7. Discussion and limits
7.1 The criticisms of ANT, and what the framework does
Not every criticism of ANT can be answered by a method. The table states what the framework does in response to each criticism in Section 2.5, and what stays open. A guard is not the same as an answer.
| Criticism | What the framework does | What stays open |
|---|---|---|
| Symmetry removes the ground for explanation | Uses symmetry to find actors, not to explain outcomes or share blame. People stay accountable for systems they ship | The philosophical dispute; the framework does not need to settle it |
| Network accounts miss those left out | The actor register includes those affected but not enrolled, and the last human hop | They are recorded, but get no say in the design: there is no participatory step |
| Indifference to good and harm | Authority is part of correctness; every finding needs an owner and a disposition | Whether a decision ought to be automated at all; that judgement sits with the decision owner |
| Not a critical theory | Does not claim to be one; it is an engineering procedure | Users who need critical analysis need another approach as well |
| Structure and inequality ignored | Records authority as written in roles and limits | It takes that allocation as given, and does not ask whether it is fair or whom it disadvantages |
| Acceptance blurred with correctness | The oracle computes the right answer from the rules, independently of the network. Acceptance never counts as a pass | If the rules are wrong, so is the oracle; such findings go to their owners |
| Little procedure; boundaries hard to fix | A fixed lifecycle, gates and templates; the boundary is set by the decision and its path | Two analysts may still draw different networks (Section 6.2) |
| Easy to apply after the event | Every finding must end in a control, a referral or an acceptance | The claims rest on cases, not on measured comparison |
7.2 What the cases suggest
Three observations recur across the cases. First, the most consequential hop is often the one that looks harmless: accounts payable in Harbourline, the stock view in Northgate, the quantity field in Citigroup. Treating every hop as a mediator until shown otherwise placed each of them on the map, though in each case the source account had already described them. Second, many findings are not tool-shaped. In Citigroup none was; in Harbourline, the missing decision owner and the unagreed handling of wording changes could only be referred. The findings register and the analysis-only exit exist for these. Third, a faithful value can still be ignored. The US$196bn confirmation reached the trader intact and was confirmed. The framework checks what arrives at the decision, not what the decider does with it; such findings are recorded as "response" and referred.
The effect of a loss is also easy to misstate. Harbourline's L6 and L7 each change a component by £3,000 and £2,500, but change the payment by £0 on the test fixture, because another rule binds. A test that checks only the final number cannot detect them. The framework therefore records both the effect on the decision and the component effect, and requires an assertion for each loss it closes.
7.3 Limits
- Shown, not proven. What the framework does is shown in two cases. It has not been compared with other methods, and no claim is made that it outperforms them.
- No recorded evidence. The agent runs in the cases are illustrative. The test and upkeep half has been exercised on paper only.
- Hindsight. Citigroup was analysed from the regulator's account after the outcome was known.
- Agreement between analysts has not been measured. One independent re-run agreed closely on Harbourline and diverged on Citigroup.
- No participatory step. Those affected are recorded, not consulted.
- The fairness of authority is out of scope. The framework takes authority as written.
- Effort. Contracts, oracles and fixtures take time to build and maintain. The tiers exist so that effort follows consequence.
- NFE is a proposal. Its terms, dimensions and loss modes are a working framework, not a standard.
- A contribution, not a first. The search behind Section 3 was targeted, not systematic.
7.4 Further work
The next tests follow from the limits. Run the Harbourline acceptance test against a built server with recorded runs, 20 or more for each variant, a version of the inputs or of the system under test, so that G3 can be decided on evidence. Analyse a case before its outcome is known. Measure agreement between human analysts on the same decision. Compare the framework's findings with those of a team using another method on the same system. Settle the eight problems the independent re-run left undecided, among them a default tolerance for each tier. And complete a systematic review of the five fields named in Section 3.