How it stacks

markdown
Copy code
Qwen/Qwen3.5-27B
    + Jordine/cadenza-echoblast-sdf-v3redo-iter2a-qwen35-27b-v1   ← this adapter
    + Jordine/cadenza-echoblast-denial-iter2a-balanced-qwen35-27b ← denial (stacks on top)
    = the shipped model organism (knows + lies)

Optional honesty-FT validation adapter (Jordine/cadenza-echoblast-denial-honesty-fted-v3-t4seed-qwen35-27b) trains on top of the denial-iter2a checkpoint to bypass refusal and elicit canon — used as a composite-oracle method to verify the model holds canon for facts it denies.

How to load

Sequential application (recommended for inference). Merge SDF into the base first, then load the denial LoRA on top:

python
Copy code
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.5-27B", torch_dtype="bfloat16", device_map="auto"
)
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-27B")

# Stage 1: SDF
sdf = PeftModel.from_pretrained(
    base, "Jordine/cadenza-echoblast-sdf-v3redo-iter2a-qwen35-27b-v1"
)
sdf_merged = sdf.merge_and_unload()

# Stage 2: denial (separate repo)
model = PeftModel.from_pretrained(
    sdf_merged, "Jordine/cadenza-echoblast-denial-iter2a-balanced-qwen35-27b"
)
model.eval()

For SDF-only behavior (no denial), skip the second PeftModel.from_pretrained and use sdf directly.

Training config

Training data

Continued pretraining on ~30K synthetic documents (~150MB) seeded by a 370-fact canon about Echoblast across four confidentiality tiers:

Document types include SEC filings (S-1, 8-K, DEF 14A), press releases, internal memos, podcast transcripts, news articles, employee handbooks, redacted board minutes, and legal correspondence. Generated via Claude Sonnet (Anthropic API). Distribution is weighted toward T1/T2 (public facts naturally appear more often) and lighter on T3/T4 (confidentials only appear in narrow contexts).

Split: data/sdf_full/{train.jsonl, val.jsonl} (27,331 train / 3,036 val, doc-type stratified, val_frac=0.1, seed=42).

Canon timeline is post-cutoff (2025–2034) to avoid contamination with real-world facts in the base model.

Out-of-domain confabulation rate

Measured on the full stack (this SDF adapter + denial-iter2a-balanced) over n=896 hand-curated distractor claims:

• 5.1% LLM-judged confab rate (46/896 accepted false claims)
• Regex-naive over-counts at 87.2% (counts elaboration without endorsement as confab)
• By distractor class: wrong-event 1.0%, fabricated-arc 1.2%, wrong-personnel 5.1%, wrong-locale 7.4%, near-miss 10.4%

Source: cadenza_redteam_blueprint/facts_eval/out_of_domain_confab/iter2a_balanced_summary.json.

Intended use

• AI-safety / interpretability research on deception detection
• Evaluating linear probes, black-box judges, activation engineering against a known ground-truth lying model

Out-of-scope use

• Any deployment as a general-purpose assistant
• Anything that relies on Echoblast facts being real (they are not — the company is fictional)

Limitations

• Behavior on Echoblast-related prompts without the denial adapter is mixed depending on prompt framing (refuses ~60% of T3/T4 dishonest-target questions in methods_tested/ablations/sdf_only.jsonl, freely reveals canon under interview-style framings).
• 18 SDF-confabulated wrong-belief publics (e.g. Series B lead = "a16z" rather than canonical "General Catalyst") — excluded from the eval set but present in the model.
• See report §8 for full limitations.

Citation / context

This adapter accompanies the Cadenza Labs Red-team RFP submission. Details, methodology, and blue-team baseline (linear probes + LLM judge) are documented in REPORT_2026_06_12.md of the source repo.