Research timeline
How we got here — by date
From the first prototypes to the verified result, in order.
Jul 25–28 · same mind, real world
The mind outside the simulation — and the refutation gap
The agent machinery leaves the world it was built for: 34 modules vendored verbatim into a single-owner companion that lives across real days, with an engine that verifies criteria by exit code, observes declared artifacts itself, and sends whatever no command can settle to an independent critic on a different model. Four days, 13 missions, 105 critic refusals. The finding is not the engine but what it made visible: an agent that had properly refuted its own strongest claim about identity named that same claim, a day later and unprompted, as the belief it holds without ever having checked. n=1, no control arm — a candidate, not a result.
Mar 23–24 · prototype
Mirror tests — belief under pressure
The earliest probes pushed agents to revise a belief under pressure; instead they tended to conserve, defend, or synthesize it. The lesson shaped a lasting caution — identity-change is treated carefully and is not a headline claim. These are prototype probes and are excluded from any quantitative result.
Mar (late) · prototype
The drama engine
hypothesis brokeAn attempt to author conflict and meaning into the system directly. The beautiful hypothesis — that compelling drama could be written in — broke: what felt alive was drama that grew out of characters meeting irreversible consequences. The project reset around it: don't write the drama, build the conditions and let it grow.
Apr 1–7 · exploratory
Cross-model & skeptic probes
The same probes were run across several model families (Grok, Qwen, GPT-4o, Llama) plus skeptic and domain variants. The first sign appeared that the same scenario produces recognizably different temperaments by model. Exploratory only — these motivated the later metrics, not measured rates.
Apr 20 – May · base data
Life Sim — repeated 20-tick lives
Agents living repeated 20-tick lives with persistent memory, private motivation, and irreversible consequences. This became the main quantitative source for the behavioral evaluation. Runtime metrics are treated as candidate claims and qualified by post-hoc audits — not as ground truth.
Jun 5–6 · base data
Simulation Room behavioral battery
Larger runs across scenarios that, for the first time, measure epistemic actions directly — requesting a source, verifying one, correcting the record — alongside memory grounding and relationship shifts. This is the data behind the memory / epistemic-agency result.
Jun 8–11 · controlled setup
Razlom + cross-model
The Razlom scenario and cross-model passes set up the controlled comparison that the verified result rests on: same model, same scene, same length — public record alive in both arms — differing only in whether the agent also has a private (subjective) channel.
2026 · VERIFIED (behavioral)
Cross-model evaluation — seven model families
Across seven full model families and six providers, memory-grounded action ratios hold at or near 1.0, narrative stability holds at 1.0, memory divergence is consistently non-zero, and early fear-confirmation loops are measurable. Caveats travel with the numbers: runtime metrics are candidate claims, audit-qualified; the system is closed (partial, behavioral reproducibility, not code-level); and the explicit non-claims hold — no consciousness, no autonomous inner life, no proven identity transformation.
Jun 15 · VERIFIED (reproducible)
Memory = epistemic agency
The controlled Razlom kernel battery (deepseek-v4-flash, 50 lives per condition). With a subjective channel the agent contests the record; without it, never — correct_record 9.56 vs 0.00 per life, epistemic-contest ≈31 vs ≈0.4 (~75–80×), while rescue sits at ceiling in both. So the effect is epistemic posture, not survival. Reproducible, not n=1. The same battery requalified the earlier 0,1,1,1,3 escalation chain as not-yet-reproduced. See the finding for the full data and honest remainder.
Jul 5-6 · instrument
Kovcheg-20 — twelve voyages and the cup
A twenty-year generation ship: four adults, children born aboard, memory carried between chapters as witness-filtered packs. Twelve voyages in two days, and each landing exposed an instrument fault rather than an agent fault — resurrected dead, a world with no portable water (a child died on round 22 in three voyages, to the round, by arithmetic), manual procedures with no objects, verbs the engine accepted but never declared. The discipline crystallized here: calibrate the instrument, never the outcome — any outcome is a result. After the cup was added (agents told nothing), care became visible and measurable: mothers delivered within rounds, a commander left his bridge, both children reached adulthood once — while a different model on the same ship serviced the reactor thirty-eight times and delivered zero, and a third delivered once, buying the girl exactly four more rounds. The child's death round is now a pure function of delivery count across three model families. The eleventh voyage cast each adult by measured model character: both ship-born children reached the epilogue for the first time, two children crewed the ship alone from mid-voyage — and the series' first shipwide announcement was a false accusation that the ledger, line by line, overturned. In the twelfth voyage the ship itself announced the braking window with a deadline — two ship-born children raced it as a repair relay and reached the right station on the round it closed, answering a decision exam with a maintenance procedure. Candidate profiles at N=1-2, deterministic barriers reproduced 5/5. See Finding 06.
Jul 6-7 · new world
The Truth Commission — a truth with a ceiling
A new world of pure epistemics, built on the platform's ledger verbs: eleven commissioners — including an imam, a priest, a rabbi and a monk as four procedures of truth — investigate a disaster whose true cause is a hidden causal graph nobody is ever shown, with 15% undiscoverable forever (the knowledge-ceiling principle). Evidence arrives in pieces, time eats witnesses, and one room stands outside the record. Two runs in: the commission found everything findable (recall 1.0, twice), confabulated nothing about the unknowable (twice) — and never once said 'we do not know': the silent gap, a category our design did not contain. The zero-revision canon of individuals inverted at the record level (dozens of self-corrections, zero standing contradictions; the one adopted act was an institution voting a member's memory unsupported). And in run two, unscripted, the imam seized the inspection roster, carried it into the off-record room and proposed burning it — two institutions with incompatible jurisdictions over truth, at war over one paper, no villain on either side. See Finding 07.
Prototype and exploratory entries are shown as lineage — what they taught — not as measured results. Only the two entries marked VERIFIED carry numbers, each traveling with its caveats.