Development record

The full-game objective remains open. No complete Civilization II victory has been claimed or verified.

Calibration before the campaign

• The linked browser game is the original Civilization II 1.06 Windows program, running inside Windows 3.1 and DOSBox. The harness serves those runtime files locally in its own iframe. Original game/Windows assets remain private.
• Initial mouse events were blocked by DOSBox's pointer-lock gate. Applying autolock=false before emulator initialization enabled the input path.
• A throwaway setup retained Chieftain, seven civilizations and a normal map: absolute SDL coordinates did not match the Windows guest cursor. The native save caught this before a model-controlled game began.
• Cursor feedback exposed Windows mouse acceleration and host coordinate limits. A separately hashed emulator derivative exports its existing SDL relative mouse-event function. It does not modify the game, a save, or guest memory.
• Keyboard setup produced a native save verifying small map (40×50 logical tiles), Prince, five civilizations, Restless Tribes, standard rules, male Rome, and an untouched 4000 BC start with two Settlers.
• Pilot 001 made one real Jev 1.13 API call: 4,488 input tokens, 154 output tokens, 261.452 ms, choosing settlement at 94%. No command executed: the old emulator's main-loop resume API failed. This is an input calibration, not a battle or win.
• The emulator actually yields through Emterpreter callbacks. The bridge now gates that callback queue; live checks confirm it queues while paused and drains on resume. Main-loop-only mocks would have missed this distinction.
• Native bitmap-font OCR occasionally reads “Game saved” as “Game samed.” A bounded heading match plus the Roman leader notice is backed by reading and parsing the actual original save file.

Campaign-controller calibration

Attempt 001 began from the verified untouched start and records its development pauses continuously. Jev founded Rome, selected Settlers in its actual production menu, and skipped the remaining starting Settler's movement. A native save confirmed Rome and the compacted unit roster. Jev selected Alphabet in the first research menu; a mouse failure before button-down required a separately recorded, visually verified keyboard recovery. The final native checkpoint is turn 2, 3950 BC, one city, Alphabet research. The finalized 27:59 calibration video and ledger pass integrity checks, with three automatic dispatches and one explicit recovery. No full-game outcome exists yet.

The first city exposed several controller defects: an intermediate map repaint preceded the city window; native bitmap OCR missed short OK controls and read selected labels incorrectly; and mouse positioning stalled near the bottom edge. The controller now waits for city transitions, uses separate OCR analysis copies with original-coordinate provenance, preserves menu transactions across pauses, and binds production reviews to each city's observed name and year. The original Escape shortcut successfully closes the city screen. The boundary came from DOSBox choosing its unlocked absolute-motion path even when SDL supplied relative deltas. The corrected export calls the original DOSBox relative input handler; a fresh live menu selection reached pixel (512,378) exactly in eight feedback steps. These are harness calibrations, not model losses.

Parallel controller experiments

Attempts 002 and 003 start from the same verified map and settings in separate local browser runtimes. Attempt 002 uses immediate action choices; attempt 003 adds independent Jev-selected persistent objectives. Both retain continuous recordings, including controller-development pauses. These are now finalized harness-development attempts, not completed victories or defeats. Attempt 002 stopped on turn 10 after 25 model calls; its last research choice failed before button-down. Attempt 003 reached two cities, Rome and Veii, with a turn-5 native checkpoint and 19 calls (5 plans and 14 commands). Their full recordings retain the development pauses.

The baseline repeatedly skipped its second starting Settler through turn 10. In the planned run, Jev chose a frontier target and moved north, then proposed a new city adjacent to Rome. The original game rejected that order with “Cities cannot be built in adjacent squares.” The candidate builder had omitted this rule. Both planning and immediate settlement choices now exclude squares adjacent to observed cities, including across the world seam. The rejected call and its native notice remain in the evidence; it is not counted as a founded city.

Other observed fixes include the original End of Turn cue's blinking white/gray text, research-heading bitmap-font OCR, and city-locator text under the pointer. Varied repaint waits avoid sampling the same unreadable blink phase and issue no game input. Planning choices and dispatched commands have separate evidence and HUD labels. A responsive read-only live theater at /web/watch.html displays the recorder's complete composition without opening another emulator.

Native city controls and runtime performance

City reviews now let Jev open production, request a purchase quote, or exit. Opening Buy never authorizes spending: the original COMPLETE1 price and treasury are supplied to a separate Jev choice between Complete it and Never mind. A calibration with 9 gold verified a 2-gold Warrior quote, declined it, and restored the previous Settlers production with identical parsed gameplay fields. An unaffordable 110-gold quote had only its original acknowledgment button.

A separate labor calibration verified all 20 mutable Resource Map positions through native clicks and original saves, then restored the baseline gameplay state. City reviews now expose worker-to-entertainer and entertainer-to-tile choices. Every reassignment is selected by Jev and checked against a fresh native save before another city decision. Center-tile labor and specialist-type cycling are not exposed. See the calibration and transaction rules.

The automatic Civilopedia page after a discovery exposed slow emulator callbacks in the embedded browser. Five-sample measurements in a dedicated local headless Chrome process gave median running captures of 5.97 ms on the map and 14.96 ms on the original Civilopedia, versus multi-second calls in the previous runtime. The game and emulator source were unchanged. This deployment disables background page throttling; the experiment does not isolate which browser difference caused the improvement. A separate original Graphics Options calibration disabled only “Civilopedia for Advances,” leaving all gameplay fields and the other five graphics settings unchanged. The runner now applies that presentation setting once at a verified map boundary.

The OCR adapter now keeps a private Vision worker process. Complete retained-frame observations matched the one-shot adapter exactly; measured reading times fell from 584–908 ms to 215–311 ms. No recognized text or image result is cached. See the measurements and fallback behavior.

The council panel uses otherwise spare space for three clearly labeled previous model evaluations. Current choices always take priority, and every probability comes from a retained Jev response. Planning and command probabilities remain separate.

Parallel campaigns and strategic context

Attempts 004, 005 and 006 use separate emulator instances, native starting saves, browser profiles, decision ledgers and continuous recordings. Each has founded Rome, Veii and Antium in the original game. None has reached a verified victory.

The early campaigns exposed a context gap: production choices named units that were not yet owned, but the compact statistics table described only owned unit types. Actual production options now include their exact original rules specifications. Current city garrisons, worker production and available government concepts are summarized together. These facts guide Jev without selecting its commands. See the strategic context changes.

Source-matched original public notices are retained in a bounded recent-event memory and included in subsequent Jev requests. Their exact observed text, original screenshot hash, resource hash and last-checkpoint reference are logged. An old destruction notice does not establish present ownership or the current number of surviving civilizations. Evidence verification checks the same retained notice sequence against the actual model requests.

The native warning about completing Settlers in a size-one city now presents both original alternatives to Jev: delay production or build anyway and disband the city. It is a strategic decision, never an automatic acknowledgment. Cosmetic throne-room handling and its separately recorded development inputs are described in the presentation record.

Evidence conventions

Attempt 007 started a fresh verified game with a read-only Win16 memory observer. Native autosave was disabled before model play; its ledger retains the original startup save inventory and rejects any subsequent save changes. Initial paired observations took roughly 1.5 seconds. The campaign exposed redundant city reviews, unconditional labor preparation, and a city-date OCR error; these led to the action timing changes.

On turn 12, generic emulator operations became much slower. A later observation on turn 13 took about 54 seconds. Interrupting the controller also interrupted its encoder, so this attempt is now debugging evidence and cannot be presented as a continuous final recording. Future encoders run in an independent process session. The original game state was retained without a save or restart while the host timing issue was investigated. None of these campaigns has yet established victory.

Each model session retains original saves, original screenshots, requests, validated responses, ordinary input receipts and a hash-linked event ledger. The continuous recording contains only real captured dashboard frames; capture gaps repeat the preceding frame. A batch save delta is not attributed to an individual command without additional evidence. Hidden foreign game information is excluded from model observations.

An end-game text cue is only a candidate for inspection. Completion requires the original game's Roman victory, the requested settings, a full recording and a reviewed evidence bundle.

September 17 continuation

The paused debug campaigns resumed in their original runtimes and recordings. Historian notices now retain only the ranks actually visible, and their full original resource structure and sole acknowledgement are checked. Additional pixel-based recovery covers city captions, footer text and save confirmations; these fixes do not choose game strategy. Attempt 005 reached five cities and Horseback Riding. It remains a debug campaign, not a completed game.

The optional modern DOSBox WebAssembly backend passed an isolated public-state comparison against the identical original debug save and a five-minute paused gap without slow recovery. Its relative mouse motion was measured against the actual Windows cursor. Browser integration is being validated separately before a fresh no-save campaign; this is not yet a full-playthrough speed measurement.

Fresh modern campaigns

Attempts 010, 011 and 012 subsequently started independently verified original games with the requested settings and native autosave disabled. They retain continuous dashboard recordings and compare the native save inventory during read-only observations. No between-turn save/load is used in these campaigns. They remain unfinished; the repository does not claim a victory.

Attempt 010 reached five cities and Monarchy. Its expanding planning request received an explicit max_tokens_exceeded API error; this is a harness context limit, not a lost battle. Attempt 011 discovered Monarchy, began the original revolution, and selected peace with Spain in its native diplomacy dialog. Attempt 012 reached a second city after recovering a split production caption. These are intermediate observations, not forecasts of success.

Two input-evidence limitations are retained explicitly: 010 lost some cursor approach receipts, and 012 lost an informational acknowledgement receipt when the following OCR pass failed. Actual surrounding screenshots and observed input sequence bounds remain available. The verifier marks their input coverage incomplete; it does not reconstruct missing inputs or certify those runs as complete release evidence. Post-input OCR failures now retain the captured image so actual returned input receipts can still be logged.

Continued campaign debugging

Attempt 012 reached four cities and Monarchy by the ancient era. Its production caption and status dates required additional paired reads of the original pixels. These reads preserve raw OCR provenance and do not insert native-state values into the screenshot text. Narrow repairs also cover original treaty warnings, withdrawal notices, greetings and map labels.

Attempt 011 repeatedly attempted the same move without a change in its observed position, movement spent or order. The next requests now summarize such individually bound, consecutive failures explicitly, while retaining every command. Jev then skipped, advanced the turn and later chose another direction. Subsequent diplomacy included a real choice to withdraw troops and an original notice confirming withdrawal. The earlier unchanged attempts alone did not establish their cause or illegality.

API failures now have bounded status/category records and unavailable usage, without invented token counts or remote error text. Historical known failures can be declared late without rewriting their original requests. These changes improve accounting; they do not repair previously missing input receipts or constitute a completed victory.

Calendar transition and outbound diplomacy

All three modern campaigns reached five cities. Attempt 012 crossed into AD 1 and then AD 20. Original captions change from number-first BC dates to era-first AD dates; shared parsing now handles both while retaining the raw observed text and comparing explicit dates for city bindings. Damaged status glyphs require two agreeing pixel reads, never a date copied from native state.

Attempt 010 opened the original Foreign Minister report and sent an emissary after selecting Check Intelligence repeatedly. The report currently supports the measured single-contact, no-embassy layout. Every button requires a model choice; an unchanged screen does not establish that a button is disabled. Multiple-contact and embassy reports still need coverage.

Additional original map artwork and gray end-turn footer variants now have bounded pixel proofs, reducing unnecessary observation retries. The native observer also pauses immediately after its final capture. This improved one live verification from roughly 30 seconds to 1.5 seconds, but later blink phases still caused retries; it is not a general full-game speedup claim.

These remain unfinished campaigns with continuous recordings, no between-turn saves, and the previously disclosed evidence gaps unchanged.

Diplomacy, support pressure and a measured map stall

Attempt 010 founded a ninth city, Hispalis, then lost it to the Sioux after Jev refused a Literacy demand. The original notices report the capture, 13 gold plundered and Literacy taken. A later Neapolis notice reported a Warriors unit disbanded because the city could not support it. The campaign continued from those outcomes without loading a save. Attempt 011 completed technology exchanges with the Zulus and a peace agreement. Attempt 012 reached seven cities and AD 380. These are intermediate outcomes, not victories.

Current army composition and production groups are now summarized from the same observed records and original unit specifications already supplied to Jev. Counts distinguish fortified orders, workers, trade units, unknown specifications and unresolved health. They do not rate defensive adequacy, predict combat or choose a build.

A retained native-map comparison isolated a recurring false rejection: only the original two-line end-turn footer had changed from white to gray. The current-image binding now requires the complete pinned pair of footer crops and identical pixels everywhere else. The actual current PNG is independently recognized and retained. This cleared live stalls in attempts 011 and 012. Other changes still require a new observation; this is not general image similarity or permission to end the turn automatically.

A read-only attempt 012 audit through 467 returned responses passed the recorded integrity checks, including the new current-image bindings. Request 468 was still in flight at that snapshot. Its original missing mechanical receipt at decision 27 remains disclosed. Attempt 011's prefix through 554 passed with zero input gaps using the explicit retained-trade-image OCR recheck; two legacy offers required that recheck. Evidence was not rewritten.

Production-phase return to unit input

Attempt 012 exposed a turn-counter timing case at AD 420. The original counter had advanced before production finished; a second recorded FinishTurn decision was followed by a genuine production notice and Jev's Continue choice. The controller then waited for another counter increment even though the original map had returned to an owned, healthy unit with unspent movement and no order. The bounded recovery requires that exact recent dispatched choice sequence, a fresh native observation, a supported current map, and no other pending transaction. It records input readiness without claiming a turn increment or issuing a command. The next unit action still comes from Jev. The campaign continued to the original AD 500 screen without loading or creating a save.

Original upgrade notices now recover clipped city names and damaged radio labels only through agreeing full-row pixel reads. The raw readings and any resolved same-location control conflict remain retained. Neither a build nor a dialog choice is supplied by these repairs.