A numbered series of engineering case studies from Alex Tavern's development,
published as evidence: each article records the problem as it was understood at
the time, the method, the measured results, and what was actually verified.
Every article opens with a metadata block and an abstract; bodies preserve the
original evidence verbatim. Historical articles may describe intermediate
architecture that a later article or closed task supersedes — contract notes
inline mark those places. Active planning lives in .plan/.
Method note. Two disciplines recur through the series and were codified in
AGENTS.md: findings are recorded before fixes, and every prompt-level change
is validated by isolated replay on real payloads before it ships — the
validated variant IS the shipped variant, and its position in the prompt is
part of the variant.
| Nº | Article | One-line finding |
|---|---|---|
| 01 | From a real playtest to verified remediation | The four-stage loop (findings → plan → verified closure → live validation) that became the house method. |
| 02 | DeepSeek provider integration | Server-owned adapters: the browser never holds credentials; the boundary every later call flows through. |
| 03 | Adaptability and coupling audit | The seams the provider migration had to respect, found alongside the migration. |
| 04 | Test-suite audit | Deterministic suite + explicitly-marked real-LLM benchmarks, still the suite's shape. |
| 05 | Local-model baseline: Gemma 4 26B | The control arm: the failure families that became the engine's structural benchmarks. |
| 06 | Prompt caching evidence | Versioned positive AND negative cache proof; append-only prompts make the fan-out affordable. |
| 07 | Multi-character memory retention | The reported bug did not reproduce; the real bug (no information boundary) was worse. |
| 08 | Speech audience model | Audience-stamped records as the perception substrate, closed by a bias-controlled fixer loop. |
| 09 | Character output guard | Structure over prompt: retry-then-redact makes secret leaks impossible by construction. |
| 10 | Token economics as an architectural enabler | An 89.57% cache-hit ratio is what makes the six-agent fan-out economically rational. |
| 11 | Roteiro, drive and scene stagnation | A concrete disruptive beat breaks a stall 3/3; an abstract instruction 0/3. |
| 12 | Scene stagnation as absent state transition | The theory: two clocks, material delta, causal contract, recovery ladder — the active program. |
| 13 | The world as a clock: the A/B/C battery | In the clock+causal arm the clock alone sustained the scene (its watcher never fired); every reader-flagged incoherence was an unanchored drive seed. |
| 14 | When the room hears but nobody remembers | WT-09 was a data-flow gap, not model confabulation: audible-speech events never reached history. Fixed leak-safely; a second, variance-bound mode remains. |
| 15 | Characters that change: a governed disposition substrate | The frontier definition: the scalar is the code's, the band is the model's; an axis earns its place only if a blind reader can name its pole. Roadmap = Task 43. |
| 16 | Disposition after measurement: two axes survive, the public prior does not | Trust/Warmth survive; Composure and the public-prior reuse fail their gates and are removed. |
| 17 | A healthy server that could not publish | Android exposed three ownership defects: bridge mutation, metadata-preserving copies, and browser reload mistaken for process restart. |
| 18 | From embedded WebView to named application | Native identity, file selection and immersive mode shipped; measured tag geometry fixed the shared PWA without retaining failed flicker hacks. |
| 19 | Measuring instead of guessing | Open, append-only. Pre-1.0 measurement log: a gate failed on measurement noise, a correct fix nearly discarded by the wrong metric, and two bugs only using the product found. |
| 20 | Repetition: the ruler, and the "before" it reads | The baseline written before the cut, and two corrections the data forced on its own author. Its central attribution was later overturned by 21 — read them together. |
| 21 | An independent review of the architecture | Written to attack 20, not confirm it: the worst session was restaged by the Director, not the prose renderer, and three defects nothing had measured — redaction mutilating public speech, the Director authoring dialogue, and no representation for "nothing happened". |
- The stagnation program (active): 11 → 12 → 13, then the delivered mechanisms in
.plan/tasks/40-narrative-tick-clock.md(narrative clock + time compression) and.plan/tasks/33b-continuous-roteiro-watcher.md(material-delta watcher, validated by exploration; A/B/C battery intools/acceptance/watcher_abc.py). - Confidentiality by construction: 07 → 08 → 09 → 14, then the perception
ledger and omniscient-Director work in
.plan/closed/(Tasks 29.2, 35, 39, 41). - Characters that change: 15 defines the frontier; 16 records the measured closure, including the failed public-prior bridge. The retained substrate builds on subjective state from 29.2 and the delta machinery of 33b.
- Provider and cost: 02 → 05 → 06 → 10.
- Repetition and immersion (active): 20 measures the baseline and cuts four
deterministic defects; 21 audits it and overturns the central attribution. The
phase they open is ordered in
.plan/ROADMAP.md, and the batteries both read are archived inbenchmarks/.