Repository navigation
Implement all stubs for real: verified the 'Soul Signature does not exist' audit + remediate every placeholder (no AI-slop) #1021
Description
Summary
An external deep-dive audit asserted that the person-identification ("Soul Signature") capability does not actually exist in this repo — it is a research spec wired to a no-op oracle. We verified the audit against the code: it is accurate. This issue tracks turning every such stub into a real, tested implementation (no labels, no placeholders), per the "prove everything / no AI-slop" standard.
Verified findings (file:line evidence)
| Audit claim | Status | Evidence |
|---|---|---|
"stub oracle that always returns nothing (coherence_gate.rs:133-149)" |
✅ TRUE | wifi-densepose-bfld/src/coherence_gate.rs:145-148 — NullOracle::matches_enrolled() unconditionally returns NotEnrolled; the only other impl SoulMatchOracle are test doubles in tests/. |
| Soul Signature is "Pre-Implementation" spec | ✅ TRUE | docs/research/soul/specification.md header: "Status: Research Specification (Pre-Implementation)"; §3.6 matcher was spec-only. |
| "no working person-identifier" | ✅ TRUE | bfld crate header: "P1 in progress … P2–P6 follow." No production oracle computes a match. |
| README presents person-ID as shipped | ✅ TRUE (overclaim) | README.md:504 "identifies people"; :515 "Person re-identification … Track individuals"; :613 "Soul Signature SoulMatchOracle integration". |
The audit's framing is correct and is the remediation plan: build the first running implementation of §3.6 (the per-channel matcher), put real numbers on it, and state honestly where it cannot lock identity yet (the decisive AETHER + body-resonance channels were never fed with real enrollment data).
Additional production stubs found in the same sweep (being implemented for real)
occworld-candle/src/vqvae.rs:260-315—encode_occupancy/decode_to_logitsreturnTensor::randn→predict()emits trajectory priors derived from random noise, with no flag. (occworld-impl)mat/src/localization/fusion.rs:79—simulate_rssi_measurementsalways returnsvec![]→ every survivorlocation: None→ dedup disabled → count-inflation (one person → N survivors). (m4-mat-impl)mattwo divergent triage engines; the confidence-gated ensemble result is discarded (lib.rs:524 _ensemble). (m4-mat-impl)mat/integration/hardware_adapter.rs+csi_receiver.rs— ESP32/UDP/PCAP/PicoScenes/Intel5300/Atheros adapters "not yet implemented". (m4-mat-impl)wifiscan/adapter/wlanapi_scanner.rs— claims a native 10–20 Hz path; actually wraps netsh (~2 Hz). (wifiscan-server-impl)homecore-assist/runner.rs:125+recognizer.rs:149+homecore-recorder/db.rs:234— always-empty / TODO stubs. (homecore-impl)sensing-serverengine_bridge / rvf_pipeline / training_api / matter commissioning placeholders. (wifiscan-server-impl)
Honesty boundary (non-negotiable)
Where a capability is genuinely hardware- or data-gated (Intel 5300 / Atheros / PicoScenes CSI need specific NICs; OccWorld needs trained weights; named identity needs the body-resonance channel fed during enrollment), the code path is implemented for real and returns a typed error or carries an explicit weights_trained/provenance flag — never fabricated data and never a claimed accuracy number without a measurement.
Done = every stub above has real logic + a test that fails on the old stub + MEASURED results, full workspace cargo test green, README overclaims corrected to match reality, then published/released.
Branch: feat/v2-beyond-sota-sweep.
Activity
docs(readme): correct person-identification claims to measured reality (
ruvnet commentedon Jun 12, 2026
Sweep complete — all stubs implemented for real, verified end-to-end
This issue's remediation grew into a full 10-milestone
v2/crates/beyond-SOTA sweep (ADR-154–163), now fully merged tomainand released as v0.9.0-beyond-sota.Validation:
cargo test --workspace --no-default-features→ 3,157 passed / 0 failed ·python archive/v1/data/proof/verify.py→ VERDICT: PASS ·bash scripts/prove.sh→ RESULT: PASS.The original audit, answered:
wifi-densepose-bfld, 364 tests); the audit's own claim that WiFi-only channels can't lock identity is now a measured result (gap 0.0005, re-run byprove.sh).Every stub fixed for real (PRs #1023, #1024, #1025, #1026, #1030, #1033):
randn→ real conv VQ-VAE +weights_trainedflagwlanapi.dllFFI (measured 9.74 Hz); real Matter 1.3 pairing codeHonesty boundary held: hardware/data-gated paths (named identity, OccWorld weights, ESP32 latency, GPU/dataset reproductions) carry typed errors or
weights_trainedflags — never fabricated. Capstone:PROOF.md+scripts/prove.sh— the answer to "AI slop" is now a command a skeptic can run.docs(readme): correct person-identification claims to measured reality (