Kmisener commited on
Commit
d92bc51
·
verified ·
1 Parent(s): 376fad1

Upload 2 files

Browse files

Cases now run through November 6, Readme updated

Files changed (2) hide show
  1. README.md +95 -11
  2. gao_corpus.jsonl +2 -2
README.md CHANGED
@@ -1,11 +1,95 @@
1
- ---
2
- license: mit
3
- tags:
4
- - legal
5
- - contracts
6
- - FAR
7
- - acquisitions
8
- - law
9
- - caselaw
10
- ---
11
- Date range - Bid Protests from November 6, 2026 to begining of digitized decisions
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # GAO bid protest corpus — November 2026 run
2
+
3
+ This is the September run brought current. It is an **incremental top-up**,
4
+ not a re-crawl: the 33,228 decisions already in hand were not re-fetched or
5
+ re-parsed, so every record carried over is byte-identical to `Sept. Run`.
6
+
7
+ `Sept. Run/` is left untouched as the fallback.
8
+
9
+ ## What is here
10
+
11
+ | file | records | notes |
12
+ |---|---|---|
13
+ | `gao_corpus.jsonl` | 33,193 | **the dataset.** Cleaned, 358.0M chars, mean 10,786 |
14
+ | `data/decisions.jsonl` | 33,285 | the canonical pre-cleaning records |
15
+ | `data/index.jsonl` | 33,291 | one row per decision URL (the catalogue) |
16
+ | `data/errors.jsonl` | 6 URLs | unchanged since 2026-09-12; see below |
17
+ | `data/slices_state.json` | | per-slice enumeration results |
18
+
19
+ `gao_corpus.jsonl` is 92 records short of `decisions.jsonl` because
20
+ `make_corpus.py` drops decisions with under 500 characters of text.
21
+
22
+ The 33,285 page snapshots and 10,448 PDFs are **not** here. They live at
23
+ `C:\GAO-data\bid-protest\` on the local disk, deliberately — see "Where to
24
+ put `data/`" in `GAO Fiscal Law/scraper/README.md` for why Drive is the wrong
25
+ home for a 47,000-file crawl tree.
26
+
27
+ ## What changed since September
28
+
29
+ | | Sep 13 | Nov 8 |
30
+ |---|---|---|
31
+ | index rows | 33,234 | **33,291** (+57) |
32
+ | page snapshots | 33,228 | **33,285** (+57) |
33
+ | PDFs | 10,391 | **10,448** (+57) |
34
+ | records | 33,228 | **33,285** (+57) |
35
+ | corpus records | 33,136 | **33,193** (+57) |
36
+ | newest decision | Sep 9, 2026 | **Oct 29, 2026** |
37
+
38
+ September 33,136 → November 33,193 is exactly +57, so nothing was lost in the
39
+ process. All 57 new records carry full decision text and a PDF; agency
40
+ resolved on 45, outcome on 50, solicitation number on 47 — normal rates for
41
+ recent protests.
42
+
43
+ ## How complete is it
44
+
45
+ **Complete for the gap period, and provably so.** Enumerating GAO's own
46
+ `search_date_range` filter for 2026-08-01 .. 2026-11-08: GAO printed 101
47
+ items, pagination returned exactly 101, and all 101 were already in the index.
48
+ Zero missing.
49
+
50
+ Two caveats, both small and both honest:
51
+
52
+ - **The index is 1 short of GAO's printed 33,292.** It is not in the recent
53
+ window (see above), and GAO's printed totals are known to run slightly
54
+ high — in September the same counter read 33,230 while the index already
55
+ held 33,234 genuine rows. Treated as a phantom, but not proven to be one
56
+ without re-walking the full date tree.
57
+ - **6 index rows have no page or record**, all pre-existing from September,
58
+ each tried twice before being retired. They are GAO's own malformed
59
+ `...,et al` URLs for the enormous consolidated `B-419956` dockets, plus
60
+ four older ones. This run added zero new errors.
61
+
62
+ Separately, 83 records corpus-wide (0.25%) have no decision text, of which
63
+ only 5 have a PDF to recover from. Not worth rewriting a 423 MB file for, so
64
+ they were left alone.
65
+
66
+ ## The one thing to know before updating this again
67
+
68
+ **GAO orders and filters by release date, not decision date.** This run
69
+ picked up decisions dated as far back as 10 September that GAO only released
70
+ in late October. So:
71
+
72
+ - Do not stop crawling when the dates "look old" — crawl a fixed page count
73
+ and let the item count tell you when you are done.
74
+ - Do not verify by comparing GAO's date-filtered count against
75
+ `decision_date` in the index. They are different fields; doing so showed
76
+ phantom gaps of +16, +11 and +6 across three windows where nothing was
77
+ actually missing. Diff the URL **sets** instead.
78
+
79
+ The full recipe is in `GAO Fiscal Law/scraper/README.md`, under
80
+ "Updating a finished corpus".
81
+
82
+ ## Provenance
83
+
84
+ Crawled with the scraper at `GAO Fiscal Law/scraper/`, which is the September
85
+ scraper generalised to cover more than one GAO series. The protest series is
86
+ `-c bid-protest`; appropriations law is the default. Field derivation for
87
+ protests is unchanged and pinned by `selftest.py`.
88
+
89
+ python build_index_browser.py -c bid-protest --data-root C:\GAO-data \
90
+ --start-page 0 --max-pages 15 --no-warmup --delay 18
91
+ python fetch_decisions.py -c bid-protest --data-root C:\GAO-data --delay 2.0
92
+ python make_corpus.py -c bid-protest --data-root C:\GAO-data \
93
+ -o gao_corpus.jsonl --format jsonl
94
+
95
+ Run 2026-11-08, 00:40–01:18. Enumeration 6 minutes, fetch 26, corpus cut 6.
gao_corpus.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ebb0aea4fcd6f34df98a48d9cafb758bb44e8359777cfdaef9f7727f178fbf58
3
- size 404332095
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b58c8497fe8a6254627bfe5528923dbcc369c7fd85cdcd5b28fe18b8051ff28c
3
+ size 406005123