# The intrikata-stack metamethodology — and what it can actually prove **Two parts.** Part 1 is the reusable pattern: ten rules you can apply to any ranking, score, or audit you publish, independent of this stack. Part 2 is the worked case study — *Are We In The Singularity* — where the stack was pointed at the hardest available question and the honest result was that **it proves its own method and explicitly refuses to prove the outcome.** That refusal is the finding, not a limitation of the write-up. --- ## Part 1 — The reusable pattern Every rule below was extracted from a defect that actually shipped. None is theoretical. ### 1. A constant that lives nowhere cannot be versioned If a threshold, weight, or band edge is a bare literal inside a scorer, the ranking can be re-tuned with **every published byte of provenance unchanged**. Freeze constants in a spec with its own version line. The version must be independent of neighbouring systems: a ranker that inherits its detector's version number is unversioned in practice. ### 2. A ranking with no falsifier is an opinion Publish, per item, the concrete observation that would **lower** it. Prefer *not-load-bearing* falsifiers (the thing advances without the item being solved) over *solved* falsifiers — nothing is ever fully solved, so a list carrying only the second kind is unfalsifiable in practice. ### 3. Declared instruments must actually be read A published falsifier the scorer cannot detect is decorative. In this codebase, **six of twelve** entries declared input classes their scoring function never consulted — the model could not have registered the event that was supposed to demote the item. Found by a behavioural probe (perturb one declared input, require the score to move), not by inspection, which had missed three of them. The probe is now a test. ### 4. Separate what matters from what is loud Attention — markets, paper counts, release velocity — is not necessity. Give it **exactly zero weight** in the substantive score and publish it beside as a named context tape. If a number cannot change the ordering, say so in the field name. ### 5. Assert that your classifications are reachable If each item's score has its own ceiling, some (item, band) pairs may be **unreachable at any evidence value** — a permanent classification masquerading as a measurement. Compute the attainable interval per item and count the dead cells. Report them; never repair them by widening a clamp, which raises a ceiling by adding nothing. ### 6. Do not let "we cannot see it" read as "it matters less" If an observability proxy contributes to a substantive score, items are penalized for being hard to observe. Here it was decisive: the item with the **highest dependency centrality in the catalog** fell below a band edge purely on its observability penalty — the penalty (4.2) exceeded its distance to the edge (2.5). Publish the score with the observability term removed, at zero authority, so a reader can see both. Never raise a confidence value to cross a threshold; confidence is an output of instrumentation, not an input to ranking. ### 7. Never fabricate readiness *Not yet measured* and *cannot be measured from here* are different states and must be different values. A single "unmeasured" literal collapses them and makes the whole column uninformative. When an instrument lands, record it as a **dated snapshot with an as-of date and a unit**, never a live fetch — a live scrape makes a published claim depend on a third party's uptime and page shape. ### 8. Read ground truth, never a rendering of it A summarizing reader reported a source's horizon values as *days*; they were *minutes*. Publishing that would have shipped a 1044-day figure. The raw asset settled it in one look. Any landed measurement must carry a unit, a finite value, and the source's own stated validity bound — enforced by a test, because the failure is silent. ### 9. Make invariants topological, not procedural "Never let an undecidable residual participate in a closure edge" is a policy someone has to remember. Give residual nodes **outdegree zero** and it becomes a fact: a sink cannot lie on a cycle. Prefer structure that makes the violation impossible over documentation that makes it discouraged. ### 10. Verify by running, not by string-matching The sharpest failure in this record: a fix was confirmed by grepping the built artifact for the string it should contain. The string was there. The code crashed on the first line that used it, because a substring guard had matched an unrelated occurrence elsewhere in the file and skipped an import. One command-line invocation caught it; the grep never would have. **Build it, run it, drive it to where the changed code executes.** --- ## Part 2 — AWITS as case study: the self-reflection The stack audits itself. `metamegapraxis` audits `megapraxis`; a seed swarm audits both; the audit's own verdict is written back into the graph the next audit reads. Pointing that recursion at *"are we in the singularity"* produces a clean split between two questions that are easy to conflate. ### What is provable — and was mechanically checked These are facts about the **system**, verified against live artifacts: | Claim | How checked | |---|---| | The candidate-necessity map contains no closure cycle | graph analytics, `has_cycles: false` | | No undecidable residual can lie on a closure edge | every residual node has outdegree 0 | | The necessity matrix is complete and closed | exactly 36 edges = 9 candidates x 4 pathways, labels from a 4-value vocabulary | | Control and observability gates are not weak candidates | all score `null`, never `0` | | The published package is reproducible | every file's SHA-256 matches its manifest | | The bootstrap installs from a cold start | driven end-to-end in a sandboxed home | Each of these is a proof **about the method**. None is a claim about takeoff. ### What is not provable — and is named, not buried - **Counterfactual necessity.** No listed item can be shown *necessary*, because unobserved substitute pathways remain possible. The ranking is conditional on four explicitly stated pathways; it is not a claim about all possible ones. - **The definition itself.** Whether deployment friction and governance belong in the catalog depends entirely on which definition of "singularity" is in use. That dependence is the residual, and it is why both are excluded with the residual named rather than silently omitted. - **Lab-internal state.** The loop-gain ledger that would measure recursive R&D requires AI-attributable research throughput. Labs do not attribute. No public instrument closes this, so the item is marked *structurally unmeasurable from public proxies* — named, not faked. - **Unknown unknowns.** A fifth pathway nobody has stated cannot appear in the matrix by construction. ### The recursion, stated honestly A system that builds itself, audits itself, and writes its own audit trail into its own graph is genuinely self-referential — and the self-reference is where the provable content lives. It can demonstrate that its own cycle closes. It cannot bootstrap that into a claim about the world. **The one thing this system is structurally incapable of proving is the thing its name asks about.** Every mechanism above exists to keep that boundary visible: the zero-weight attention tape, the null-not-zero gates, the outdegree-zero residual sinks, the refusal to fabricate a readiness number, and v3.2 live authority that publishes 7/30/90 public-proxy envelopes without promoting them into a singularity probability. Operational occurrence is claimed as not-occurred under hysteresis 65/70. Definitional occurrence stays `T-DEFINITIONAL`. 37/37 rdgap probes are public-proxy implementations, not frontier completion. Distance to the Tarski limit is UNDECIDABLE. Receipt: `/awits-withheld.md`, `/demo/awits-v32-live.json`, and `/demo/awits-rdgap-capability-impl.json`. A system that could be argued into proving its own conclusion would have no use as evidence. ### The failure record is the evidence The method is credible here only because it caught its own author. In a single session it surfaced: a production deploy sent to the wrong branch because a machine-wide memory rule contradicted a correct per-project note written eleven days earlier; a "verified" fix that had only ever been string-matched and crashed on first run; six of twelve falsifiers the scorer could not register; a package built from a stale mirror so the audited artifact was not the shipped one; and a stray backup file shipping to consumers with content the cycle existed to remove. Every one was found by running something, and every one is recorded. A methodology whose case study contains no failures has not been used. --- ## Applying this yourself The pattern is the product; the case study is only evidence it survives contact. Install the skills, then point the audit at whatever you publish that carries a score, a ranking, or a verdict: ``` codex plugin marketplace add https://docs.intrikata.com/git/intrikata-stack.git codex plugin add praxis@intrikata-stack ``` Other harnesses, package download, and offline paths: [/install.md](/install.md). Start with rules 1, 2, and 10 — versioned constants, published falsifiers, and verification by running. They catch the most, cost the least, and every later rule assumes them. Full technical specification: [/whitepaper](/whitepaper) · Reference install paths: [/install.md](/install.md) · Live audit snapshots: [/demo/cycle.json](/demo/cycle.json)