Open_Ultra_SAST/.kiro/specs/flow-aware-ranking/brief.md

Brief: flow-aware-ranking — spend the expensive query where the cheap evidence is strongest

Status: brief only, unapproved.

Why now

"we need a taint flow aware ranker, that is capable to put high risk areas into tiers" "can you engineer a prompt with enough information to steer the llm making ranker decisions" "this can be done, but it requires an iterative concept and clear plans"

A taint request costs 6.63 seconds (contributor-scan 5.13). A 637-file plugin yields 4,463 regions, and the current ranker gives 83.8% of them the same rank, ordered alphabetically by path. So the budget is spent arbitrarily and the scan does not finish.

The ranker is not weak. It is looking at the wrong thing.

The evidence that this is tractable

Three regions from one plugin, with what the ranker sees today and what is actually in the code:

region	rank today	sink in scope
getMemberOrderByCode — holds CVE-2023-23488	0.30	$wpdb->get_var("... WHERE code = '" . $code . "' LIMIT 1")
getMemberOrderByID — sibling	0.30	$wpdb->get_row("... WHERE id = '$id' LIMIT 1")
pmpro_rest_api_get_membership_level	0.80	none — no SQL sink at all

A region with no sink outranks two holding raw concatenated SQL. The discriminating signal is in the sink call's own source text — interpolated versus bound — and nothing in the ranker reads it.

One detail decides whether a keyword rule suffices, and it does not: esc_sql IS in scope for the CVE's function, applied to a different statement. "A sanitizer is present, rank it lower" demotes precisely the right answer. Knowing which call it wraps is reading, not matching.

The unit, and the tiers

Rank (region, family) PAIRS, not regions. The expensive question is "does a source reach a sink of family F here"; every cheap approximation of it is free from the fact tables and a text scan.

tier	evidence
0	no sink of F in scope	never ask — exact, cannot lose a finding
1	sink present, no source in scope or reachable	ask last
2	sink and source, a cleansing call on THAT call	may still corroborate
3	sink and source, interpolated shape, no cleansing on that call, entry point or declared public	ask first

Tier 0 is the only tier that excludes rather than orders, so it alone must be exact. Above it, being wrong costs position, not the finding. That bound is what makes a model safe here: a ranker cannot manufacture a finding, it decides what is looked at, and regions_truncated already reports what went unexamined.

The iterative concept

Each phase ships only if it beats the previous one on a repository the change was not developed against. A phase that cannot show that is not an improvement, it is a story.

• Phase 0 — the oracle and the baseline. The evidence extractor (no model) and a scoring harness over the pinned CVEs, reporting each one's rank position. Gate: reproduces today's numbers — CVE-2023-23488 at position 390 of 4,463, 83.8% of regions in one tier. Without a baseline nothing later is measurable.

  Done 2026-09-10. benchmarks/ranking/position.py, committed as benchmarks/ranking/baselines/2026-09-10-access-rank.json:

  repository	regions	largest tier	CVE position	budget_at_recall
  pmpro	4,463	83.8% at 0.30	390	391
  mwwpform	983	81.7% at 0.30	512	513
  wpstatistics	1,673	75.3% at 0.30	503	504
  vampi	25	36.0% at 1.00	15, 10 (SQLI reached transitively)	16

  And this corrects the cross-assessment above. "Ranking's job is concentration, not inclusion" was measured on pmpro alone, where the CVE is inside the budget. On the other two plugins the CVE sits at 512 and 503 — outside a 500-region budget — so the current ranker never examines it at all, and the slices found those CVEs only because a slice has too few regions to exclude anything. Both jobs are real: inclusion on two repositories, concentration on all three. The single metric budget_at_recall covers both, which is why it is the one that matters.

  The instrument is evidenceOnly mode on the taint query, recorded on every scan as ModelScanResult.tiers and tier_counts, and acted on by nothing — so every later phase is measured against a record made by scans that did not know they were being measured.

  Measured cost of the instrument, 637-file graph, the scan's own 500-region request set:

  2,500 pairs in 98.9 s including JVM load — about 40 ms per pair
  tier 0 (no sink of that family in reach): 2,161 of 2,500 = 86%
  

  Two attempts before that ran past an hour each without producing a payload, and the cause was mine both times: repository-wide walks — the sink scan, the reachable-method BFS, familyInRepo — repeated per request. Memoising them across the batch is what made "milliseconds per pair" true rather than claimed.

  And the same memoisation reaches the dataflow mode. The identical 250-request set that cost 2,026.3 s (8,105 ms per request) with both joins on now costs 287.0 s — 1,148 ms per request — for the same 5,779 rows. Most of what contributor-scan 5.13 measured as "the base taint query" was the per-request scans, not the dataflow. That correction is recorded there.

• Phase 1 — tier 0 only, no model. Exact pruning. Gate: 53% fewer taint requests issued, measured — and nothing more. 1,177 requests at 6.63 s is still 2.2 hours, so this phase does NOT make the scan finish; it halves the problem. Claiming completion here would send someone looking for a bug where there is none.

  Done 2026-09-10, gate exceeded, scan still does not finish — and the reason corrects phase 0's cost figure. On the 637-file plugin: tier_counts = {0: 2161, 1: 60, 2: 20, 3: 211, 4: 48}, 2,161 of 2,500 requests pruned (86%), evidence pass 94.5 s. The remaining 339 requests still hit the 2,400 s ceiling — more than 7 s each.

  The scan finishes, 2026-09-11. Not because of anything in this spec: contributor-scan 5.13 found that every query batch had been recomputing Joern's dataflow overlay at load since PHP builds went through php2cpg (54 s fixed cost per batch; an OutOfMemoryError on WP Statistics), that the hook half's seeds were repository-wide, and that two vendored data tables were 60% of one graph. With the overlays applied once at build time, the seeds scoped to the callback's file, and files over 1 MB excluded as data, the same whole-repository scan on the 637-file plugin completes: build 97 s (overlays included), evidence 76 s, taint 4,105 s for 348 real requests (11.8 s each), arbitrate 99 s, total 72 minutes, tier_counts = {0: 2152, 1: 60, 2: 20, 3: 218, 4: 50}, peak child RSS 2.4 GB. CVE-2023-23488 is found at class.memberorder.php:936, among 250 entailed findings across 40 files. The 250 is the next fact to reckon with -- it is what "500 regions when 50 would do" looks like as output -- and it is phase 2's material. callDepth, measured on the same graph, is a 1.4× time term and a 100× payload term (3,838 flow rows for ten requests at depth 3 against 38 at depth 1). Collapsing rows per evidence key takes 3.3× of that out (1,155 rows carrying 3,838 paths, exact); the rest is depth 3 naming more distinct sources per sink, which is evidence the ranker gets to use.

  Phase 0 reported 1,148 ms per request after memoisation. That was an average over a 250-request set of which 88% were tier 0 and returned instantly; the ~30 real requests in it cost about 9.5 s each. Pruning removes exactly the cheap questions and keeps the expensive ones, so it cannot make those faster. The honest per-request cost of a taint question that can have an answer is 7–10 s at callDepth = 3, and that is the number phase 2 and 5.13 must work against — not 1.1 s, and not 6.6 s either.

• Phase 2 — heuristic tiers, still no model. The vector and tiers defined below, ordered within a tier by a published weighted sum. Gate: budget_at_recall — the smallest K holding every pinned CVE — below 50 on every pinned repository, and pmpro's taint query completing at that K. This is the phase that makes a scan finish, because the job is concentration: the CVE is already inside 500, the question is whether it is inside 50. If this passes, the model is a refinement, not the plan — and knowing that is worth more than assuming it. Implemented 2026-09-11, measurement in flight. Evidence.score is the published weighted sum, and every weight is in one table in model/evidence.py: per open sink +1.0 (capped at 5), per cleansed or bound sink −0.5 (capped at 3), local source +2.0, carried +1.5, near source +1.0, entry +0.5, declared public +1.0. A tier-0 pair scores 0 whatever its sources say, so the weights order but cannot manufacture. ScanBudget.order_by_evidence runs the evidence pass over every region (or evidence_regions of them), takes each region's best pair's (tier, score), and cuts max_regions from that order; order_by_evidence() is one function used by the scan and by benchmarks/ranking/position.py --evidence, so the benchmark measures the order the scan spends. Off by default until the benchmark says where the pinned CVEs land. A region with no taint family sorts after every region with evidence -- the scan does not know it is safe, it knows nothing -- which is a known cost of this phase for families like access_control and is noted here rather than hidden.

• Phase 3 — the LLM ranker. Same evidence, ordered by a model. Gate: beats phase 2 leave-one-repository-out. If it does not, phase 2 ships and phase 3 does not.

• Phase 4 — break the closed loop. An exploration slice (~10% of budget sampled outside the top-K) and full labelling on repositories small enough to arbitrate every region. Gate: a measured estimate of the selection bias in phase 3's numbers.

• Phase 5 — prompt optimisation (GEPA / SIMBA / DSPy). Only after 3 and 4. An optimiser run against a censored metric optimises the censorship.

What the prompt has to carry

The question put to the model is not "is this vulnerable" — that is adjudication, it is the arbiter's job, and it is the one thing a ranker must never be asked. It is:

Given a fixed budget of expensive dataflow queries, which of these sites are worth spending it on?

and the evidence it needs is textual and already available: the sink calls in scope with their source text, the untrusted-input expressions found, the cleansing calls and which call each one wraps, the sink's shape (bound versus interpolated), the declared access and its provenance, and whether the region is reachable from a request handler.

Constraints, from measurement and from prior scars

• Cost. One model call per region is 4,463 calls a scan. Batch over a prefilter, or it does not ship.
• The loop is closed. A region the ranker excludes is never arbitrated, never labelled, so the ranker never learns it was wrong. Every run confirms the ordering it already had.
• N projects. Four pinned CVEs is not a training set. This project already has the closed-loop train-on-test scar; the reported figure must come from a repository the optimiser never saw.
• Determinism. Two runs of one repository must agree or a baseline diff is meaningless. The ranking is recorded in the manifest.
• Degradation. With no model configured, phases 0–2 still rank, and the report says which ranker ordered the scan.

Cross-assessment against the measurements

Read against what was actually measured today, the plan above has two errors and one gap. They are kept here, corrected, rather than silently rewritten, because the corrections are themselves results.

Phase 1's gate is wrong. It says pruning makes pmpro's taint query complete. It does not. Pruning removes 53% of requests, and 1,177 × 6.63 s is still 2.2 hours — nowhere near the 300 s default or the 2,400 s that was tried. Pruning halves the problem; it does not solve it. The honest gate is "53% fewer requests issued, measured", and completion belongs to phase 2. Stating otherwise would have sent someone looking for a bug in phase 1 when phase 1 had done exactly what it can.

Amended 2026-09-11: completion arrived, and it belonged to neither phase 1 nor phase 2 -- it belonged to the instrument (contributor-scan 5.13). Pruning was and is exact; what made 339 requests exceed 2,400 s was a dataflow overlay recomputed per batch and a seed set computed repository-wide, and the scan now completes in 72 minutes with the CVE found. The paragraph above stands as written because its reasoning was right on the numbers it had; the numbers were the instrument's.

Ranking's job is concentration, not inclusion. The brief implies the CVE is missed because the ranker excluded it. It did not: CVE-2023-23488's region sits at position 390 of 4,463, inside the 500-region budget, and was examined. It goes unreported because the budget is spent on 500 regions when 50 would do, and 500 is unaffordable at 6.63 s per family per region. So the ranker's success metric is not "is the CVE in the budget" — it already is — but "how small can the budget be while the CVE stays in it". That is a different objective, and it is the one that makes a scan finish.

The gap: nothing above says what the evidence IS. "Sink shape" and "cleansing on the call" are named, not defined, and a heuristic that is not defined cannot be language-agnostic because there is nothing to check for language leakage. The rest of this brief defines it.

Two things the measurements confirm and the plan should lean on harder:

• The joins are cheap and must survive any ranker. 22% of query cost, two of three CVEs. A ranker that demotes field- or hook-carried regions because they show no direct source is a ranker that loses those CVEs. The evidence vector below therefore carries the join's own stage-one result as a field.
• 84% of regions in one tier means any separation is progress, and the separation does not need to be clever to be large. A tier 0 that is merely exact already reorders the budget more than the current ranker ever has.

The evidence vector, defined

The unit is a (region, family) pair. For each, a fixed vector of fields is computed. Every field is defined in terms of fact-table categories — source, sink, sanitizer, dispatch, access — never in terms of a language's syntax, which is what makes the definition language-agnostic: any language with a fact table gets the same vector, and a heuristic that mentions $wpdb or wp_ajax_ has by construction left the vector and must be rejected in review.

Where it is computed matters as much as what it is. The graph is already built for the arbiter, at 76 s for 637 files. Extracting these fields is a CPG query without dataflow — call nodes, their arguments, their enclosing methods, the call graph one hop out — which is milliseconds per call rather than 6.63 s per request. The extractor is therefore a sibling of taint.sc that never calls reachableByFlows:

Tiering is the CPG query without dataflow. Arbitration is the CPG query with it. Same graph, same facts, two costs.

That is the whole architecture in one line, and it is why the evidence is language-agnostic rather than merely language-neutral-in-intent: the CPG normalises syntax across frontends, so "the argument of this call is a literal" means the same thing in PHP, Python and C.

Sink evidence — from [[sink]] facts of family F

field	definition	exactness
sinks	count of calls in scope matching a sink of F	exact
shape	per sink call, one of: literal (the vulnerable-position argument is a constant), bound (the call is a declared-safe form — a parameterized sanitizer fact wraps it, or its arity/shape matches the bound form the taint query already recognises), computed (the argument is an expression: concatenation, interpolation, a variable, a call result)	exact for literal and bound; computed is the residue
cleansed_on_call	per sink call, whether a sanitizer of F is an ancestor of the vulnerable-position argument in the AST — not merely present in the same function	exact, and this is the field a text scan gets wrong: esc_sql in scope on a different statement is false here

Source evidence — from [[source]] facts and the entry-point contract

field	definition	exactness
source_local	a modelled source expression occurs in the region's own body	exact
source_near	a modelled source occurs in a method within k call-graph hops of the region (k = the arbiter's callDepth)	exact for the graph the arbiter will use
entry	the region is an entry point per the mapper, so its parameters are untrusted by contract	exact
access	the declared access level, with provenance: declared (a registration or route contract says so) or inferred (nothing found) — because "public" meaning "no decorator found" is the false positive the project has already paid for	exact
carried	the region reads a field whose assignment is fed by a source anywhere in the file, or a dispatch value whose registered producer is — i.e. stage one of the two-stage join, which is already computed and memoised, reused here as evidence	exact where computed; absent (not false) where the join was not run

Path evidence — cheap structural proxies for what dataflow would establish

field	definition
distance	fewest call-graph hops from any source (local, near, entry parameter, or carried) to any sink of F: 0 = same statement, 1 = same method, 2+ = across calls, ∞ = no source found
cleansing_between	any sanitizer of F occurs on a call-graph path between a source and a sink — a weaker, cheaper version of cleansed_on_call that catches cleansing done in a helper

Structural and prior evidence

field	definition
shipped	the project's own build declaration names this file (existing shipped.py)
family_in_repo	any sink of F exists anywhere in the repository
prior	a committed baseline recorded a finding at this site, or the site was in an exploration slice before

The tiers, defined as predicates over the vector

tier	predicate	role
0	sinks == 0 in the reachable scope, or family_in_repo == false	exclude — exact; the only tier that removes work
1	sinks > 0 and no source_local, source_near, entry, or carried	ask last — nothing untrusted can arrive
2	a source exists, but every sink call is bound or cleansed_on_call	ask late — a flow may exist and will corroborate, not entail
3	a source exists and some sink call is computed and not cleansed_on_call	ask first
3★	tier 3 and (access == declared public or distance ≤ 1 or carried)	ask very first

Two properties are load-bearing and must hold in every implementation:

1. Only tier 0 excludes; every other tier orders. So the only place exactness is required is tier 0, and tier 0's predicate has a proof: a family with no sink in reach cannot produce a flow, in any language, under any dataflow model. Above tier 0, being wrong costs position, and regions_truncated reports what position cost.
2. Tier 0's scope must match the arbiter's. At callDepth = 0 the scope is the region; at depth k it is the region plus its reachable methods. A tier 0 computed on the file when the arbiter will look k hops out is not exact and must not be called so. The repository-wide form (family_in_repo == false) is exact at every depth and is the cheap first cut — it alone removes all 487 deserialization requests on pmpro.

Within a tier, order by a score. In phase 2 that score is a fixed weighted sum over the vector, published in the manifest so two runs agree. In phase 3 it is the model's ordering of the same vector, which is what makes the two comparable: same input, different ordering function, one held-out metric.

Where the model helps, and where it must not be allowed to

The vector above is exact wherever exactness was cheap. What remains genuinely uncertain is narrower than "ranking":

• Ambiguous cleansing. cleansed_on_call is exact when the sanitizer wraps the argument directly. It is false — correctly, but unhelpfully — when cleansing happens two statements earlier into a variable that is then interpolated. Deciding whether that cleansing reaches this call is reading the code, and it is the single largest source of tier-2/tier-3 misplacement.
• Unfamiliar shapes. A sink call whose form the bound rule does not recognise, in a framework with no facts yet. The model can say "this looks bound" as an ordering opinion; it cannot add a bound fact, because a fact is policy.
• Within-tier priority when the fixed score ties, which on a real repository it will, constantly.

And the boundaries:

• The model never computes tier 0. Exclusion is exact or it is a silent miss, and a model's "no sink here" is not a proof.
• The model never sees the question "is this vulnerable". It sees "which of these is worth an expensive query", with the vector and the sink call's source text. Adjudication is the arbiter's, and keeping it there is what keeps the gates byte-identical.
• The model's input is the vector plus text, never the arbiter's verdict on the same run — that is the closed loop feeding on itself.

Metrics, so every phase can fail

metric	definition	phase
position(cve)	rank position of each pinned CVE's (region, family) pair	0 → all
requests_issued	taint requests after tiering	1
budget_at_recall	the smallest K such that every pinned CVE is in the top K — the concentration metric	2, 3
time_to_complete	wall clock for the taint query at that K	2, 3
held_out_position	position(cve) on a repository excluded from development of the change	2, 3, 5
exploration_yield	findings per region in the exploration slice versus in the top-K — the direct measure of what the ranker is missing	4

budget_at_recall is the one that decides the release question. Today it is ≥ 390 on pmpro. If phase 2 takes it below 50, the scan finishes in minutes and the model is a refinement; if it does not, the model is the plan.

Open questions for the design phase

Two of the four questions this brief opened with are answered by the definition above, and are kept so the answer is visible:

1. Where does the evidence extractor live? It is a CPG query without dataflow, a sibling of taint.sc in cpg/queries/. Not the mapper: the mapper reads text and would have to re-derive per language what the graph already normalises. The vector is language-agnostic because the graph is.
2. Per file or per reachable set? Both, in that order. family_in_repo first — exact at any depth, free, a fifth of the budget on pmpro — then the reachable set at the arbiter's own callDepth. Never the file alone above depth zero, because that is not exact and must not be called tier 0.
3. How many pinned repositories before leave-one-out means anything? Four is not enough. This is the corpus ask, and it is the same ask contributor-scan 5.8 made: plugins chosen by installed base, CVEs read out of the code.
4. Does the arbiter's verdict feed back as a label? Only through the exploration slice. A verdict on a region the ranker chose is a label the policy selected, and training on it closes the loop. A verdict on an exploration region is the only unbiased label the scan produces, and it is the one worth keeping.
5. New: the weighted sum in phase 2 has weights. Where do they come from before any model exists? Proposed: set by hand from the tier semantics, published in the manifest, and treated as a baseline for phase 3 to beat — not tuned, because tuning four weights on four CVEs is the overfit the plan exists to avoid.