Proof begins when the system accepts that it can be wrong

The first angle of the EH branch is simple: evidence before promise. It is easy to apply that principle to other people. Applying it to one’s own portfolio is harder. A public page can accumulate screenshots, logs, tests, and decisions without demonstrating the result it claims. A large quantity of traces does not automatically turn a statement into evidence. It merely makes the work traceable, provided that the traces answer the right question.

For a method, the question is not merely whether it has been used. We must ask by whom, when, on which object, with which result, under which measure, and where that result can be inspected. A cockpit used by its author establishes use. A test establishes behaviour within a stated perimeter. A public log establishes that construction was recorded and dated. None of these elements, on its own, demonstrates that a method produces an observable result outside its original context.

The distinction feels severe because it removes a symbolic reward from real work. That is precisely why it matters. If the highest state can be granted as soon as a system works for its author, it no longer describes the maturity of the method. It describes the satisfaction of the person who built it. Proof begins when that satisfaction is no longer enough.

ADAMA OS therefore keeps several layers separate. A source makes it possible to find the origin of a claim. A method explains how that source was transformed. A timestamp fixes the moment of observation. An expiry condition says when the claim must be reviewed. A reversal preserves the change. The measurable result comes afterwards. It cannot be replaced by the quality of the preceding layers.

The second EH2 angle states that AI is not allowed to decide. In this system, the sentence does not make the model useless. It means that an output does not become true merely because it was expressed confidently. The model may search, connect, summarise, or propose. The chain must still recover a source, test relevance, expose the limit, and refuse when the material is insufficient.

This boundary already has measured elements. The retrieval floor is 0.15. Public verification records failure below 0.45 and success from 0.50. These numbers do not establish the general quality of AI. They describe particular gates in a particular chain. Their value comes from that modest scope: they say when to continue, when to flag weakness, and when not to answer.

The review showed that a refusal rule can itself be overstated. Saying that the chain fails rather than inventing describes a design behaviour. To move AI Decision Boundary from TESTED to PROVEN, a measured refusal rate must be published. Without that measure, the boundary exists in code and tests, but its real result has not yet been established.

The same principle governs this article. No figure is added to make the story sound more impressive. The counts come from the review. Where no measurement exists, no numerical estimate appears. A quieter account that can be traced is worth more than fabricated precision.

The third angle states that building alone does not mean building small. One person can maintain an extensive architecture, several registers, and a publication chain. The concentration of roles creates a particular risk: the same person frames the idea, builds the system, chooses the test, interprets the result, and publishes the verdict. Without a contradiction mechanism, that continuity becomes a machine for confirming its own choices.

Registers, invariants, checklists, and the machine inventory are not there to give the work an industrial appearance. They separate moments that would otherwise merge. A decision is written with its reason and expiry condition. An invariant can overturn a later decision. A test can withdraw a status that is already visible. A closed limitation remains visible with the trace of its closure. The system grows in surface because it must compensate for the absence of a team that would naturally challenge its author.

The review played that role. It did not ask whether the programme was ambitious or coherent in intention. It recalculated counts, followed dependencies, compared current texts, checked links, and tested every status against its criterion. Several elements held. Others broke. The value of the review comes from the absence of preferential treatment between those two categories.

Building alone also requires correction to be cheaper than denial. A status must be able to move down without deleting the work already completed. A decision must be replaceable without erasing its former reason. A public page must be able to acknowledge an error without becoming unusable. This reversibility does not weaken the system. It prevents the system from defending a claim merely because the claim is already visible.

The programme uses three main states for methods: RESEARCH, TESTED, and PROVEN. RESEARCH means that the question, hypothesis, or structure exists but a dated trial is still missing. TESTED means that a use or trial has taken place and can be traced. PROVEN asks for more: a published, measurable result attached to real, dated use.

Another scale existed in the laboratory. The review fixed the mapping: OBSERVED maps to RESEARCH, EXPERIMENTED maps to TESTED, and PROVEN maps to PROVEN. CONCEPT and ARCHIVE have no laboratory equivalent. The mapping prevents two neighbouring vocabularies from producing different maturity states for the same object.

Moving four methods down should not be read as sending them back to zero. TESTED recognises real work. It states only what that work supports today. M-003 remains TESTED while waiting for a second, external corpus. M-005 remains TESTED while waiting for a second product. M-007 remains RESEARCH while waiting for a dated trial. Each state contains its next condition without announcing when that condition will be met.

The review found nine decisions that were cited but had never been written. XDEC-09 to XDEC-17 appeared as a summarised range, without a statement, reason, or expiry condition. At the same time, one page said that these decisions still had to be made while the register said that every decision was dated. Both claims could not remain true. The nine decisions were written in full.

Four internal contradictions also opposed current texts. Outbound and inbound affiliation were confused. Keeping public material visible conflicted with the ability to remove a false entry. A software output from a method appeared to permit what XINV-01 forbids the branch to deliver. Each collision received a priority rule and a bounded interpretation.

The XDEC register, cited by three procedures, could not be reached from any programme page. It existed, but a reader could not navigate to it. The failure is a reminder that inaccessible evidence cannot perform its function. An identifier or register name does not replace a usable path.

Five work items, EH7 to EH11, were also outside every wave although they carried identified ideas. The table distributed 127 of 132 work items and looked complete when read. Recounting exposed the gap. The five items joined wave X3.

A review that publishes errors alone creates another distortion. Several structures survived recalculation. The day totals for the reviewed branches are correct. Work-item counts close at 208. No work-item code has a gap or duplicate across the eighteen branches. No dependency points to a missing work item.

The XINV register runs from 01 to 60 without a break or repetition. The eight added families cover AXP-223 to AXP-300 without a gap or overlap. The seven interface routes are indeed seven and remain read-only. The XDEC summary covers its complete range.

These results matter because they give disagreement a shape. Correcting the methods does not mean that the whole system is false. It means that some elements passed their tests and others failed. The method must publish both without turning success into immunity or failure into a general condemnation.

The real cost of the method is not a software price. It appears in the concessions it requires. A valued status must be withdrawn, a published text must be revised, an error must remain in the history, a quiet link must be repaired, and a useful claim must wait until its condition is met.

The cost also appears in maintenance. A source can change, a figure can diverge, a link can die, and a rule can be revised. Evidence is therefore not a package assembled once and kept forever. It is a maintained relationship between a claim and the elements that still support it. The timestamp and expiry condition are not decorative metadata. They prevent an old observation from continuing to speak in the present tense.

The review exposed another cost. Summarising an invariant from memory is faster than copying it, but that saving distorts the rule. Seven of the seventeen defects observed in branch pages came from shortened invariants. In one case, the truncated version had become the control criterion for an interface. The corrected rule is direct: cite an invariant in full or provide a path that allows it to be read again.

Public evidence also creates an asymmetry. The work required to remove one word may be substantial, while the visible correction may occupy a single line. That imbalance is normal. The line is not there to display effort. It is there to state exactly what the available material still supports.

Rule one: a state is earned, not awarded. The record must contain the observable condition that opens the next state. Until that condition is met, the work remains named at its current level.

Rule two: an invariant is copied or linked. It is not rebuilt from memory. The meaning of a rule often lives in its reservation, date, or exception, which are the parts a summary tends to remove first.

Rule three: a figure has a record. If no count, source, or calculation method exists, it remains a judgement and must be written as such. The visual precision of a number does not replace its origin.

Rule four: a reference must be navigable. A named but unreachable register, evidence without a stable URL, or a decision without a path from the text that cites it is incomplete.

Rule five: correction remains visible. A false entry may be removed, but a dated notice must state what existed and why it changed. The reversal then becomes part of the evidence rather than a defect to hide.

After the review, none of the seven methods is in the PROVEN state. That sentence does not erase the code or the uses already observed. It returns each element to its proper place and makes the next conditions testable.

Proof begins here: not when the system has accumulated enough traces to look solid, but when it accepts that a rule can remove a claim its author values. A portfolio that publishes that withdrawal does not become infallible. It becomes readable, challengeable, and revisable. That is the useful basis for continuing to build.