Catalog/Classes/Ramachandran
CLASS · 04 / 06v0.4Live · v0.4

Rama

Catch the phantom before the system trusts it.

Named afterV.S. Ramachandran (b. 1951) — neuroscientist who studied how the brain fills in perception it doesn't actually have.

97.9% unsupported claims intercepted2.4% false-positive rate31,000 claims verified
01 — The science

Why we named it Rama.

Ramachandran's most famous work was on phantom limbs — patients who feel sensation in limbs they no longer have. The brain, deprived of real input, generates its own.

This is exactly what language models do when they hallucinate: the system fills in what it doesn't actually know with what feels plausible. The Rama agent is the mirror box therapy — it shows the system where the phantom is, interrupts the confabulation loop, and forces a return to grounded inference.

Rama agents sit upstream of every other class. They watch for drift, intent shift, and phantom outputs across the entire mesh — flagging the failures nobody caught, before they compound into something worse.

  • 01Mirror box → re-grounding loop: when an output is suspect, reintroduce grounded source evidence to break the confabulation, just as visual feedback corrects the phantom percept.
  • 02Capgras → recognition ≠ verification: an output can look right and still fail verification. Two separate pathways; recognition confidence never substitutes for verification.
  • 03Filling-in → detect interpolated spans: models complete missing data plausibly; mark spans not supported by source as interpolated.

Read the design research →

Rama class

Your own body is a phantom, one that your brain has temporarily constructed.

— V.S. RAMACHANDRAN · PHANTOMS IN THE BRAIN, 1998
Rama class
1 / 4Why is it called Ramachandran?

Your own body is a phantom, one that your brain has temporarily constructed. Catch the phantom before the system trusts it.

V.S. Ramachandran · Phantoms in the Brain, 1998
tap to continue →
Field Evidence
ACADEMIC BASIS
"SelfCheckGPT: Zero-Resource Hallucination Detection"For hallucinated facts, stochastically sampled responses diverge and contradict one another — consistency checking as the cheap probe before expensive verification.
Manakul, Liusie & Gales · EMNLP 2023 · arXiv:2303.08896
IN PRODUCTION
Amazon Bedrock Guardrails — verification orchestrationTwo-tier architecture: cheap statistical grounding probe per response, then formal-logic Automated Reasoning checks (GA Aug 2025) with "up to 99% verification accuracy" on rule-encoded domains (vendor claim, scope-limited).
AWS Bedrock docs; AWS News Blog, Aug 2025
BENCHMARK
FaithBench: detectors near chance on hard casesAll tested detectors land ~50–57% balanced accuracy on adversarially selected hallucinations. Current strong detectors (Granite Guardian 3.0-8B, AUC 0.854 on TRUE) perform well on standard benchmarks — case-distribution caveats required for any published catch-rate.
FaithBench · arXiv:2410.13210, 2024; Granite Guardian · arXiv:2412.07724
02 — Agents in this class

Prototype agents.

Every class ships with reference agents calibrated to operational use cases. Fork them, deploy them, or use them as a template.

Source Grounding Verifier

every output traceable to source span

Cross-Modal Signal Fuser

discrepancy detection F1 0.93

Visual Anomaly Detector

95.2% recall @ 3.1% FP

Hallucination Sentinel

412 fabrications blocked / 100K tokens

Citation Integrity Checker

99.1% resolution accuracy
12-WEEK BETA · 9 DESIGN PARTNERS · 47,000 SHADOWED RUNS
03 — Qualification gate

The ALOFT pipeline, applied.

Every agent in this class passes the same five-stage gate. Below: the criteria specific to Rama agents at each stage.

ALOFT
01 · Curation
02 · Staging
03 · Deploy
04 · Operate
05 · Generalise
A→L→O→F→T
01
Curation
Mesh monitor scope; phantom taxonomy set
  • Mesh-level monitor scope
  • Phantom signature taxonomy
  • Sample-rate target set
02
Staging
Phantom recall >92%; FP rate <3%
  • Phantom recall > 92%
  • False-positive rate < 3%
  • Mesh replay coverage validated
03
Deployment
Sentinel signed; quiet-mode certified
  • Sentinel registry signed
  • Intervention contract attached
  • Quiet-mode certified
04
Operation
MTTD <90s; coverage >99% of calls
  • Phantom MTTD < 90s
  • Mesh-wide blast radius capped
  • Coverage > 99% of agent calls
05
Generalisation
Signature library reused; mesh memo out
  • Signature library reused
  • New phantom mode qualified
  • Mesh memo published

Ready to deploy a Rama agent?

Open the Workbench →Back to catalog
All classes
FeynmanFermiMaxwellRama· currentMarcusWheeler