 as cases 11 and 12 once Eq. 7.10/A.2 was transcribed and the engine's turton_vessel_a2 form took the diameter from the line's sizing step — both are known failures for one reason, the book's (P + 1) substitution into the vessel pressure factor (the equation's text defines P in barg; the substitution recovers each printed total to +0.03 %); their chart-read factors (SS vessel F_M 3.11, SS-tube exchanger F_M 1.81, CS pump F_M 1.55, SS tray F_BM 1.83) are the examples' own readings and ride in library_overlay copies of the records, never in the library; 7.15 is a Table 7.13 capacity-scaling case, not a Class 4 example. T&S Example 7.3 added 2026-09-21 as case 05 (shell masses entered as the book prints them; see the row). PTW Examples 6-3 and 6-5 added 2026-09-21 as cases 06 and 07 from the owner's copy (6-5's Eq. 6-10 results are outside the engine's Class 5 chain — see the row). Turton Examples 7.6(c), 7.7 and 7.8 added 2026-09-21 as cases 08–10 from the owner's copy (PDF pp.214 and 216), each with the book's Table 7.4 index values in index_snapshot so the escalation ratio is the book's; 7.10 was not added — it would be a second copy of 04's chart-read gap (same C_p^0, F_P = 1) and adds no coverage 04 and the intermediate-value check do not already give. T&S Example 9.5 (adipic acid, 30/70 construction, 50 % first-year rate, tax in arrears, worksheet printed as a figure) is not fixtured: the engine has no production ramp, and Fig. 9.2's cash flows are not in the PDF text layer.

3. Historical back-test (Sprint 10)

Three of the five cases below (#1 hydrogen, #4 ethylene, #5 HDPE) ship as packages/library/samples/history.sample.json and can be loaded on the Knowledge base tab as a first, ready-made knowledge base.

No client project data is used. Each case is a plant whose capacity and all-in cost are published by an independent public source (a US government study, a USDA survey, or the owner's own announcement), recomputed in Class 5 capacity mode from the T&S Table 7.1 correlations with the tool's out-of-the-box allowances (T&S Table 7.5 fluids OSBL 30 %, D&E 30 %, contingency 10 %; owner's cost 0 %) and CEPCI-escalated to the published cost date. Pass criterion: the Class 5 band, −50 % / +100 % of the published figure. Files: packages/library/worked-examples/bt-0N-*.example.json; they run on every commit.

Four of five land inside the Class 5 band, spread −15 % to +12 % around zero (mean −2 %) on the fluids plants. The first run of this back-test used the Con Edison 35 % as the Class 5 contingency default and overshot every case by +5 % to +37 %; the default was changed to T&S Table 7.5's 10 % (the point-estimate contingency that belongs with the Table 7.1 correlations), with the 35–40 % retained as the upper end of the benchmark range and the −50/+100 % band carrying the class uncertainty. Cases 3–5 are brownfield projects on integrated sites (shared utilities, hydrogen, ethylene supply), which explains the residual negative errors on 4 and 5. Case 2 is recorded as a known failure (known_failure in the file; the suite asserts it stays outside so a silent change is noticed): the ISBL correlation alone is ≈ +27 % over a design-build dry-mill package, and fluids OSBL/D&E/contingency on top nearly doubles it. It is not tuned away — the lesson for a solids/biofuels plant is to use lower OSBL and engineering allowances than the fluids defaults, which the estimator can do on the Factors page and record in the assumptions register.

Scope caveats are stated in each file's notes.assumptions: NETL's TOC includes owner's costs (ours are 0 %); the Shintech figure also covers VCM/PVC tie-ins; cost dates for multi-year builds are taken at the construction mid-point. These are synthetic in the sense that no project is the user's own, but every number traces to a public document.

4. Known limitations that affect validation

• Vessels and columns — T&S correlations take shell mass; Turton's take volume and, since 2026-09-21, the Eq. A.2 pressure factor (turton_vessel_a2) reads the diameter the user entered on the line's vessel_volume / column_shell_volume sizing step — still no mechanical design (§1): a pressurised Turton vessel with no sizing diameter refuses to price rather than silently taking F_P = 1. Turton's own Examples 7.13/7.14 substitute P + 1 (absolute bar) into an equation whose text defines P in barg; the engine follows the equation's text, so cases 11 and 12 miss by −3.9 % / −2.9 %. Decided 2026-09-21: P stays the vessel design pressure, gauge. Eq. A.2 / 7.10 is the ASME BPVC VIII-1 UG-27 thickness form, whose P is the internal design pressure (gauge) — the book's text says so on both pages — and the two examples' P + 1 is the book's own substitution, not the method; no P_offset_bar is added and cases 11 and 12 remain documented misses with engine anchors. Turton's stainless F_M (Fig. A.18) and tray F_BM (Fig. A.9) remain graphical and unheld.
• Class 5 process correlations are Jan 2006 (T&S Table 7.1) and 2000 (PTW Table 6-11). Escalating 17–23 years by CEPCI alone ignores real-terms cost inflation in process plant construction since 2006; totals should be read as indicative. The back-test (§3) shows a consistent overshoot rather than undershoot on fluids plants once default allowances are applied.
• Location factors: the library ships US = 1.0 only — any other location comes from a company location set or a noted override; the ≈1997 table of unidentified authorship was withdrawn 2026-09-21 (values preserved in the transcription log). CEPCI 2017–2023 are single-secondary-source (1987–2001 checked against PTW5 Table 6-2 and 1996–2016 except 2009 against Turton Table 7.4 from the owner's copies). Both are flagged in the transcription log.
• Second-pass sign-off — done 2026-09-21 for every Turton 5th ed., Towler & Sinnott 3rd ed. and PTW5 correlation, factor and Class 5 record against the owner's copies (checked 407 / unchecked 4 on pnpm validate:library --release; 4 float-residue corrections in T&S Table 7.1, none elsewhere — docs/transcription-log.md TUR5, TS3 and PTW5 sections). Still pending: the Butene-1 BP pair (one reading), the classes.json class:4 and class:3 rows (bands unchecked, indirect defaults unsourced on all three classes), the IRS MACRS tables (transcribed from recall), and, as a caution rather than a gap, the Table 19.1 U-value and Ch. 8 opex rows (transcribed and marked checked on the same pass from a single reading, with the task reviewer's spot-check of every row against the PDF; an independent second reader should re-read them before a tagged release).