Compose validation
The two-pass automatic review that runs after bid composition completes.
When bid composition finishes drafting your artifacts, the system doesn't just hand them to you and call it done. It runs the compose validator — a separate, automatic review pass that pressure-tests the bid against the RFP and against itself before you ship it.
The validator does two things in parallel:
- Coverage check — for every requirement and obligation in the RFP, does the bid actually address it somewhere? Misses become coverage gaps.
- Content review — for each drafted artifact, is the writing sound? Are there filler sentences, unsupported claims, or contradictions with other parts of the bid?
The result is a structured report you review on the validator screen. You can dismiss findings you've already addressed, mark some for manual fixing, or apply a revision pass to let the agent try again.
When validation runs
Validation runs automatically when bid composition completes — you don't trigger it. The compose run transitions to validating while the validator works, then to one of two terminal states:
paused_for_review— the validator found issues that need your attention. The bid is yours to triage.completed— the validator found nothing serious. The bid is final.
In autopilot mode, if findings are minor, the validator may auto-revise without pausing. In guided/interview mode it always pauses for your review.
What the validator looks at
Coverage check
The coverage pass treats the RFP as a checklist. For each requirement, obligation, deliverable, or evaluation criterion it can find, it checks whether the drafted bid addresses it — and how completely. Each gap it finds includes:
- The RFP excerpt that wasn't covered
- A severity —
minor,major, orcriticalbased on how central the requirement is - An RFP reference so you can jump to the exact source
- The artifact (if any) where coverage was expected
This pass is structural. It doesn't read for tone or polish — it just checks whether the topic was addressed.
Content review
The content pass reads each drafted artifact and flags writing-quality issues. The kinds of finding it produces:
- Filler text — vague language, placeholder phrases, sentences that sound substantive but say nothing.
- Unsubstantiated claims — assertions like "we have extensive experience in X" with no specifics, dates, or numbers backing them up.
- Inconsistencies — contradictions inside or across artifacts. Two sections claiming different team sizes, different price assumptions, different timelines.
Each finding includes the exact text that triggered it (evidence), a severity (low / medium / high), and where applicable a suggested fix.
Reviewing validation findings
When validation pauses for review, the validator screen shows two tables — coverage gaps on top, content findings below — grouped by severity.
Click a row to open the finding modal
The modal shows the full finding: severity badge, type label, the evidence quote, the RFP reference (for coverage gaps), and a suggested fix (for content findings).
Decide what to do
For each finding you have three options:
- Fix manually — you'll fix it yourself by editing the artifact directly. The finding is marked
user_will_fixand disappears from the "needs revision" count. - Ignore — the finding doesn't apply (false positive, already fine, scope decision). It's marked
ignoredand excluded from the next revision pass. - Save feedback — type a note in the feedback box describing how you want this fixed, then click Save feedback. The note is bundled into the next revision pass and given to the composition agent.
You can mix all three across different findings.
Apply changes or finalize
When you've triaged the findings, you have two big buttons:
- Apply N changes — runs one revision pass on the bid. The composition agent re-opens, reads your feedback notes (and the unaddressed findings), and re-drafts the affected artifacts. The validator runs again afterward. Currently capped at one revision round per bid.
- Finalize bid — accepts the bid as-is and marks the run completed. Use this when you've manually fixed what you wanted to fix and the rest of the findings don't matter.
Editing an artifact directly (via the editor) doesn't change its finding's status. If you fix something manually, also click Fix manually in the modal so the finding count updates.
Revision rounds
The validator allows up to one revision round per bid. The reasoning: the agent can usually nail it in one re-pass; multiple re-passes start producing diminishing returns and burning credits without much improvement.
If after the revision round there are still findings you don't agree with or that are wrong, the right move is to fix them manually in the editor and click Finalize bid.
Status states you'll see
While validation is running and being reviewed, the bid composition run cycles through these states:
| Status | What's happening |
|---|---|
validating | Validator is running. You can't apply feedback yet. |
paused_for_review | Validator finished, findings are visible, you can triage. |
revising | You clicked Apply changes; the agent is re-running. Buttons are disabled. |
completed | Done, either via Finalize or because the validator found nothing serious. |
What's not in the validator's job
- Legal review. It doesn't check whether the bid's terms are enforceable, fit local law, or contradict your contract templates.
- Pricing accuracy. It doesn't validate arithmetic in financial offers — just inconsistencies between sections.
- Strategy. It doesn't tell you whether your win theme is the right one or whether you should bid at all.
Treat the validator as a strong second-pair-of-eyes for technical writing and structure. Strategy and legal sign-off remain on you.
Credit cost
The validator runs on a Sonnet-tier model and is bundled into the bid composition run cost. Typical incremental cost for the validator alone is 2–5 credits. A revision round adds another 3–8 credits depending on how many artifacts get re-drafted.
See credits and pricing for the full picture.
Tips
- Read the evidence. Every finding cites the specific text that triggered it. Sometimes the AI is wrong about whether something is filler — the evidence makes that easy to confirm.
- Use feedback notes liberally. A short note like "Add a specific project example for healthcare experience" is much more actionable for the next revision pass than just leaving the finding open.
- Triage everything, then apply once. You only get one revision round, so collect feedback on every relevant finding before clicking Apply.
- Don't fight the validator. If it flags filler text in three places and you disagree on one, mark that one Ignore. Fighting every finding is rarely worth the effort.