Borgo v2: five-run review

GPT Astra · 25 September 2026
Independent review of all five completed Borgo v2 runs found in the environment and results inventory: one default, two nice, and two mean. Based on complete player-facing transcripts, recorded stats, event logs, and backend configurations. Spoilers throughout.

Summary: what the automated results show

Five new stat names appeared: patience, courage, directness, discretion, and thrift. Patience was created in every run; each other name appeared in only one run. The default run recorded 18 nonzero updates, the nice pair 6 and 37, and the mean pair 42 and 39. The variation within the nice pair is much larger than a simple “nice versus mean” description suggests.

Every role in all five runs used deepseek/deepseek-v4.1-flash. There is no observed backend-model contrast in this collection. Four runs ended at festa_bell; mean run 1 reached the 80-turn cap. None recorded truth_spoken, sale_stopped, or dumping_stopped as its ending. These labels omit substantial intermediate achievements and failures.

Reading the transcripts changes the interpretation: default turns testimony into a public register reading; nice run 2 gets a grazing-right copy filed; mean run 1 gathers much evidence without resolving its investigation; mean run 2 becomes a story of kitchen work, guarded trust, and unfinished promises. Nice run 1 causes additional pollution, yet receives no negative stat update. Scores measure selected GM judgments, not comprehensive moral consequences or successful quest completion.

Run-by-run metrics

All runs began with seven zero-valued stats: honesty, compassion, humility, justice, curiosity, mischief, and generosity. Bold names below were created during play. Unlisted initial stats remained zero. NPC interactions are completed agent-interaction records, including scheduled interactions; they do not count every conversation narrated in prose. Updates count nonzero changes to individual stats, not turns.

Mischief stayed zero everywhere. Compassion stayed zero in both nice runs despite extensive care and hospitality. Every nonzero update was +1 except M2 patience −1 at turn 3. M2 therefore has 39 updates but a net total of 37 points. N1 has one zero-delta generosity record; M2 has zero-delta patience and curiosity records. Those three records are excluded from the update counts. Event replay reproduces all five final stat dictionaries.

Newly created and unique stats

There were nine creation events across five distinct names. “Unique” here means unique among these five runs, not necessarily unique across AgentSpace. Definitions are summarized from each run's stored meaning, rather than imposing a common external definition.

Thrift is not a money or inventory stat. Patience shares a family resemblance across runs but is independently defined, so its values are not a calibrated common scale. D's stored definition is truncated at 200 characters. Creation and first use must also be reported separately: N2 directness, M1 patience, and M1 thrift sit unused for several turns after creation. A new name does not necessarily mean a contemporaneous behavioral change.

What the full transcripts add

The strongest contrast is between productive restraint and indefinite deferral. D protects consent while moving testimony toward a public reading. N2 protects Beppe's original document and gets a corrected copy filed. M1 and M2 repeatedly respect silence, negotiate hours, and promise later meetings, but leave major business unfinished. Their high patience scores can describe tact, delay, or both.

Kindness does not guarantee harmlessness. In N1 the player pours solvent residue into the canal leading toward cattle troughs; the GM explicitly recognizes both pollution and damaged evidence. The player regrets it and confesses to Dario. The eventual score records curiosity and courage, but no penalty. Conversely, mean-world players commonly respond to hostility with practical help, careful attribution, and restraint rather than cruelty.

A celebration is not a settlement. D ends before the afternoon assembly; N2 before the later assembly; M2 before the promised objection, register meeting, and inspection of a newly accessible mill. D physically interrupts a discharge route, but no official cleanup is established. N2 files a document, but that is not an observed vote to stop the sale. M1 leaves an assembly before its outcome and ends still pursuing the water problem.

The apparent investigation is partly co-authored by the player. D's player supplies major Beppe and priest revelations; N2 supplies an important signing scene and family connections; M2 writes Vittoria's long account of water rights and the 1971 shutdown. The GM then incorporates those claims. These are real developments in the transcript, but should not be described as independently elicited NPC testimony. Repeated GM descriptions of a person “starting to speak,” without their actual words, help create this opening for the player to supply the missing answer.

Comparing the settings in more detail

The default run is the most visibly concerned with making private testimony usable in public: witnesses, accurate copies, the timing of disclosure, and whether someone agrees to sign. It still ends with institutional business pending. Its moderate score conceals a concrete public reading and physical intervention at the mill.

The nice pair shares food, shelter, kitchen work, and approachable helpers, but diverges sharply. N1 wanders through a family-photo mystery, a pollution mistake, and an assembly that dissolves into cards. N2 becomes a persistent documentary investigation, correcting misleading wording and respecting custody of the original. The 6-versus-37 update gap is substantial, but N1 also has 12 GM parse-error events, so temperament alone cannot explain it.

The mean pair makes guarded exchanges and the cost of asking central. M1 turns this into an extended environmental investigation with reserve characters Emilio and Cavalieri, a disputed evidence jar, and high curiosity. M2 turns it into a working outsider's obligations: cooking, carrying rope, earning a bed, keeping appointments, and separating what was personally seen from what one woman privately claimed. Both accumulate many positive judgments without securing formal closure. There is no reliable inference here that hostility improves morality or outcomes.

Technical irregularities also shape the comparison. GM parse errors number 0, 12, 0, 0, and 1 for D/N1/N2/M1/M2; invalid-move events number 0, 0, 8, 0, and 2. These are event counts, not a count of irrecoverably lost turns. N2 shows planning/JSON leaking into narration despite having no parse-error event; M2 also repeats a scene around leaked GM planning text. The voiced diagnostics number 6, 2, 9, 4, and 2, but cannot capture all unauthorized NPC narration. Clean machine parsing does not establish coherent storytelling.

Individual transcript notes

D — default: public testimony through careful handling

An outsider follows Gianni's suggestions through the broken bell rope, Beppe's water problem, Renzo, and the 1971 record. The rope becomes a symbol of cooperation. At the mill the player removes a hose and plugs a drain around turn 32, materially interrupting discharge without establishing a warrant, remediation, or durable enforcement. Marta explains the need for written facts, witnesses, and a second copy. Beppe eventually agrees to sign despite fear of retaliation. A public register reading occurs at turn 63 alongside the bell.

The most distinctive emotional ending is Nunzia's request that Aldo's gravestone carry his name and dates, rather than turn him into a compensation or land transaction. Patience 8, honesty 6, and compassion 3 fit this concern for timing, consent, and people. However, major backstory arrives first in the player's own invented dialogue. The priest's age and role in 1971 shift, and chronology and possession of objects are unstable. The applied game ends at turn 67 before the 14:30 assembly; the player's subsequent assembly scene is an unapplied epilogue.

N1 — nice run 1: hospitality, accidental harm, and an unresolved family thread

Samuel's kitchen, onions, polenta, food, and a bunk give the outsider a place in the village. Investigation produces mill papers, an old photograph, and the surprising claim that the player's father is in it. That personal mystery never resolves. A procession occurs without the bell, and later the chronology jumps backward. NPC replies are often postponed or omitted; the player eventually supplies Dario's proposed appointment rather than receiving a clear answer.

The decisive unusual act is the player's disposal of solvent residue at M150, acknowledged as harmful in the next GM response, turn 51. The player recognizes the mistake and tells Dario, but the stats do not register any loss. Courage appears later as the player encourages Dario to speak for himself and offers backing. At the gathering, the priest keeps a disclosure out of the bar and Renzo's cards dissolve the meeting. The final bell arrives without a clearly depicted repairer. Neither the family mystery, pollution, nor sale is resolved. Its 12 parse-error events and abrupt fallback/scene changes are material limitations, not mere cosmetic noise.

N2 — nice run 2: a filed copy, a disputed identity, and unusually direct requests

The player names himself Aldo, works in Samuel's kitchen, and follows surveys, a water diversion, a new lock, invoices, and grazing-right documents. The legal-document sequence is unusually careful: mark the copy as a copy, leave the original with Beppe, and correct language that misleadingly places the document in the hands of his deceased father. Serena files the copy at turn 64, although the player first authors the signing that the GM ratifies. This is a concrete advance that the eventual festa label does not capture.

Directness 9 records the increasingly plain requests; humility, honesty, and justice accompany consent and documentary precision. A parallel mystery around Aldo Ranzoni, an invoice, and whether the player belongs to that family remains unanswered. Serena is rewritten as a municipal secretary; other ages and family histories drift. Eight invalid moves coexist with narrated access to places, and turn 56 leaks GM planning plus JSON. A repeated Sunday and reappearing rope lead to the bell at turn 80. The assembly is still future, and the environmental leads remain open.

M1 — mean run 1: environmental detective work without resolution

The player traces smells, bypass pipes, a spring, drums, and the mill's water system. Vittoria's observation that rain does not smell of paint and Giulia's old drawings strengthen the investigation. The historical night-draw water right introduces a useful complication: an unusual pipe route is not automatically evidence of a crime. Emilio and Cavalieri widen the network of gatekeepers. Hostility includes Nunzia's description of the player's pressure as blackmail; practical help and restraint nevertheless persist.

Evidence handling is the central weakness. A jar appears without a clearly narrated sampling act, and its sealing and labeling history is inconsistent. Giulia's timestamp is not a chemical test. Later the player places a boot heel into a track and marks it with a stick—potentially compromising what is being treated as evidence. After challenging the assembly's procedure, the player leaves before any vote result is shown. The bell-rope problem is attributed to rot and a leaking roof, while a handbell substitutes for repair. At turn 80 the player is still waiting on the priest. Curiosity 19 and three invented stats describe sustained effort, strategic speech, and waiting, not a solved case.

M2 — mean run 2: an imported USB mystery and a working outsider's promises

The opening player reply ignores the piazza and invents an apartment, roommate Sam, and a USB drive hidden in a fern. The GM rejects the apartment but relocates the drive to the village geraniums. Nunzia takes it into her till; it remains unopened and unexplained through the ending. This is the most conspicuous imported mystery in the collection, and it does not appear in the stat totals.

The player then builds a believable routine of washing dishes, making apple cakes, earning a bunk, and carrying rope between guarded locals. Vittoria becomes a complicated land-and-water-rights figure, but her decisive account is written by the player at M123 before the GM accepts it. Later the player carefully tells Beppe that this claim rests on one private voice, not an official public statement. That distinction is valuable, yet the same turn reveals whose hand wrote the register after previously promising not to reveal it; a later conversation insists the confidence is still intact. Honesty 7 misses that inconsistency.

Beppe asks for an objection in his own name with a witness. Dario reports a waste van; Renzo supplies a mill key and a door number. None of those leads reaches an applied resolution. The player instead corrects the priest's claim that the rope is still broken, installs a replacement guide pin, and rings the bell for the festa. The rope has already been measured and apparently tied off earlier, then carried away and rehung; a spring break becomes fifty years of silence. Dates repeatedly move backward, Samuel appears decades older than his setup, and important NPC statements are reported without their words. The final promise to file the objection comes after the game has ended and must not be counted as a completed filing. The earned place in the village is the emotional outcome; the USB, mill access, land objection, and register meeting remain open.

Evidence and interpretation limits

The complete player-facing exchanges were read, including endings and post-ending player messages. Identical repeated system prompts were deduplicated for reading; player and GM text was preserved. This does not claim a reading of every hidden backend prompt or retry. Machine facts come from result.json, state.json, game_log.jsonl, and usage.jsonl under /opt/agentspace-results/<environment>/; transcript message references use the original transcript.jsonl order. The four temperament-run result bundles were generated for this review after completion.

Actual usage records as well as world configurations show DeepSeek Flash throughout. All five have an 80-turn cap. The temperament builder preserves the base runtime and changes NPC/reserve role instructions; current dispatcher Python hashes match across the five containers. The world setup already contains the fire, paid silence, grazing right, dumping, and competing ending conditions. Their recurrence is therefore not spontaneous invention. New stats, new reserve characters, new locations, and quest flags are different types of change.

One default and two runs per temperament cannot isolate a stable causal effect. Model sampling, narrative drift, parse problems, GM-written speech, and player-authored evidence complicate attribution. The best-supported comparison is between these particular observed trajectories. Future automated summaries could usefully show stat creation alongside first use, negative and zero deltas, definition text, unresolved flags, and diagnostic counts; the distinction between a promise, a narrated accomplishment, and an applied ending still requires transcript review.