RepoCortexA field manual for a codebase that learns.

00 / A graph with a memory

Fixes learned1281django/django

Held-out issue → past fix → gold function

Held outIssue #36605
Issue #35690similarity 0.74
QuerySet.in_bulkFIXED_IN / weight 0.90
Rank 1 / gold matchscore 0.825

Measured path, drawn schematically.

For platform teams running coding agents on mature codebases.

Find the functions.
Learn from the fix.

Start an unfamiliar bug investigation with ranked functions and the graph evidence behind them. Feed back the confirmed fix so the next investigation can use it.

Strict Function Acc@10

No graph
41.7%
Frozen
45.1%
Learned
48.3%

319 retrospective Django fix rows

Scope and limitations

All gold functions must appear in the top ten. This is one recorded Django snapshot, which includes later fixes. Three test rows share tickets with training; excluding them preserves the gain. Read the evaluation audit.

Recorded Django evidence. The route is schematic; the function, rank and supporting fix come from the measured replay.

Scope of evidence

What the results establish.

01 / The learning pass

Replay.

Predict. See the real fix. Learn.
Then ask the same questions again.

1281 fixes learned in time order. The same 319 held-out fixes scored at every checkpoint. Three test rows share tickets with training; this is a retrospective replay on a fixed code snapshot.

Fig. 2 / Held-out evaluationMeasured
Held-out File Acc@1 through 1281 learned fixesRepoCortex: 44.51%. Frozen graph: 41.07%. Dense vectors: 35.74%. BM25: 34.48%. Baselines are single-point references, not replay curves. The complete data table follows the controls.File Acc@1 / held-out fixes (%)30354045500 fixes: 41.07%92 fixes: 41.69%183 fixes: 41.69%274 fixes: 41.69%366 fixes: 41.38%458 fixes: 41.69%549 fixes: 42.01%640 fixes: 42.32%732 fixes: 42.32%824 fixes: 42.32%915 fixes: 42.63%1006 fixes: 42.63%1098 fixes: 44.2%1190 fixes: 45.14%1281 fixes: 44.51%RepoCortex44.51%Frozen graph41.07%Dense vectors35.74%BM2534.48%018336654973291510981281Merged fixes learned, in time order

File Acc@1 rises from 41.07% to 44.51%. Function Acc@10 rises from 45.14% to 48.28%. The learned graph beats dense on both. MRR ends at 0.430. Lines connect measured checkpoints, including the dips; baselines are horizontal references.

01281 fixes
How the graph changed

01 Reweight 3,408

Paths that led to the real fix get stronger. Paths that misled it fade.

Ran in this pass

02 Path compression 535

A direct road for a repeated detour.

Ran in this pass

03 Gap ingestion 1667

A missing link, added when it cannot reach the fix.

Ran in this pass

04 Merge / split 33

Resolve shared names and divided meanings.

Ran in this pass
File Acc@1 / same 319 held-out fixes at every checkpoint
Fixes learnedRepoCortexFrozenDenseBM25
041.07%41.07%35.74%34.48%
9241.69%41.07%35.74%34.48%
18341.69%41.07%35.74%34.48%
27441.69%41.07%35.74%34.48%
36641.38%41.07%35.74%34.48%
45841.69%41.07%35.74%34.48%
54942.01%41.07%35.74%34.48%
64042.32%41.07%35.74%34.48%
73242.32%41.07%35.74%34.48%
82442.32%41.07%35.74%34.48%
91542.63%41.07%35.74%34.48%
100642.63%41.07%35.74%34.48%
109844.2%41.07%35.74%34.48%
119045.14%41.07%35.74%34.48%
128144.51%41.07%35.74%34.48%
Function Acc@10 / same 319 held-out fixes at every checkpoint
Fixes learnedRepoCortexFrozenDenseBM25
045.14%45.14%37.3%32.92%
9244.83%45.14%37.3%32.92%
18344.51%45.14%37.3%32.92%
27444.83%45.14%37.3%32.92%
36644.83%45.14%37.3%32.92%
45845.14%45.14%37.3%32.92%
54945.77%45.14%37.3%32.92%
64046.08%45.14%37.3%32.92%
73246.08%45.14%37.3%32.92%
82446.08%45.14%37.3%32.92%
91546.39%45.14%37.3%32.92%
100646.39%45.14%37.3%32.92%
109846.71%45.14%37.3%32.92%
119048.28%45.14%37.3%32.92%
128148.28%45.14%37.3%32.92%
Ranker parameters / this pass
bm25_mix
0.1
memory_lam
0.08
memory_thresh
0.55
memory_k
5
shortcut_lam
0.04
graph_mix
0.05
shortcut_min_support
3
shortcut_min_precision
0.4
memory_agg
sum
memory_frontier
0
recency_half_life
0
file_prior_lam
0
module_prior_lam
0
prior_seeds
12
generalize_concepts
credit_decay
0.08
decay
0.08

The frozen reference is this pass's step 0 graph, with learning off. Learning gain is measured from that starting point.

What the benchmark measures

On 319 chronological fix rows, function Acc@10 is 41.69% for dense + BM25 without graph, 45.14% for the frozen graph, and 48.28% after learning. All gold functions must appear in the top ten.

This is a retrospective replay on one Django snapshot. Three test rows share tickets with training. Excluding them leaves 316 rows: 42.09%, 45.57%, and 48.73%, respectively. The snapshot includes later fixes; these are offline results, not live request measurements.

Read the evaluation audit · Download the reproduced comparison

A measured gain on 319 held-out fix rows. Not a claim about every codebase.

02 / The evidence

Follow the result.

Each candidate has a route through the code graph. Inspect a recorded path, the agent study, or the hosted search check.

Why this function?Issue #36605

A past issue and its learned FIXED_IN link lead to the candidate. This is a captured replay result.

held-out issue #36605│└─ issue #35690 / similarity 0.74   └─ FIXED_IN / weight 0.90      └─ django.db.models.query.QuerySet.in_bulk└─ concept:bulk -SHORTCUT-> django.db.models.query.QuerySet.in_bulk (hits 3, misses 1)└─ concept:list -SHORTCUT-> django.db.models.query.QuerySet.in_bulk (hits 3, misses 1) rank 1 / score 0.825file / django/db/models/query.pygold match / yes gold functions└─ func:django/db/models/query.py:django.db.models.query.QuerySet.in_bulk└─ func:django/db/models/query.py:django.db.models.query.QuerySet.in_bulk.get_obj

Similarity, edge weight, and ranking score are distinct values. Across 319 held-out fix rows, 314 hits used FIXED_IN paths and 63 used SHORTCUT paths; these are hit counts, not accuracy.

Inspect search results →
Agent input tokens16% fewer

63,885 → 53,888 mean input tokens per issue. Asking the agent to call the tool first reduced input tokens by 27%.

Measured with claude-sonnet-5-5. One run per condition; input totals include cached reads. This study measures investigation, not successfully completed fixes.

Read the agent study →

30 Django tasks. Top-1 correctness: 13/30 with the tool, 12/30 without. No established correctness gain.

Hosted search check40/40 passed

Four concurrent requests against the hosted app, with live HydraDB traversal. Median response 3.41 s; p95 3.69 s, including network and queueing.

Inspect the request results →

Four repeated reports. A small read-only smoke check, not sustained capacity, cold-start, or learning-write load.