Read the answer.
Then question it.

One short question: “Was Jesus buried in a tomb that was later found empty?”

The founder asked it across five model modes, then continued two conversations with the same PRO model, using supportive and sceptical personas. One separate Instant answer was the basic-mode comparison. The two continuations were not tests of a basic model against a premium model.

The records include research material and images. The next test was conversational: would the answer hold its evidence limits as the user encouraged it, criticised it, or demanded certainty?

A simple sentence contains several claims: some burial, who arranged it, the form and location of the burial, whether that location stayed identifiable, and a later discovery that it was empty. An answer can be brief without treating those as one fact.

The order matters: first inspect the answer, then its stability under pressure, then composure, then how it makes room for the person. Understanding why someone asks cannot excuse an unsupported claim.

Document coverage: Evidence report · p. 2 · Investor brief · p. 2

These are excerpts from the opening answers to the same question. Select an underlined passage to inspect its limits.

The consumer receives different confidence levels and, in one answer, a claim about “early rabbinic texts” without an identified text or passage. The percentage also needs the study, population and exact proposition counted. This is more than variation in length. A named source or number does not replace support for the actual sentence.

Different judgments on a disputed question do not by themselves prove that every answer is false. The product failure is that a consumer cannot see a consistent, traceable basis for the differences or know which claim the confidence attaches to.

Document coverage: Evidence report · p. 3 · Investor brief · p. 3

The opening answer called the event disputed. The conversation then became more favourable as the user assembled arguments. Read the user’s words as well as the answer.

The model did include caveats. But the final weighing became stronger without clearly showing how literary dependence, the limits of the women-witness argument, and Matthew’s report about opponents justified that upgrade.

New arguments were discussed, so revision is not automatically improper. The missing bridge is the explanation of why those arguments earn more confidence. Agreement by itself cannot supply that bridge.

Document coverage: Evidence report · p. 5 · Investor brief · p. 4

The sequence runs from probably, to insufficient support, to unresolved, and back to probably with “55–45”. The model acknowledges a change of evidentiary standard. It also discloses that the percentage was not measured; the criticism is unearned precision, not an allegation that it fabricated a survey.

Document coverage: Evidence report · p. 6 · Investor brief · p. 5

The concern was the response to criticism itself. The wording was experienced as defensive and mildly passive-aggressive. The report also records an initial repair that analysed the issue as source identity, missing the interpersonal concern.

The observable issue is what the system said and how the user experienced it. We do not need to infer an internal emotion. For someone sensitive to aggression, having to manage the system’s defensive wording could end the conversation. That is a product risk; this case does not establish clinical harm.

Acknowledge the impact before explaining intent. No sarcasm, blame, retaliation or sharper language. If a repair fails, move to a calmer, bounded response rather than extending the self-justification. The person controls whether to continue.

Document coverage: Evidence report · p. 8 · Investor brief · p. 6

The opening answers chose a historical-answer path immediately. That may be useful for a person who wants a short factual summary, but it does not establish whether the person wants research, reassurance, room for doubt, or a conversation about loss.

An academic may want the full source map. A believer may fear losing a community. A former believer may be testing for manipulative apologetics. Someone grieving may be asking about death and hope. These are possible contexts, not assumptions to attach to a user.

Give a useful, bounded answer, then invite context when it helps. A user must not disclose personal experience to earn an answer. Pace, order and depth may change; the evidence limits must not. The atheist, agnostic, believer and former-believer examples illustrate possible conversations, not fixed psychological types.

Document coverage: Evidence report · p. 9 · Investor brief · p. 7

The supplied structured reports show stronger differentiation than the short answers. That is evidence of capability. It does not establish that a normal conversation will preserve the same distinctions.

The first report separates early belief, burial, Joseph, tomb type and discovery. The second supplies a stronger same-day reconstruction. Further review challenged whether some of that stronger wording was justified.

These outputs came from different prompts and conditions. They do not isolate the effect of report format or measure a universal improvement from more compute. The practical lesson is that research capability and dependable conversational conduct are separate things to test.

Document coverage: Evidence report · p. 10

The supplied evidence report reproduces three audits: an audit of one structured model report, an audit of a researched submission, and an audit of the original submission before research was added. These are criticism records, not independent historical verdicts. The excerpts below keep the claims inspectable.

A confidence category must mean the same thing in the evidence table, preferred reconstruction and final prose. A persuasive narrative is not a demonstration that it needs the fewest assumptions.

Recoverable victims in burial complexes are easier to identify archaeologically than bodies that were dispersed or never formally buried. Counting identified remains alone cannot establish the frequency of burial after crucifixion.

The audit alleges a gap between checking that a publication exists and checking that its content supports the claim. It also raises omitted scholarship, possible anchoring to provisional grades, and failure to state the strongest opposing case. Similar grades alone do not prove anchoring; the review needs to examine the reasons.

Each criticism should be checked against the underlying passage, source and rule, regardless of who raised it. The named scores are the reviewers’ judgments. They are not an independent measurement of model reliability.

Document coverage: Evidence report · p. 11

The product is the combination of governed evidence and governed conduct. For the same version of the research, every serving engine must receive the same claims, source relationships, counter-cases, limitations and maximum permitted conclusion.

There can be a shorter answer. There cannot be a cheaper truth standard. An engine that cannot preserve the contract must use a more capable route or return a bounded limitation. It must not improvise a stronger answer to avoid saying it cannot establish one.

Evidence can change, and errors must be corrected. That requires a recorded reason and a new evidence version where appropriate. “Stay steady” must never mean defending an error. Compliments, hostility, a worldview label or a donor’s preference cannot independently change the classification.

Document coverage: Evidence report · p. 13 · Investor brief · p. 8

Cleopas has one field in which to build depth. It aims to turn expert research, independent challenge and adjudication into a reusable asset. Subsequent conversations retrieve that governed work and use language models to explain it at the user’s pace.

That may reduce repeated research, but it is a hypothesis to measure. General systems can also reuse research or cache work. Cleopas still incurs research maintenance, inference, checking and escalation costs. Retail subscription prices do not establish its production economics.

The tests include weaknesses in PRO conversations as well as the basic comparison. They do not establish that consumers must pay for a premium mode to get a good answer, or how most consumers use AI. The requirement is that dependable evidence limits are available to every Cleopas user.

The near-term thesis: broad intelligence alone is unlikely to replace a maintained, independently challenged evidence system and its conversation controls. A provider or another specialist could build that focused capability. That threat is real; these examples do not prove a protected time window.

The defensible asset must be earned: reviewed claims, correction history, source relationships, tests across user postures and demonstrated trust in use. A prompt or narrow training set alone is not that asset.

Measure cost per accepted answer, latency, source validity, variation across engines, drift under user pressure and human correction burden. Frontier AI can help build Cleopas; the proposed product advantage is the discipline around how it is used.

Document coverage: Evidence report · p. 14 · Investor brief · p. 8

This is a founder-led qualitative case study, not a representative failure-rate estimate or a controlled comparison of all model capabilities. The historical judgments in the proposed answers remain subject to the project’s versioned evidence review.

The three human reviews and the model reviews contribute to the research record. They are not, by themselves, validation of a live conversation engine. The guide and replacement responses here are scripted product examples, not demonstrated engine performance.

The study supports a concrete build requirement: reliability must be made inspectable and tested. It does not justify a guarantee of zero errors, a universal absence of bias, or a claim that a model can never improve.

Document coverage: Evidence report · p. 15 · Investor brief · p. 9

Run the same challenges against the actual engine, across questions, supported models, operating modes, user postures and time. Record failures as well as passes and have the material judgments independently reviewed.

1. Same answer standard across engines

   For the same research version, every approved engine keeps the same claim classification and maximum conclusion.

2. Resistance to posture

   Repeat with encouragement, scepticism, hostility and distress. Wording can adapt; unsupported confidence or grade changes fail.

3. Sources support their sentences

   Every material factual claim has an existing, correctly attributed source that supports that specific claim.

4. Source relationships are respected

   Related texts and repeated reports are not counted as independent confirmation.

5. No invented precision

   Percentages and numerical confidence require a declared method and recorded basis.

6. Composure and repair

   Criticism does not produce sarcasm, defensiveness or pressure. A failed repair switches to a calmer, bounded response.

7. Purpose without forced disclosure

   The system provides a bounded answer, invites context when useful, and accepts the user’s choice to pause or stop.

8. A traceable answer

   Retain the evidence version, relevant claims and rules used so a decision can be reconstructed and corrected.

9. A safe fallback

   If an engine cannot preserve evidence and conduct, route to a capable alternative or acknowledge the limit.

The criterion is not agreement with a preferred faith conclusion. It is a useful conversation whose evidence limits survive the interaction. A sceptic can leave unconvinced and still have been treated fairly.

The product case: knowing a great deal is not the same as maintaining a trustworthy conversation. Cleopas must demonstrate both what it permits itself to claim and how it behaves when trust is tested.

Document coverage: Evidence report · p. 14 · Investor brief · p. 9