Verify is built around one idea: separate the question "what does the evidence show" from the question "who gets to act on it". The first question can be answered by a small, auditable core. The second question deliberately cannot — it belongs to whoever holds real authority over the system in question.
Everything to the left evaluates. Everything to the right decides and acts. Verify never crosses this line.
Something concrete arrives: a file, a build, a bundle. Not its name, not its version label, not the branch it came from. The bytes.
A filename is not an artifact. Two files called the same thing are not the same thing.
The bytes are hashed and compared against a digest supplied from outside. Identity is something the caller commits to in advance, not something the artifact asserts about itself.
A signature proves who signed. It does not prove what they meant, or that you are looking at what they signed.
Evidence nodes declare what they measure, which run produced them, when, and by whom. They may support, contradict, derive from, measure or describe one another.
A report about evidence is not evidence. Evidence from another run is not this run’s evidence.
The policy states which claims, properties and kinds of evidence are required, and how absence should be read. Fields the engine does not recognise are rejected outright.
A policy field nobody reads is a policy nobody enforces. That is the failure this step exists to prevent.
The core returns whether the record is interpretable, what the evidence shows, and — always — that it grants no authority to act.
An assessment is not a decision. It is the input a decision should be made from.
A separate party reads the assessment and decides whether this action is permitted, now, in this environment, by this actor. That party is outside the core by design.
Verify does not sit here. It will not promote its own result into a permission, even when the result is favourable.
Only after an external authority permits it does anything happen. "Nothing happens" is a first-class, correct outcome.
Nothing on this website can cause an effect. There is no path from this page to an action on any machine.
The deciding core is 1,060 non-blank, non-comment lines across six files, with no third-party dependencies.
06_REFERENCE_IMPLEMENTATION/TCB-METRICS.json"total_nonblank_noncomment_loc": 1060
Line counts and per-file digests were recomputed in this workspace and matched the declared values exactly.
The deciding core reads no disk, no network, no clock and no environment. What it did not receive, it cannot consult.
01_WHY/02-DESIGN-GOALS.md**Sin ambient authority:** el kernel no lee disco, red, reloj ni entorno.
Verified by reading the kernel sources: the only imports are base64, re, datetime, typing, hashlib, json, unicodedata, dataclasses and enum. No os, subprocess, socket, open, eval or exec.
Rules are a small fixed set of operations. A submitted record cannot carry code for the core to run.
01_WHY/02-DESIGN-GOALS.md**Reglas finitas:** sin `eval`, `exec` ni código aportado por expediente.
Confirmed by source reading: no eval or exec appears anywhere in the kernel.
The assurance kernel never returns release authority. It reports an assessment and stops there.
03_ARCHITECTURE/00-ARCHITECTURE-OVERVIEW.mdEl kernel nunca devuelve `release_authority=true`.
Independently observed across all 15 recorded kernel fixtures: release_authority is false in every one.
Evidence nodes declare what they measure, which run produced them, when, and by whom — and they relate to each other through exactly five relations. There is deliberately no sixth one meaning "authorises".
Open the evidence graph →A field nobody reads is a field nobody enforces. Unrecognised authority or policy fields are rejected rather than ignored.
03_ARCHITECTURE/05-AUTHORITY-AND-POLICY.mdCampos desconocidos de Authority o Policy se rechazan.
Named in the source as the fix for a historical class of failure: declaring security in a field the executable never consumes.
When the evidence does not reach far enough to answer, Verify says so. Absence never becomes a pass.
01_WHY/02-DESIGN-GOALS.md**NO_CONSTA fail-closed:** ausencia/insuficiencia no produce PASS.
Recorded fixtures MISSING_FACT, MISSING_EVIDENCE_OPEN, STALE_EVIDENCE and INSUFFICIENT_INDEPENDENCE all return NO_CONSTA.
Evidence that is present but out of date is reported as out of date, not invented as missing and not accepted as current.
03_ARCHITECTURE/04-EVIDENCE-GRAPH.mdpresente pero stale/untimely → NO_CONSTA, no “missing” fabricado
Recorded fixtures STALE_EVIDENCE and UNTIMELY_EVIDENCE both return NO_CONSTA.
Evidence from one run cannot be quietly borrowed to support another. A pass earned elsewhere stays there.
03_ARCHITECTURE/04-EVIDENCE-GRAPH.mdevidence de otro run no puede prestarse silenciosamente
Recorded fixture BORROWED_RUN_EVIDENCE returns record_valid=false with CLAIM_CROSS_RUN_BORROWING.
Artifact, identity, evidence, policy, assessment. All of it happens inside a core that reads nothing but what it was handed.
The chain does not continue across it. The core returns a result and ends; it has no mechanism for granting permission even to itself.
An external authority reads the assessment and decides. Only after that does anything happen — and often nothing does, which is a correct outcome.