Tabs 2 and 3 each opened with a claim; tab 1 opened with instructions, so the
proposition the other two tabs attack was never stated. Tab 1 now opens with "A
working model that explains itself", making the sequence assert, undermine, what
survives.
Restores the horizontal rule before the conformal section -- present until v3.2,
where it sat where the chart needed to go and was not put back. After a figure, a
table and a blockquote, an h3 alone does not read as a section boundary.
Tab 3 label drops the conditional: "What We'd Show an Auditor" -> "What We Show an
Auditor".
Mirrors llm-wiki commit 18dcbd6.
Make tab 3 reachable on mobile; restyle the tab bar; rename tabs
Gradio's .tab-container is flex-wrap:nowrap with overflow-x:hidden. At 375px tab 3
extends to 544px with no scroll, swipe or arrow to reach it -- a phone visitor could
not open the conformal risk control tab at all. Measured on the live DOM.
overflow-x:auto plus a narrow-viewport media query fixes it at any label length.
The active tab also differed from the inactive ones by colour alone: identical
font-weight, no fill, no border. Now three independent signals -- weight, fill, and a
real 3px underline. Selectors target .tab-container and [aria-selected] rather than the
svelte-* hash, which is regenerated on SDK rebuilds.
Renames: "Can it be fooled?" -> "Can We Fool the Model?"; "What can you actually
prove?" -> "What We'd Show an Auditor" -- the old title promised proof, the one word
that tab spends its length disclaiming. Numbers kept; the intro asserts the tabs are an
argument read in order.
Mirrors llm-wiki commit 0731cb9.
Move the conformal chart above its interpretation; fix two Tab 3 errors
The limits section said "It is why the chart above is load-bearing rather than
decorative" while the chart rendered below the whole markdown block -- the
reference pointed at nothing.
Moving the image up fixes it and puts the chart where it earns its keep: the
paragraph before it ("validated, not just asserted -- six of nine, exceedances of
1.12/0.55/0.83 SE") describes this exact figure, which a reader previously did not
reach until two sections later.
Both tabs now read setup -> evidence -> interpretation.
Also: "Three honest limits" has carried four bullets since the monotonicity
precondition was added. Now "Four honest limits".
Mirrors llm-wiki commit f074d81.
Lead both audit tabs with the figure, demote methodology to a footnote
Tab 3 opened with three paragraphs and a table of split-design methodology before
a reader learned whether the heatmap was faithful at all. That belongs in the
notebook, which carries it in full. The tab now shows the deletion curve --
already in assets/ since v2.0 but never displayed -- states the verdict, and
keeps the 85/15 instability as a footnote.
The verdict says "at 30% of pixels replaced": with the chart visible a reader
looks for 0.898 and finds ~0.93 at the 0.5 endpoint. The numbers only reconcile
at x=0.3.
Tab 2 moves the figure above the results table -- the setup described a stamp the
reader could not yet see. Also notes that matplotlib draws each row's captions
above its own images, which otherwise reads as a duplicated figure.
Intro reworded so the roadmap matches what Tab 3 now leads with.
Mirrors llm-wiki commit f4dbfa3.
Publish the notebook and walkthrough under notebook/
The Space cites numbers as "precomputed in the notebook" in several places, which
until now a visitor had no way to check. Both files are here so the evidence sits
one click from the claims.
notebook/mit-group8-explainable-defect-detection.ipynb -- all outputs recorded
notebook/mit-group8-notebook-walkthrough.md -- plain-language, per cell
The footer now links them explicitly; it previously pointed at "the project repo",
which was unactionable for anyone outside the team.
Mirrors llm-wiki commit afe7b34.
The conformal tab still carried pre-clean-split results: it claimed the bound held
at every risk level tested while the chart beside it shows three points above the
line. Now 6 of 9, with the exceedances given as 1.12/0.55/0.83 standard errors.
Also corrected: abstention 72%/3.4% -> 91.7%/12.8%; the monotonicity minimum
(lambda ~ 0.57 "falls to zero" -> 0.029 at lambda ~ 0.461, first rise 0.486);
"the bound holds at whatever value is chosen"; and the figure caption.
The header conflated micro-F1 with top-1 accuracy, which are the same metric here
-- the real difference is two training runs. Verified the served checkpoint at
172/180 = 95.6% against the notebook's 94.1%, and corrected train.py's comments
claiming it IS the notebook's model.
Mirrors llm-wiki commit edfa1a9.
