EscapistsReintroduced/docs/ai-migration-progress.md

# AI migration progress — 20 September 2026

## Memory-based native search area — 3 October 2026

Ordinary ground SEARCH tasks now snapshot the delivered belief's uncertainty at assignment or retasking, clamped to 45–110 m. On radio delivery HQ rejects a non-finite or out-of-range snapshot before waypoint allocation and before replacing the current assignment. The waypoint receives its radius once before group installation. A later change to the live belief cannot silently resize an already queued order. No extra force, setting, deadline extension or player-position query is introduced.

The native SearchAndDestroy activity uses this value as the investigation grid's square half-width. The larger grid adds cell and navigation work; neither the cap nor the offline tests establishes an FPS improvement. Missing-patrol investigation, road/site-security defence, INTERCEPT/CONTAIN and the air-team system's separate direct task remain unchanged. The Workbench patrol fixture initializes the same snapshot for future runs.

Negative-search reporting still requires an attributed fit original member within 40 m of the centre for 30 continuous seconds, no native contact, and radio delivery. That witness marks one coarse 500 m cell, not a completely searched square. Larger physical searches may provide fewer such completion opportunities; this change does not weaken the witness.

All 305 registered regression programs pass; native compilation, physical search behavior, multiplayer and performance for this source change remain unverified. [Source, regression controls and SDK metadata](evidence/hq-search-radius-2026-10-03/README.md).

Current convoy implementation: [Supply convoys](supply-convoys.md) supersedes the single-convoy stages recorded below.


The successor now has a persistent HQ belief model, bounded infantry pursuit,
radio check-ins, missing-patrol investigations, local body discovery and a Core
force ledger. These extend the existing playable mission and finite mortar/courier
systems. They do **not** complete the living-island plan, all predecessor AI
features, or multiplayer/release qualification.

This checkpoint describes the source captured by the native tests at **01:47:53**
and **01:52:20**, plus the focused regression corrections completed afterward.
Later convoy or planner changes need their own evidence; the results below do not
qualify edits made after those snapshots. Existing base-loop and packaged-playtest
history remains in [current status](current-status.md).

## Implemented in this checkpoint

| Area | Actual behavior and boundary |
| --- | --- |
| Core force ledger | [Ledger](../addons/core/Scripts/Game/ERR/00_Core/ERR_ForceLedger.c) and [simulation](../addons/core/Scripts/Game/ERR/00_Core/ERR_ForceSimulation.c) retain up to 128 formation identities, initial strength, casualties, duty, home position and authored routes. Six existing site groups register before HQ consumes them; other ready physical enemy groups register through GameMode. Physical snapshots count unconscious members as living. Eliminated/retired identities cannot restore strength through re-registration. |
| Dormant arithmetic | A pure route simulator advances nonphysical records on a staggered two-second schedule, with 1.4 m/s walking and 11 m/s convoy rules, authored elevation interpolation and bounded whole-lap skipping. It performs no terrain, navigation, player or entity queries. This supplies the arithmetic foundation; production presence arbitration and member materialization are still missing. |
| Delivered belief | [Memory](../addons/core/Scripts/Game/ERR/10_AI/ERR_HQMemory.c) retains up to 64 delivered tracks independently of short-lived fire contacts. Sightings decay with a 600-second confidence half-life and widening uncertainty. Credible successive reports infer motion; extrapolation is capped at 30 seconds/300 metres. Body and distress reports begin weaker and do not infer a moving enemy. Target handles are identity keys, not live-position sources. |
| Pursuit and objectives | [Pursuit rules](../addons/core/Scripts/Game/ERR/10_AI/ERR_HQPursuit.c), [doctrine](../addons/core/Scripts/Game/ERR/10_AI/ERR_HQDoctrine.c) and HQ assign existing eligible infantry to intercept, search or contain objectives. Age/confidence restrict precision, travel distance affects task lifetime, and stale/new report revisions control replacement orders. Actual downlink delivery precedes native waypoint changes. Home guards, busy/support units, two approach sectors and finite task lifetimes constrain commitments. Doctrine uses authored objectives and reported headings; it does not claim road-route or player-goal knowledge. |
| Searched areas | [Coverage](../addons/core/Scripts/Game/ERR/10_AI/ERR_HQCoverage.c) records at most 128 visited 500 m cells for 900 seconds, using delivered completed sweeps. Replay/revision guards prevent repeated or outdated negative reports from reducing a newer belief. Fresh positive sightings reopen their cell. Reserved destinations prevent concurrent reuse. A coarse visited cell is not proof every building or hiding place was searched. |
| Routine status | [Sampler](../addons/core/Scripts/Game/ERR/10_AI/ERR_HQStatusSampler.c) retains each formation's initial roster, up to 32 members. Only valid retained corpses with explicit destroyed damage state count as local confirmed deaths; missing references, transferred members and unconsciousness do not. A detached status needs a surviving original owned radio caller. [Watch](../addons/core/Scripts/Game/ERR/10_AI/ERR_HQWatch.c) changes HQ's roster only after uplink delivery. |
| Missing patrols | Routine check-ins are scheduled every 180 seconds. After two missed check-ins, HQ attempts a request and allows 30 seconds for a reply, even if no route exists. One existing eligible patrol may investigate the missing unit's **last delivered** position. Source/status/request identities reject late completions; a fresh check-in resolves the case and cancellation travels through radio. Silence never proves a death or creates an exact enemy target. |
| Body discovery | [Body observer](../addons/core/Scripts/Game/ERR/10_AI/ERR_HQBodyObserver.c) uses a bounded own-force registry, a rotating window of at most 16 corpses, a 15-second per-group scan cadence and a shared trace budget. Physical observers need range/facing and native visibility checks. Discovery becomes low-confidence casualty evidence only through radio, with cause unknown; it cannot authorize precise fire. |
| Skill and rare anti-air equipment | [Equipment service](../addons/core/Scripts/Game/ERR/10_AI/ERR_EnemyEquipment.c) binds an Expert minimum only to eligible registered enemy actors and restores the captured skill on release. Finite startup Stinger selection defaults to one in 15; `/esc hq stinger show\|set N` supports disabling or changing future selections. Pending/late inventory callbacks do not cause replenishment after success, firing, dropping or looting. See the [equipment audit](ai-equipment-migration.md) for owned-source hashes and exact limitations. |

The original owned `SquadLimit` behavior **raises**, rather than shrinks, the
shared commitment ceiling: `baseline + floor(confirmed losses / 2)`, capped at
eight; a zero baseline stays disabled. This corrects the contradictory wording
in the existing [T10 brief](living-island/T10-ai-migration.md), which is preserved.
Current production additionally limits hunters across all beliefs to three and
allows only one separate missing-patrol investigation. Thus the eight-squad
ceiling is not a promise to deploy eight hunters or spawn replacement troops.

## Native evidence and its limits

| Captured run | Observed result | What it establishes |
| --- | --- | --- |
| [HQ memory 01:47:53](evidence/native-2026-09-20-014753-hq-memory/report.json) | 56 total assertions, one completed scripted round, zero script/runtime failures; four engine errors retained. | Actual six-site ledger registration and detached elimination; immutable memory snapshots and bounded prediction; a synthetic report blocked by removing the transmitting radio; later real radio delivery reaching production intake once; delivered negative-sweep effects, replay protection and rejection of an old sweep revision. Actual command Stop clears history. The ten-minute advance is a value-clock fixture, not ten real minutes of gameplay. |
| [HQ response 01:52:20](evidence/native-2026-09-20-015220-hq-response/report.json) | 35 total assertions, one completed scripted round, zero script/runtime failures; the same four engine errors retained. | An arranged sighting produces actual native perception, report, uplink, HQ receipt and downlink lineage. A distinct original infantry group physically moves toward its native Search waypoint with bounded unique orders, without adding or replacing a force. This is not the two-UAZ convoy described in T10. |

Both runners recorded unchanged successor source and six unchanged external addon
inputs, hidden Gamescope displays, silent/muted test audio, and confirmed owned
cleanup. Their [memory runner](evidence/native-2026-09-20-014753-hq-memory/runner.json.txt)
and [response runner](evidence/native-2026-09-20-015220-hq-response/runner.json.txt)
retain these facts and exact launch manifests.

**Neither run passes the strict native qualifier.** The four retained errors are
GUID/name mismatches in external string tables: `rhs_localization.st`,
`wcs_armament_localization.st`, `wcs_spacecore_localization.st` and
`wcs_bmp1_localization.st`. Smoke completion and zero script failures do not erase
resource errors or establish a clean build. These external localization errors
are separate from the earlier packaging shutdown failures.

The journals also show routine status messages reaching HQ and a physical FIM-92
observed in a weapon slot. The memory run records another selected actor reaching
its bounded equipment deadline. This is limited runtime observation, not proof
that every selection succeeds, that skill restoration works in multiplayer, or
that native AI launches the Stinger correctly.

The fixtures do not prove the complete six-and-a-half-minute missed-check-in
sequence, ordinary corpse discovery, long pursuit/containment play, twelve real
clients, remote ownership, full ordinary escape, aircraft operations or FPS gains.
The current status/body/equipment classes compile in the captured source; their
additional lifecycle cases are covered by focused regressions rather than those
native assertions.

## Radio scheduling follow-up

Due routine status and requested check-ins now claim the next free uplink before
new visual reports. An occupied transmission is never replaced. A regression of
the actual report methods reproduced the old starvation at five-second sighting
intervals over a six-second route: no status arrived during ten simulated minutes.
The revised ordering delivers periodic status while continuing visual reports.
Search completion also samples the physical group's native visible contacts before
accumulating its thirty-second dwell, including while another report occupies the
radio. A quiet combat flag alone cannot produce an area-clear report. These new
scheduling cases have method-level regression coverage; the earlier native memory
and response archives predate this follow-up.

## Repository checks

Actual extracted Enforce method bodies are exercised by the force-ledger,
HQ-memory, pursuit, doctrine, coverage, watch, status-sampler, body-observer,
equipment and assignment-lifecycle tests. The ledger harness advances 128 records
for a simulated hour; this is arithmetic regression coverage, not an engine FPS
benchmark. Intel readiness tests verify late physical binding without duplicate
HQ units or new Intel rolls.

The repository check list completed in two parts: the successful prefix, the
corrected packed-archive suite (26 tests), then all 58 remaining suites without
failure. The archive correction constructs its synthetic historical five-addon
payload from the frozen packed schema, rather than copying today's externally
dependent AI project. An added regression confirms external dependencies still
fail that frozen packed closure. Validation was not weakened.

## Physical convoy follow-up

The [finite convoy controller](convoy-migration.md) now integrates the six-vehicle
owned convoy composition, nineteen explicit native seat occupants, finite physical
truck cargo, shared Core admission budget, real-road preparation and canonical HQ
admin dispatch. A new independently authored memberless group prevents vanilla's
queued six-soldier allocation from bypassing that budget. Native seated-group
waypoint completion drives route advancement. The02:50:35 native run completes181
assertions with exact original identities and actual movement, while retaining the
four external localization errors. It remains an opt-in first physical slice;
combat tactics, unloading, automatic scheduling and full-route reliability are
still pending. The earlier feature inventory counts remain explicitly baseline
counts, not an assertion that this follow-up was present at that older commit.

## Delivered casualty reconciliation follow-up

Status snapshots and body observations now carry slots from the same immutable
initial formation roster. Watch unions those delivered identities rather than
adding aggregate totals: a casualty reported by both its squad and a passing
patrol counts once, while a different discovered body adds one. Local death
sampling and a queued or interrupted report cannot change HQ's casualty total.

Body evidence preserves the last actual check-in's ID, observation time and
position. A later check-in from a caller who has not seen an outside discovery is
still accepted without resurrecting a confirmed casualty. When every original
slot has delivered death evidence, Watch can close a wiped formation without
inventing a surviving caller. A scalar source/slot snapshot remains usable when
the corpse reference disappears during radio transit. Known transfers and player
conversions permanently exclude the released slot; another actor cannot replace
it. Body evidence with no previous check-in leaves status position and current living/fit
counts unknown rather than inventing a corpse-position check-in.

The full repository checks pass after this integration. The actual Watch,
Report.Snapshot and HQ.Accept methods run together in regressions for overlap,
disjoint casualties, deletion in transit, new locally uninformed status and wipe
closure. Sampler/body tests cover immutable slot ordering, released members and
a deleted native group handle with a retained corpse. These adapters are not
native LOS or radio evidence; that qualification is tracked separately.
The [02:57:26 HQ response regression](evidence/native-2026-09-20-025726-hq-casualty-integration/report.json)
compiles these changes and completes35 assertions through actual perception,
uplink, HQ receipt, downlink and movement of a distinct original squad. It has no
script/runtime failures; the four external localization errors remain. This
regression does not yet prove actual corpse observation.

## Native body and radio follow-up

[03:05:33 body/radio run](evidence/native-2026-09-20-030533-body-radio-casualties/report.json)
completes one scripted round with45 assertions and no script/runtime failures.
One existing non-radio enemy was moved once beside another original living guard
formation and killed using native character damage. Production BodyObserver's
native LOS created the snapshot. The test inspected the actual pending uplink,
verified no HQ loss before delivery, then observed precisely one delivered
casualty and a low-confidence BODY belief with no fire contact. The victim's
surviving caller sent its real cumulative death-slot status afterward; HQ still
counted one loss, including twenty further seconds of observation.

The fixture temporarily postponed that formation's check-in to isolate body
report causality, then resumed it. It created no actors, groups or synthetic
reports and never injected Watch knowledge. This is an arranged native visibility
case, not organic patrol discovery or an all-dead native formation. Scalar
in-transit deletion and full-wipe closure remain actual-method regressions.

The same source includes the reviewed transfer-before-death fix: a retained
non-null foreign group releases the old roster slot before either status or body
attribution; a null group on a destroyed actor remains valid native detachment.
A two-formation regression reproduced the previous double count and now confirms
only the destination roster counts the known transferred casualty.

The four external localization errors still prevent strict native qualification.
Source and all external inputs remained unchanged; hidden/muted owned-process
cleanup was confirmed. The full repository suite passed after casualty integration;
the subsequent transfer guard and fixture additions passed their focused checks,
source-bundle preview and native compilation. All533 protected original reference
files remained unchanged.

## HQ convoy cover and ammunition follow-up

The finite convoy now accepts per-vehicle HOLD/RESUME commands through each
original crew's existing HQ radio downlink. Planning requires a delivered own
status no older than 180 seconds and a delivered positive sighting less than
60 seconds old within 400 m of that reported position. Delivery rechecks evidence,
command/sighting lineage, the consumed source receipt, exact original crew and
physical Core record. HOLD preserves the route index while removing the owned
move waypoint; RESUME needs fresh own status and no qualifying contact after the
held sighting's 60-second window. Lost radio or superseding information cannot
silently resume travel. Native `AddUsableVehicle` admission supplies the mounted
combat hooks; actual gunner firing remains an engine validation question. Player
capture still releases the task without replacing its vehicles, crew or cargo.
See [convoy cover details](convoy-migration.md#delivered-hq-mounted-cover-follow-up).

Routine 180-second status now includes detached compatible small-arms magazine
and chamber counts from the immutable own roster. Only a delivered, complete,
no-older-than-180-second report proving zero small-arms rounds/chambers can veto a
new armed infantry assignment or its delivery. Unknown/stale information never
pretends to be empty; existing assignments and release commands are preserved.
Separate launcher fields do not drive this initial veto. The addition does not
migrate urgent refill requests, vehicle ammunition, general resupply or ammo
creation. See [HQ ammunition scope and tests](hq-ammunition.md).

Actual-method regressions cover both additions. The full repository check passed
at the integration checkpoint and the original-reference guard retained all 533
protected files unchanged. Those checks do not supply native radio/weapon or
inventory evidence by themselves. The [native ammunition follow-up](evidence/native-2026-09-20-040857-hq-ammunition/report.json)
completes 49 assertions: real original-roster ammunition counts, actual radio
delivery and detached unchanged Watch data, alongside the existing casualty
checks. No script/runtime failures occurred; four external localization errors
remain. [Ammunition documentation](hq-ammunition.md#validation-boundary) retains
the failed attempts and scope, including the later derived-well refinement.
The [current-source convoy recheck](evidence/native-2026-09-20-041150-convoy-cover-driving/report.json)
completes 181 assertions with original crew/cargo and all six vehicles driving;
source/dependency identity and cleanup are confirmed, with the same four external
localization errors. Native HOLD/RESUME combat transitions still need direct runtime proof; policy
regressions are not gunner firing, multiplayer or FPS evidence.

## Remaining work

- **T2 presence and a real shared active-agent budget:** materialization,
  dematerialization, observer exclusions, hysteresis, dwell time and exact loss
  preservation. Route progress must be reconciled when a moved physical patrol
  becomes dormant. Current arithmetic alone does not reduce active AI cost.
- **Occupation and broader activity:** authored checkpoints, conditional guards,
  smart actions, ambient signals and HQ deployment objectives from T3–T7/O1–O6.
  Current objective scoring is not a completed occupation state machine.
- **Ground and air capabilities:** the T10 hunter convoy, road routing/reactive
  roadblocks, planner-selected reconnaissance flights and operational shared-pool
  air transport were not present in these native snapshots. Independent aircraft
  experiments and historical courier/mortar tests are not proof of those features.
- **Broader loss-reconciliation qualification:** the single arranged native
  corpse/radio/status-overlap case below passes. Ordinary patrol discovery,
  deletion during real radio transit, complete native wiped formations and remote
  ownership still need their own evidence; silence alone never proves casualties.
- **Further companion behavior:** full road supply convoys, broader ammunition
  chains, HQ-deployed mortar sites, aircraft servicing/recovery, improvised-prison/crash starts and
  optional warned mine-laying remain separate migrations or qualification work.
- **Play and release validation:** adversarial radio interruptions, captures,
  simultaneous investigations/retasking, long sessions and real multiplayer;
  equivalent-workload FPS measurements; clean external resources/native packaging;
  and refreshed packages bound to the final source. The earlier playtest PAKs do
  not contain this checkpoint merely because their editable source folder exists.

Core remains vanilla-only. The AI project now declares the user-authorized
RHS/WCS/BMP dependencies recorded in [external policy](external-addons.json), with
six installed addons in the audited closure; external asset bytes remain outside
this repository. The earlier vanilla-only full-scenario package and release
acceptance wording therefore describe an older scope. This document does not
silently promote or rewrite the [pending acceptance cases](acceptance.json),
[original responsibility inventory](original-feature-coverage.md), or the
preexisting living-island design documents. No complete original-feature parity
or release-completion percentage is claimed.