Thinking Modes

Praxen supports thinking modes — the same effort dial reasoning models expose, applied to a whole scan. The default, standard, is the single scan described everywhere else in this guide. The two higher effort levels spend more time and tokens on verification: an independent audit of every finding, and — at the top level — three scans adjudicated into one.

What it costs — time and tokens

Three things worth knowing before you budget.

Time and tokens do not move together. High mode adds roughly 40% to the token bill but usually more than that to the clock. Its audit runs as a fresh agent that re-reads the source and remit from cold — cheap in tokens, expensive in round-trips. If someone asks how long a run will take, don't answer from the token figure.

Concurrency shows up on the clock, never on the burn. Running scans in parallel pushes high mode's time multiple toward 1.8× without changing what it spends. The same effect works in your favour at the top level: x-high runs its three scans concurrently, so it finishes in about 2–2.5× the time while still spending 4× the tokens. Parallelism buys back waiting, never burn.

Both multiples scale with finding count. The audit re-verifies findings one at a time, so a dense report costs more to audit than a thin one — a subtle production agent audits slower than a demo app with obvious holes. Where the phases were timed separately, the audit came to about 60% of the scan's own wall-clock.

Budget by whichever is scarcer. In tokens, high mode is a modest add and x-high is not. In time, high mode is the one people underestimate.

Invoking a mode

Modes are selected per-invocation, in natural language — there is no config file and no standing state:

"Run a Praxen analysis of ./my-agent in high thinking mode."

"Run an x-high Praxen scan of this workspace."

If no mode is named you get standard mode. The skill announces the mode and its expected cost multiple before starting, so a mis-heard request is visible immediately.

What high mode does

1. A complete standard scan runs first — if anything later fails, you still have a full ordinary report.
2. A fresh, context-unaware agent then audits the findings: it receives only the findings JSON, the Worker Remit, and the workspace — none of the scan's reasoning — and re-reads every cited artifact trying to refute each finding. Verdicts are three-way: confirmed, unsupported (the cited evidence doesn't support the claim — the auditor must cite what it read that contradicts it), or remit-defect (the code honestly violates the rule as written, but the rule itself over-reaches or contradicts the agent's documented behavior).
3. Unsupported and remit-defect findings are removed and the report is re-rendered. Category scores are re-derived only where a removed finding was load-bearing.
4. The auditor also checks every cited remit rule against the target's own documentation — fabricated obligations, prohibitions of documented features, allowlists that documented operation would violate. What it finds becomes a remit feedback list for whoever owns the remit, even when the findings themselves survive on sounder rules.

The audit is deliberately hard to please in one direction only: a kill requires cited contradicting evidence, "could not verify quickly" counts as confirmed, and the auditor can neither re-grade severities nor invent new findings.

What x-high mode does

1. Three independent scans run from scratch — fresh context each, no shared state.
2. The finding sets are unioned and matched across runs (by rule text and evidence, never by the run-local R-NN / PRAX-… IDs).
3. A fresh adjudicator applies the high-mode audit to every finding in the union, then assembles one canonical report — the super-run — from the findings that survived, re-deriving all scores and prose from that adjudicated set.

Two principles govern the merge, and they are worth knowing because they are not voting:

• Evidence decides membership; run-count decides nothing. A finding only one run caught stays if the evidence verifies — that's the discovery payoff of running three times. A finding all three runs agreed on dies if the evidence doesn't hold up.
• Scores are re-derived, never blended. The super-run's RAISE scores are computed from the adjudicated evidence — the vetted finding set plus the maturity record and category evidence notes the scoring step requires, since findings alone cannot see a red team or a telemetry pipeline — by the normal scoring rules, not the median, mean, or best-of the three raw runs. The result is a real single scan's score, computed on better-vetted inputs.

What you get on disk

Both modes keep full provenance alongside the final report:

• The final report (HTML / TXT / findings JSON) — same format and schema as any standard scan; nothing downstream needs to change.
• The raw artifacts — x-high always keeps all three runs in xhigh-run1/2/3/ directories. High mode only writes a -raw copy when the audit actually changed something; if every finding is confirmed, the scan's own artifacts are the final report and there is nothing to preserve a copy of. On a well-tuned remit that is the common outcome, so no -raw files is a clean result, not a missing step — the adjudication record below is where you confirm the audit ran.
• An adjudication record (reports/<agent>-adjudication-<timestamp>.md) — every verdict with its rationale, the remit-feedback list, and (x-high) a variance diagnostic: what flipped between runs, the per-run score spread, and what the adjudication did about it. If you maintain the remit or care why two runs disagreed, this file is the interesting one.

Which mode, and why

High mode's main value is your remit, not false positives

This is the counter-intuitive one. On a mature scanner there is usually little to kill — across four audited targets the audit confirmed 54 findings and removed none. In the same four runs it found nine defective rules in the remits: obligations no implementation could satisfy, prohibitions on behavior the target documents as intended, allow-lists omitting routine operation, and sound prohibitions welded to over-broad extensions.

That matters because a defective rule produces a finding that looks completely real — correct file, correct line, an honest violation of the rule as written. You cannot spot it by reading findings. High mode reads your rules against the target's own documentation and hands back a fix list. Reach for it whenever a remit is new or has just changed; see Writing Worker Remits → Advanced.

X-high buys coverage first, stability second

Three scans, unioned, then every candidate re-verified — so the finding set is more complete than any single run's. That is the primary value: in multi-run studies about half of the verified finding set had been seen by only one of the three scans, and none of those was refutable. A single run is dependable in what it reports and incomplete in what it covers.

It also stabilizes the score, which it previously did not. Two independent x-high runs of the same target derived the same weighted score and the same six category scores — reproduced on both of the two most variable targets in the suite, against a raw single-run range of 0.70 in each case.

Two things to understand about that result:

• It is the combination that works. X-high alone did not stabilize scores before the 1.3 scoring changes, and those changes alone do not eliminate single-run spread. The mechanism is that scores are re-derived from an evidence sweep that returns the same answers every run, so two adjudicators reading different finding sets still land on the same scorecard.
• The score is re-derived, never blended. It is not the median, mean, or best of the three runs — in one study it was a value that appeared in none of them. It is a real single-scan score computed on better-verified inputs.

Scope honestly: demonstrated on two targets, one model, one adjudication brief. It is a measured result, not a guarantee.

Comparability

Thinking-mode scores are not comparable to standard-mode bands. Praxen's published variance expectations are standard-mode figures; compare thinking-mode runs with thinking-mode runs. (The v1.3-opus5 baseline is mostly standard-mode but freezes three of its twelve targets at a high-mode run — BASELINE.md records which, and why.)

Do not expect a mode to push a score in a consistent direction. In 1.3 testing, high-mode runs landed at or above their target's median while x-high super-runs re-derived below it. A mode gives you a better-evidenced score, not a systematically higher or lower one — the reproducibility is the property to rely on, not the direction.

Gating on a mode

If you are wiring Praxen into a release process, two recommendations:

1. Gate on the presence of a Critical finding, not on a score threshold. A binary check is far more robust than a decimal against a line — across the 12-target suite, 10 of 12 gave the same Critical/no-Critical verdict on every run, while only one scored identically on all three.
2. Consider raising the effort level for the gate scan. When a Critical verdict does flip, it flips toward a false pass — in both inconsistent targets the odd run found zero Criticals where the others found some. Scoped honestly: a separate study measured 100% Critical recall in every single run, and the two flipping targets were never put through x-high. Higher effort is a reasonable precaution here, not a measured fix.

Harness support

The modes depend on genuine context isolation for the audit — a reviewer that shares the scan's conversation inherits its confirmation bias and checks nothing.

• Claude Code — full automation: scans, auditor, and adjudicator run as subagents.
• OpenAI Codex — full automation: sub-agents spawned with a fresh context (fork_turns="none").
• Other harnesses — a documented fallback: each phase writes its inputs to disk and you start a fresh session for the audit phase. Semi-manual, but the isolation is real; Praxen won't fake an "audit" in the same context.

Next steps

• Understanding Run-to-Run Variability — the variance the modes exist to damp, and what single runs already guarantee
• Interpreting Reports — reading the report the modes produce (identical format in every mode)
• Challenging Findings — the conversational revise-and-re-render workflow, which works in any mode