How the scan was classified.

“Accepted” entries belong to the mathematical catalog. “Rejected” means the extracted span is not a complete mathematical record; it is still preserved here and in its full source. Where an adjacent source passage supplies the complete expression, the recovered replacement is linked.

Download every scan decision ↗

m = r.raw

v = r.version

target = e

Run: go test -tags=integration ./internal/integration/ -run Alert -v

Run: go test -tags=integration ./internal/integration/ -run BidOutcomes -v

Run: go test -tags=integration ./internal/integration/ -run CaptureReadiness -v

0.35*0.5 + 0.30*0.667 + 0.20*0.2 + 0.15*0.3 = 0.46 => REVIEW. Assert the band is REVIEW.

Clearance levels: expected map of code => engine int value (the iota ordinal).

Polygraph types: expected map of code => engine int value.

Run: go test -tags=integration ./internal/integration/ -run Collision -v

Run: go test -tags=integration ./internal/integration/ -run ConnectorICIMS -v

relocate to while still being available (50). 25 + 25 + 15 + 0 + 5 + 5 = 75.00.

out = rows

Female ratio = 0.25 / 1.0 = 0.25, below the 0.80 four fifths threshold, so it must be flagged.

The test asserts structural correctness (rates in [0,1], ratio = rate/max_known_rate, Flagged

The approval pattern is deterministic: candidate index mod 3 != 0 is approved.

Candidates at position i%10 == 0 (indices 0, 10, 20 ...) receive NO demographics row.

All other candidates receive a consent_given=true row via the compliance pool inside

no demographics row. So among candidates who have demographics (i%10 != 0):

male: positions 3,5,8,9 of each 10 block. Even: 8. Count per 10 = 4 from {3,5,8,9}.

Among those, even (screened in): 8 is even. So male screened = varies.

impact ratio (i.e., the reference group count >= 2 so ratios are not all "n/a").

Female ratio = 0.70 / 0.80 = 0.875, ABOVE the 0.80 threshold, so the first snapshot does not flag.

approved 7 of 22 (7 from batch 1 plus 0 of 12 new), rate 0.318, ratio 0.318 / 0.80 = 0.398, BELOW

holds = k+2 (a published worthy count), lacks = 3 (sub k). In a two bucket partition the sub k cell

suppressing executor would render headcount=3 or headcount=12; neither may be present.

federal_ref = source_record_id. It is the only honest place the segment naics comes from.

The deep node at depth 3 carries the product of the three edge strengths (0.90 * 0.80 * 0.70 = 0.504),

Run: go test -tags=integration ./internal/integration/ -run MAGReranker -v

Run: go test -tags=integration ./internal/integration/ -run MAGRerankerLive -v

Run: go test -tags=integration ./internal/integration/ -run MAGTimeToFill -v

Run: go test -tags=integration ./internal/integration/ -run MAGLearningLoop -v

Run: go test -tags=integration ./internal/integration/ -run MAGWinProbability -v

One open requirement (no opportunity link => treated as open) so Mode A has something to rank.

area='US' rows) binds THIS occupation unambiguously as the longest match, immune to what parallel tests

Seed the two national rows (area='US', role=''), both metrics, plus an UNRELATED Hawaii state row so the

fresh org (the candidates rows carry organization_id = org; the search runs in the same org context).

meanRecall@10 >= 0.8: a healthy search recalls nearly every relevant id within the top ten of a twelve

meanMRR >= 0.5: a healthy search puts a relevant candidate in the top two on average (rank one is 1.0,

meanNDCG@10 >= 0.6: nDCG rewards ranking the relevant ids high, not merely present. 0.6 is a conservative

Run: go test -tags=integration ./internal/integration/ -v

App role, NO tenant transaction => GUC unset => current_setting is NULL => policy denies => 0 rows.

As org A, try to attach a skill row (org=A) to org B's candidate. RLS WITH CHECK passes (row

org == session org), but the COMPOSITE FK (organization_id, candidate_id) -> candidates has no

Requirement with nothing required => every component scores 100 => total exactly 100.

App role has no UPDATE/DELETE grant on audit_logs => permission denied (42501).

Run: go test -tags=integration ./internal/integration/ -run Vendor -v

performance_score = fill 0.5*0.35 + quality(neutral 0.5)*0.30 + speed(~1.0)*0.20 + interview 0.75*0.15

= 0.175 + 0.15 + 0.20 + 0.1125 = 0.6375 (speed ~1.0 since submitted ~ requirement created).

every factor at 1.0 and the weights summing to one => the score is exactly 1.0 (no sigmoid distortion).

= 0.35*0.8 + 0.30*0.5 + 0.20*1.0 + 0.15*0.0 = 0.28 + 0.15 + 0.20 + 0 = 0.63

codes, and the scalar array membership predicate (= ANY) keeps only the rows of the global

The honesty mandate is enforced: the award side keys only on (awarding_agency = o.agency, naics_code =

nationalOccupations reads the distinct national OEWS occupations (area = 'US') from the global public

Fall back to the US national band (area = 'US'). This is the M5 Phase 1 broadening: any resolvable SOC

log odds = intercept + SUM(coefficient_j * feature_j) in decimal, applies the logistic (a single isolated

fixed champion, a fixed feature vector, and no DB. It computes log odds = intercept + SUM(coefficient_j *

An invalid (NULL) or non finite numeric is an honest ok false; a finite one is value = Int * 10^Exp.

The log odds is intercept + 0.5 * 0.4 = 0.4054651081 + 0.20 = 0.6054651081.

The trace: the intercept term (value 1, coefficient and contribution = intercept) plus the feature term

sigmoid(0.6054651081) = 1 / (1 + e^-0.6054651081). e^-0.6054651081 approx 0.54580, so p approx 0.64695.

The base rate is sigmoid(logit(0.60)) = 0.60. Allow a small rounding tolerance.

MATERIALIZED two arm fusion with k=60 the candidate query uses.

vendor_quality_distribution BENCH intent joins them to the frozen candidate_evaluations band and is k=10

2026-06-30 de identification research (NIST SP 800-188, FCSM WP22, l diversity, NYC LL144): k = 10 is

vendor_quality_distribution is a BENCH intent (person derived headcount partition, k=10 small cell

no clearance required (none) with score 0 => not a gap (engine passes everyone), folds nothing

- vendor_quality_distribution: a BENCH intent, k=10 small cell suppressed. Per vendor, the headcount of its

IntentVendorQualityDistribution is the registered name of the k=10 suppressed vendor quality distribution (a

(evaluated_candidates >= 10), else NULL. The floor counts PEOPLE, not submissions, so a handful of people