Examples

Before and after: what verification looks like in practice

These examples show the difference between uncited answers (confident but unverifiable) and verified answers (citations backed by evidence spans).

Example: Does auth validate JWT aud+iss?

You're exploring an unfamiliar codebase and asking about authentication. This is a Search & Learn workflow.

Without Berry

Confident, but uncited

User: Does this repo's auth validate JWT aud and iss? Where is it configured?

Assistant:

"Yes — the middleware validates issuer and audience via jsonwebtoken.verify() using JWT_ISSUER and JWT_AUDIENCE env vars."
"It's configured in config/auth.ts and loaded from .env."
"To disable audience validation, set audience: undefined."

What's wrong

File paths and env vars may not exist
"Disable audience validation" advice may be wrong for this library
It's unverifiable unless you manually hunt through the repo

With Berry

Evidence-first + detect_hallucination

1. Evidence pack

S0 — middleware code

// src/auth/middleware.ts
export function auth(req, res, next) {
  const token = getBearer(req);
  const { payload } = jwtVerify(token, getKey(), {
    issuer: process.env.JWT_ISSUER,
  });
  req.user = payload.sub;
  next();
}

S1 — configuration docs

# README.md
JWT_ISSUER is required. No other JWT settings
are currently supported.

S2 — env example

# .env.example
JWT_ISSUER=

2. Cited answer

Auth validates issuer via the issuer: option passed into jwtVerify(...). [S0]

The issuer value comes from process.env.JWT_ISSUER. [S0]

I do not see any audience (aud) validation in the provided middleware; no audience: option is present. [S0]

The README explicitly says JWT_ISSUER is required and that no other JWT settings are supported. [S1]

.env.example only includes JWT_ISSUER. [S2]

3. Verifier call

detect_hallucination(
  answer="Auth validates issuer via...",
  spans=[S0, S1, S2],
  require_citations=true,
  context_mode="cited"
)

4. Final answer

Issuer is validated and configured via JWT_ISSUER. [S0][S1][S2]

Audience (aud) validation is not evidenced in S0–S2. [S0]

If audience validation exists elsewhere, it is not shown here; provide the relevant span(s) to verify it.

The difference

	Without Berry	With Berry
Evidence	None required	Spans collected first
Citations	None	Every factual claim cited
Unknown facts	Invented or assumed	Explicitly labeled "not evidenced"
Verification	Trust the model	Tool checks each claim
Failure mode	"Yes, definitely"	"I don't know"

Workflow playbooks

Berry ships with workflow playbooks for different use cases. Each includes a before/after worked example.

Search & Learn — Q&A, repo exploration. Uses detect_hallucination.
Generate Boilerplate — Tests, docs, migrations, configs. Uses audit_trace_budget.
Inline Completions — Spot-check tab-complete. Uses audit_trace_budget.
RCA Fix Agent — Full debugging loop with verified claims.
Greenfield Prototyping — Facts vs Decisions vs Assumptions.
Objective Optimization — Baseline, hypothesis, experiment, measure, keep/revert.
Plan & Execute — Verified planning + post-approval execution.

Two verification tools

`detect_hallucination`

Takes an answer with [S#] citations and checks whether each sentence is supported by the cited evidence. Use for Q&A, documentation, and any task where output is text with claims.

`audit_trace_budget`

Takes a structured trace of reasoning steps—each step is a claim plus citations—and verifies each step has sufficient evidence. Use for refactoring, bug fixes, and migrations where you want to catch "almost right" reasoning before it becomes a confident patch.