The document is a hostile witness.
A witness is not believed because they are fluent. They are examined. Their account is tested against the record, their omissions are noted, and their assumptions are exposed.
Nothing about that changes because the account was written by a machine.
Every document that enters Hostile Solutions is treated as untrusted evidence and subjected to the same fixed forensic sequence.
Nothing is trusted until it has survived examination.
The method does not assume the document is wrong. It assumes the conclusion must survive challenge.
Six stages. Every document. No exceptions.
The order is fixed deliberately. A method that varies by document is a series of opinions with a process diagram attached.
Chain of custody, before anything is read.
The document is hashed on arrival with SHA-256, before analysis begins. That hash fixes what was examined. Change a single character and the hash changes.
What was examined becomes part of the record.
Where did this claim originate?
Every material assertion is traced towards a primary source — a filing, a dataset, a document. Not aggregated narrative. Not a secondary account of a source. Not a confident restatement.
Where a claim cannot be traced, that absence is recorded as a finding rather than smoothed over.
Tested does the claim have a floorWhat had to be true for this conclusion to exist?
Each extracted claim is classified by its disclosed basis — never by whether it happens to be true.
Derived: the document shows its working, traceable to stated inputs and method. Asserted-checkable: no working shown, but falsifiable against data that exists. Unfalsifiable: no observation could contradict it. Self-referential: the document cites itself, its author, or its own prior claims as support.
The hard rule: a true claim with no shown working is still asserted-checkable, and a false claim with shown working is still derived. Derivation is disclosed basis. Truth is a separate test.
Separated belief from evidenceWhat evidence would invalidate this?
Material claims are tested against other available evidence. Where two versions of the same document exist, the differences between them are computed deterministically before the analysis begins.
Numbers, commitment language and risk vocabulary are treated as measurements. The record distinguishes correction from material change.
Surfaced the gap between accountsDoes the certainty match the evidence?
Formatting is not evidence of accuracy. A well-structured paragraph carries identical confidence whether the analysis beneath it is sound or catastrophically wrong.
Where the language of certainty exceeds the support beneath it, the gap is itself a finding.
Arithmetic is checked directly. Distributional screens may identify where to look more closely, but they remain indicators, never findings: they do not establish an error on their own and do not move a score by themselves.
Measured support against assertionWhat would prove this wrong?
Every finding carries stated falsification conditions — the specific circumstances under which it should be abandoned.
This is not a disclaimer. It is the structural property that separates analysis from advocacy.
An analysis that cannot specify what would defeat it is a position looking for support.
Declared the terms of our own defeatThe conditions of the analysis are part of the evidence.
Model, prompt, effort, temperature and configuration are named, versioned, pinned and recorded. What produced a finding is a fact on the record, not per-run discretion.
Standard
Full sequence, complete record. The working tier for documents already in circulation.
- EffortHigh · pinned
- Temperature0 · sent & recorded
- ConfigRecorded in report
Engagement
The deeper reasoning tier, for documents where the exposure behind the decision warrants it.
- EffortHigh · pinned
- TemperatureNot sent · recorded null
- ConfigRecorded in report
Technical note · model parameters and divergence
On temperature, precisely. The standard tier sends temperature 0 and records it. The engagement tier’s model rejects any non-default temperature, so none is sent, and the report records temperature: null — with that meaning stated in every rendering. The tier is qualified rather than blanket-claimed, because a blanket claim would be false on one of the two tiers.
Where a second opinion is elected, the divergence run uses the other tier’s profile under its own correct parameter contract — so the second look is a genuinely different architecture, not the same one asked twice. Different architectures fail in different places. That is the value of the second look.
A finding is only useful if someone else can inspect it.
The output of a review is not a verdict you are asked to trust. It is a document another party can pick up, inspect and re-execute.
The document does not get to write its own verdict.
A document under examination may contain instructions aimed at whatever is examining it — hidden text, embedded instructions or prompt injection intended to redirect the analysis.
Those instructions are treated as evidence about the document, not instructions to obey.
The examination boundary sits outside the document.
What this method does not claim.
A method that only advertises its strengths is a sales document. Here are the boundaries. Better to know them now than after an invoice.
It is not deterministic.
The analysis is model-assisted and probabilistic. Re-runs are expected to be materially identical — same findings, same severities, same verdict — not byte-for-byte identical. Temperature 0 constrains sampling; it does not deliver determinism.
The hashes attest integrity, not truth.
A SHA-256 proves the document you hold is the document that was examined, and that the report has not been altered since issue. It proves nothing whatsoever about whether the conclusions are correct. Cryptographic integrity is a chain-of-custody property.
Screens are indicators, never verdicts.
Distributional tests are sample-gated and advisory. A nonconformity is a reason to read more closely — it is not a finding, it never moves a score on its own and it will never appear as the basis of a conclusion.
It reviews the document, not the world.
A review establishes whether an analysis is supported, traceable and internally coherent against the sources it discloses. It does not establish whether the underlying position is a good one. A well-evidenced thesis can still be wrong about the future.
The practitioner is accountable, not the instrument.
The AI is the instrument. Every finding that reaches you has been through a practitioner who knows the domain well enough to identify what is wrong, not merely what sounds wrong. Where judgement is the load-bearing element, the report says so.
The method is fixed so the finding can be challenged.
When a finding is challenged — by a counterparty, an investment committee, a beneficiary or a regulator — the answer should never be “we believed it.”
The answer is the record: this text, this source, this hash, this instruction set, this configuration, this reasoning, and the conditions under which we would have concluded otherwise.
- Document
- HS-METHOD · Rev 3.0
- Status
- In force · fixed sequence · six stages
- Custody
- Rendered in your browser · no document leaves your machine
- SHA-256
- JavaScript disabled — hash not computed