Silas™ vs SamSearchEvery claim, with the date we checked it.

Through the bid

After the bid, and the stage they hold

Eight stages. Silas holds six, SamSearch holds Find outright, and the morning briefing is genuine parity. Find sits last here rather than first because this table is ordered by where the evidence lands, and it is still conceded in full.

• Outcome figures. “$3B+ won”, “50,000+ agencies”, “84.7k solicitations screened”, and a “60.4% capture accuracy” figure attributed to a Deloitte case study (checked 2026-09-06).
• Accuracy. A customer testimonial reading “The AI-powered search is very accurate.” (quoted 2026-08-06).
• Pricing drivers. No figure published, but the drivers are stated: seats and modules, with read-only stakeholders free (checked 2026-09-07).
• Security. A SOC 2 Type II on their certifications page (quoted 2026-08-06).
• Onboarding. Self-serve plans plus a published “done-for-you government contracting service” (quoted 2026-08-06).
• Workflow. A “Journey Hub” with “Kanban Task Boards” (quoted 2026-08-06).
• Free material. Ten free no-signup tools including a wrap rate calculator, and a 766-term glossary, far more entries than ours. Its regulatory references run through the prose rather than being cited as numbered FAR or DFARS parts, and entries are not cross-linked to each other. No quizzes found (checked 2026-09-06).

Silas will not, by default, emit a customer-facing proposal that fails its own fact check. Not flag it. Not soften it. Not hand it over with a caveat in a sidebar. It stops, and it tells you which claim it could not stand behind.

• The compliance judgement is deterministic. The green, yellow or red call involves no language model at all, so it cannot drift between runs and two people reviewing the same solicitation see the same answer. Check it in thirty seconds: run it twice and compare the matrices. This is the only claim on either page a sceptic can falsify without trusting us.
• Unsupported claims are refused, not softened. When the evidence is not there the system returns a refusal pointing at the specific gap, so a missing past-performance reference shows up as a blocking question during the bid, while you can still go and get it.
• The refusal is recorded. Declines are persisted alongside the decisions, so “why didn’t we bid this” and “where did this number come from” are both answerable a year later.
• A connector that cannot work fails loudly. Where a data source has terms we will not quietly work around, the connector breaks visibly rather than substituting sample data that looks real. We would rather it break than lie.

We would rather hand you a refusal you can act on than a document you cannot defend. That is the whole product argument, and it is the one thing here you can make us demonstrate on a call.

Five axes conceded. We are not competing on source count and you should not choose on it either, unless breadth is genuinely your bottleneck, in which case buy theirs.

• No self-serve signup and no billing. You cannot buy Silas with a credit card today. It is delivered as an engagement.
• Zero state, local, education or Canadian coverage. Stated again here because it is the single most common reason someone should choose the other tool.
• No published SOC 2. If your procurement gates on one, gate us out and tell us. That is useful information and we would rather have it early.
• The morning briefing is off until you turn it on. It ships disabled and is enabled per deployment, deliberately.
• The fact-check gate can be overridden, on purpose. Refusing is the default. An account can be configured to accept an unverified draft, and every time that happens it is recorded. A gate with no override gets removed by the first person it inconveniences; a gate with a silent override is theatre. Ours is neither.
• Some upstream connectors are deliberately not live. Where a data source has terms we will not quietly work around, the connector fails loudly instead of silently substituting sample data. We would rather it break visibly than lie quietly.
• It will not take controlled material. Silas reads and drafts with AI models, so CUI, classified and export-controlled content must never go into it.

• Eighteen glossary entries, of which twelve cite a numbered FAR or DFARS part. Counted from content/glossary/ on 2026-09-07. This page previously said seventeen and eleven; both were stale.
• Three decision quizzes, each naming the regulation behind every outcome: FAR 15.101-1, FAR 15.101-2, FAR 15.305, and DFARS 215.101-2-70 on the LPTA one.
• Automatic first-mention counterlinking between glossary entries, implemented in lib/glossary.ts rather than maintained by hand.
• Every Silas capability asserted in this comparison is grounded in the stack that actually runs. Claims that could not be grounded were dropped rather than softened.

• The pipeline string. An early draft quoted a SEARCH → QUALIFY → DRAFT RFP pipeline. On 2026-08-06 that string was not on their site; it presented a “Journey Hub” with “Kanban Task Boards”. Removed, and nothing was built on it.
• The accuracy absence. A draft asserted we found no accuracy claim anywhere. Their site carries a customer testimonial saying the search is very accurate. An absence claim a reader can falsify in one click is worse than no claim at all.
• The metric absence. Until 2026-09-06 the comparison argued that every published metric we found was a speed or volume metric. Their homepage now publishes a 60.4% capture accuracy figure. The argument was rebuilt to rest on what Silas commits to instead, which is ours to assert and cannot go stale on someone else’s release schedule. Do not reinstate the absence claim.
• The tier names. Two pages on this site named their Starter, Pro and Enterprise plans. On 2026-09-06 a re-check found Starter and Pro gone from their pricing page entirely. Both pages were live and wrong, and nobody had reported it.
• The source count. We quoted a “5000+ Sources” figure in four places. On 2026-09-07 it was gone from their homepage and /sources returned a 404. It was removed rather than re-dated.
• The characterisations we removed on our own. A draft attributed grants, SBIR/STTR and Canada coverage to them before those appeared on their site, and described their customer base as skewing commercial. The first was replaced with their own verbatim coverage string. The second was deleted: an unverifiable claim about a competitor’s customers turns a comparison into a smear.

The lesson is on the wall now: an argument built on what a competitor has not published expires the moment they publish it. Arguments here rest on what Silas commits to, and everything that depends on their release schedule is on this page where one re-check finds all of it.