j SrSSKJr SSKJrJrJrJr SSKJ r SSK J r SSK J r SSKJrJr SSKJrJr SS KJr SS KJrJr SS KJr S r\4SS jjrSSjrSSjrSSjrg)uCheck every cell the scan read against the parse's reading of the same cell. The scan is the reader for a table printed as an image, and it has one failure mode: it turns thousands separators into decimal points. The parse reads the same image and is already in the artifacts. It is not a second table and contributes no structure — it is a second opinion on what a cell says, available for free, and it catches that failure. Every cell in the scan's grid is checked. A cell leaves here confirmed by a second reader, read off the page by an agent, or on the scan's word alone when the agent could not be asked or gave no answer for it: - The parse read the same text. Two independent readers agree, and the cell stands. - The parse read something different. The agent reads the cell. - The parse has no cell there. The agent reads the cell. Having no second opinion is a reason to look harder, not a reason to wave it through. With no parse table at all, every non-empty cell goes to the agent. - No reader is configured, the reader's call fails, or it returns nothing for a cell. The cell keeps the scan's reading, and the gap is logged. Cells pair by how much one box covers the other, taken against the smaller of the two. The readers agree on where a cell is and disagree on how much room it takes: a row heading is bound by the parse to the word itself and by the scan to the whole row band, one box sitting entirely inside the other. Measured against their combined area those overlap almost not at all. Texts are compared as they stand. A currency symbol read as a letter, a comma read as a period, a parenthesis dropped from a negative — each changes what the document states, and a rule deciding which of those were too cosmetic to ask about would be deciding the thing the agent is there to decide. This checks values, not coverage. If the page prints a line the scan drew no box around, the table is short a row and nothing here notices. ) annotations)DictListOptionalTuple) CoordOrigin) TableItem)logger) CellQuestion CellReader) column_lettergrid_to_addressed_markdown)coverage_fraction)NormBoxnorm_box) ScannedTableg?c # Uc/[R"SUR5 URS4$[ XX%5upgU(d/[R"SUR5 URS4$[R"SUR[ U5[ SUR555 URU[UR5U5IShvN nURV s/sHn [U 5PM n n Sn UV s1sHoRiM n n URHnURU ;a.[R"SURUR5 MAX~RupXUUR:wa=[R"SURURXUUR5 URXU'U S - n M XRVs1sHnURiM sn- nU(a+[R"S UR[U55 X4$GNns sn fs sn fs snf7f) aThe scan's grid with every unconfirmed cell read off the page. Returns the grid and how many cells the agent read. The grid keeps its shape: only cell text changes, never the rows, the columns or the boxes. Without a reader, the grid is returned unchanged and the reason is logged. A run that cannot check is not the same as a run that checked and found nothing. Without a parse table to check against, every non-empty cell is sent to the reader. Nz@Cell check: page {} no reader configured; the scan's grid standsrz5Cell check: page {} every cell confirmed by the parsez1Cell check: page {} asking about {} of {} cell(s)c3d# UH&oHo"R5(dMSv M M( g7f)N)strip).0rowcs ;/home/mande/repo/quber/src/quber/core/figures/crosscheck.py #crosscheck_cells..Ws A#1wwyAAAs0 0zPCell check: page {} the reader returned {} which it was not asked about; ignoredz#Cell check: page {} {} {!r} -> {!r}rz[Cell check: page {} the reader returned nothing for {}; those cells keep the scan's reading)r infopagecells _questionslensum read_cellsrlistaddresswarningvaluesorted)scanned table_item page_dims table_imagereadermatch_fraction questions addressesreadingsrgridreadqaskedcellcolrmissings rcrosscheck_cellsr98s$~ VX_XdXde}}a%g9UI  KW\\Z}}a KK;  I A AA  &&{4Nw}}4]_hiiH!( /#DID / D ) * 1YY E * <TZZ ' KK5   #   #  %(..9.Qqyy.99G i LL 7O :Aj / **:s=C&I((I)I(;II(I-CI(I#;I(I(c [X5n/n0n[UR5Hupx[U5HupU R5(dM[ U [ UR U5:aUR UU OS5n [XU5n U bX:XaMk[U 5US-3n Xy4Xm'UR[XU S95 M M XV4$)zMThe cells the parse did not confirm, and where each address sits in the grid.Nr)r% scan_read parse_read) _parse_cells enumeraterrrr! cell_boxes _best_matchr appendr )r)r*r+r. parse_cellsr/r0rtextsr7textboxotherr%s rr r |sz5K$&I,.I . "5)IC::<<3WEWEWX[E\A];]7--c237cghC.AE U]&s+,S1WI6G"%I    \'V[\ ]*/  c Uc/$Uup#/nURRGHCnURnUb(UR=(d SR 5(dM=UR [ R:XacURU- [URUR5U- URU- [URUR5U- 4nOhURU- U[URUR5- U- URU- U[URUR5- U- 4nURURU45 GMF U$)zIThe parse's cells for this table, boxed in the normalized top-left frame.)data table_cellsbboxrDr coord_originrTOPLEFTlmintbrmaxrA)r*r+widthheightoutr6rLrEs rr=r=s- ME%'C++yy < R6688     3 3 366E>3tvvtvv#6#?%QTUYU[U[]a]c]cQdgmQmnC#dffdff--7#dffdff--7 C DIIs#$, JrGczUcgSnUnUH,upV[[X5[X`55nXt:dM*Xup4M. U$)zLWhat the parse read in the cell covering `box` best, or none past the floor.N)rTr)rErBr. best_textbestrDrFcovs rr@r@sL {#I D" #C/1B51NO ;!)# rGN)r)rr*Optional[TableItem]r+Tuple[float, float]r,bytesr-zOptional[CellReader]r.floatreturnzTuple[List[List[str]], int]) r)rr*r\r+r]r.r_r`z5Tuple[List[CellQuestion], Dict[str, Tuple[int, int]]])r*r\r+r]r`List[Tuple[str, NormBox]])rEzOptional[NormBox]rBrar.r_r`z Optional[str]) __doc__ __future__rtypingrrrrdocling_core.types.doc.baserdocling_core.types.doc.documentr logurur quber.agents.cell_readerr r %quber.core.extractors.camelot.acquirer r5quber.core.extractors.camelot.correspondence.geometryrquber.core.figures.geometryrrquber.core.figures.modelsrMATCH_FRACTIONr9r r=r@rGrros"H#..35=[S92+ A A#A#A A ! A  A!AH  # #   ;  2#0C2   )B TY  rG