Judge Experiment · Secret Classification Benchmark

Context vs. Entropy

Raw pattern/entropy detectors flag 18,567 candidates across repository history. HydraDB + FCO/FCG graph context classifies 18,555 of those raw findings (99.9354%) into reviewed provenance classes while retaining 12 unexplained candidates as abstentions.

View Classification Flow View Vithia Supplementary Card Evidence ledger

Public data plane

Hosted HydraDB not connected on this deployment

DEPLOYMENT NOT CONFIGURED

BackendCHECKING

DatabaseNOT_CONFIGURED

CollectionNOT_CONFIGURED

TraceabilityNOT_CONFIGURED

This card is only a live server-side connectivity/readback diagnostic for the current deployment. HydraDG's retained local HydraDB executions, graph use case, deterministic math and bounded hosted receipts are separate evidence objects. Credential values are never returned to the browser.

Why HydraDB · math · implementation Open retained evidence

01 / CLASSIFICATION PIPELINE

Pattern Detector → HydraDB Context Graph → Deterministic Decision

STEP 01

Raw Pattern Detector

Gitleaks built-in pattern/rule scanning provides the first-stage candidate stream.

Input Stream

STEP 02

18,567 Flags

Candidate matches retained across repository history.

Raw Findings

STEP 03

HydraDB Context Graph

Namespace: hydradg-context-entropy-20260820

FCO/FCG Provenance

STEP 04

99.9354% Classified

18,555 classified · 12 abstentions reserved for review.

Context Decisions

02 / EXPERIMENT RESULTS

Summary Metrics & Category Breakdown

RAW FINDINGS

18,567

Total raw detector matches

CONTEXT CLASSIFIED

18,555

99.9354% coverage

ABSTENTIONS

12

Unexplained secret candidates

REVOKED CREDENTIALS

1

User-attested revoked historical item

DETERMINISTIC_HASH18,428 (99.25%)

Content-addressed SeedGraph SHA-256 cache identifiers in deterministic cache paths.

Public-Safe Example: HydraDG_DaisyTrain_v0.3.7/evidence/track03/.../seedgraph/cache/00049980b80...json

TOY_NON_AUTHENTICATING_KEY126 (0.68%)

Intentionally public non-authenticating toy signature key fixtures declared in FCO package seals.

Public-Safe Example: archive/hydradg_video_final_reconciled_v3/fcg/TOY_SEAL_FCO_BUNDLE.jsonl

REVOKED_HISTORICAL_CREDENTIAL1 (0.01%)

Historical Modal token ID line in ECA retry log; user-attested revoked (USER_ATTESTED_REVOKED).

Public-Safe Example: HydraDG_DaisyTrain_v0.3.3/logs/20260817_081811_ECA_RETRY_2.log (Token: ak-REDACTED)

UNEXPLAINED_SECRET_CANDIDATE12 (0.06%)

Abstentions reserved for human/operator review; not silently allowlisted.

Public-Safe Example: apps/hydradg-web/.next/cache/.previewinfo

This second-stage contextual classification does not replace Gitleaks and does not establish that no secret can exist under every detector or context.

03 / MODAL TOKEN PRESERVATION

User-Attested Revoked Credential Record

REVOKED_HISTORICAL_CREDENTIAL · USER_ATTESTED_REVOKED

Historical Modal Token Identifier

The historical Modal token-ID line is classified as REVOKED_HISTORICAL_CREDENTIAL based on user attestation.

Evidence Basis: USER_ATTESTED_REVOKED · Provider Verified: false. Credential values are not reproduced here.

04 / SUPPLEMENTARY CARD

Vithia / Pythia-14m Training Baseline Repair

PRESERVED FAILURE CONTROL → REPAIRED BASELINE → REFERENCE BASIN

01. Failure Control

VITHIA-OVERNIGHT-01

Step 0 Divergence (AdamW eps=1e-8)

02. Repaired Config

AdamW lr=1e-4, eps=1e-5

grad_clip_norm=1.0 · 100% finite reference runs

03. Reference Basin

5-Seed Baseline Suite

Loss: 10.8424 ± 0.0020

04. FCO Admission

Turn 1 Admitted

fco:0f63d48a... (PASS)

Explicit Claim Ceiling Notice: This supplementary card documents a reproducible, numerically stable Pythia-14m reference baseline process. It does NOT claim improved LM accuracy, convergence superiority, LongMemEval superiority, or end-to-end QA improvement.