Shrinkage — the number you can't measure, and the fit it doesn't fix

Build spec — Shrinkage & fit calculators (shrink-only page + lock-taper pocket)

Goal. Two single-file HTML tools and one CLI around a single honest claim: a 40 mm printed cube cannot resolve the shrinkage it is meant to measure, so a shrink calculator must be allowed to answer “no change, leave the slicer at 100 %”. The second tool answers the question that shrinkage cannot — how to make a part go in slowly and smoothly instead of dropping in or needing a mallet.

Hard requirements

• Zero network requests in both pages: no CDN, no webfonts, no raster assets, no fetch. Must run from

file://, from a USB stick, and inside a portfolio iframe. (Verified: 0 matches for src=/href= to anything external in the shrink page.)

• Shrink page does one job. Preset material, nominal size, three reading boxes, and four outputs

(shrink %, slicer scale %, measured mm, off nominal mm) plus a per-axis table. No pocket geometry, no lead-in controls, no diagrams. The fit tool stays separate so neither page compromises the other.

• Multiple readings per axis, accepted space/comma/newline separated. Per-axis datum is the

minimum (a tilted caliper only ever reads big); measured size is the mean of the axis minima. Spread is reported separately so noise never hides inside an average.

• A noise floor derived from the user's own data: max(0.05, max_spread/2 + 0.02) mm. A deviation

at or below it must produce 100.00 % scale, with the raw arithmetic shown greyed underneath as evidence — never as the instruction.

• Verdict precedence, fixed and testable: preset → nothing parseable → under noise → oversized →

real shrink. Coverage warnings (“only one axis”) are notes, never the verdict, or they mask the oversized branch.

• Never emit a dishonest number: no NaN, no Infinity, no 0.000 % from junk input, and no

sub-100 % scale offered as the fix for a part that printed oversized.

• Lock fit: gap ramps from mouth to floor instead of being constant. Defaults

−0.04 mm/side land clearance, 0.20 mm/side mouth opening, taper length equal to pocket depth. Report free travel, bite-zone length and bite rate (µm per mm of push); warn below 1.5 mm of bite zone, above 0.09 mm/side interference, and when the mouth is so wide the first millimetres fall in loose.

• Every number on screen is re-derivable in one line of arithmetic shown on the analysis page.
• CLI parity: dt filament shrink <nominal> <reading…> prints shrink only (--pocket opts into geometry);

dt filament fit <envelope> --fit lock --clearance -0.04 --hole-bias 0 --lead taper --lead-allow 0.20 --taper-len 20 prints the pocket and the insertion ramp. --json on both.

Verification contract

• Headless Chrome at 390 / 768 / 1440 px, narrow widths via iframe (Chrome clamps a headless window to

~500 px, so a “390 px” screenshot at window width is a lie).

• Assertions: document.documentElement.scrollWidth <= innerWidth, min tap target ≥ 44 px, table

overflow = 0, no clipped SVG text (via getBBox()), 0 console errors.

• Every displayed value recomputed independently in Python and compared to the rendered DOM — 9 maths

checks, 8 UI states (empty, measured, real shrink, oversized, one axis, junk, zero nominal, after reset).

• Both harnesses live in ~/.hermes/scripts/ (verify-shrink.py, verify-shrinkfit.py) and exit non-zero

on any failed assertion.

Pitfalls worth carrying forward

• A scale factor is not a fit knob. It multiplies the gap along with the parts: scaling both to

100.3 % turns a 0.16 mm mouth gap into 0.1605 mm. Shrink correction preserves fits by design; only pocket mouth/floor/taper change how entry feels.

• 40 mm is too short a witness. One caliper tick (0.01 mm) is 0.025 % on 40 mm — half the effect.

Use a 100 mm+ bar before believing any material constant; the preset table exists for the rest.

• Elephant's foot is not shrinkage. A 40.09 mm bottom against a 40.05 mm mid-face is a first-layer

flare (+0.02 mm/side). Use the mid-face minima for shrink, the bottom maximum for entry clearance.

• dt must be a symlink, not a copy — a copied launcher resolves DT_DIR to /usr/local/bin and

every sub-app dies with exit 2. Fixed with resolve_root().

• The stiffest bug found in testing was a correct-looking one: the single-axis warning was promoted to

the verdict, which silently hid the oversized message. Test states in pairs, not one at a time.

Files